Integrative multi-omics analysis identifies AEBP1 and EFEMP2 as key regulators of immune heterogeneity and therapeutic response in glioblastoma

Abstract

Objective

Glioblastoma (GBM) is a highly aggressive brain tumor with complex immune microenvironment and molecular heterogeneity. This study aimed to characterize immune infiltration patterns and identify prognostic biomarkers in GBM.

Methods

We classified GBM samples using immune-related gene sets and correlated subtypes with molecular features. Key genes were identified through Cox and LASSO regression analyses. Multi-omics approaches including single-cell RNA sequencing, transcriptional network analysis, and molecular docking were employed to investigate therapeutic targets.

Results

Two immune subtypes (c1/c2) were identified, with c2 showing mesenchymal features and poorer prognosis. Four immune-related genes (RPL39L, AEBP1, EFEMP2, GALNT12) were prognostic markers, with AEBP1 and EFEMP2 overexpressed in GBM. Single-cell analysis revealed five tumor subtypes, with MES-like being most malignant. NFIA and BATF3 were key regulators. Seven potential drugs (e.g., Bosutinib) were identified with stable target binding.

Conclusions

This study reveals immune-molecular interactions in GBM, identifies AEBP1/EFEMP2 as prognostic markers, and proposes targeted therapies for personalized treatment.

Supplementary Information

The online version contains supplementary material available at 10.1007/s12672-025-03500-4.

Introduction

Glioblastoma (GBM) is the most common and aggressive form of primary malignant brain tumor in adults, characterized by rapid proliferation, extensive invasion, and dismal prognosis [1, 2]. Despite decades of clinical advances—including maximal surgical resection, radiotherapy, and temozolomide-based chemotherapy—the median survival for patients with GBM remains approximately 12 to 18 months [3]. A key contributor to treatment failure is GBM’s profound molecular complexity, which enables tumor cells to dynamically reprogram their transcriptomes, remodel the microenvironment, and resist therapeutic pressure [4].

Among these adaptations, immune evasion has emerged as a hallmark feature of GBM [5]. However, unlike tumors with overt immune infiltration, GBM tumor cells often orchestrate an immunosuppressive state through intrinsic gene expression programs that limit antigen presentation, interfere with immune activation, and shape myeloid-dominant infiltration [6]. These tumor-intrinsic mechanisms—rather than immune cell dynamics per se—may represent a central axis of immune resistance and therapeutic failure in GBM. Nevertheless, the transcriptional networks that govern such immune-related programs in tumor cells remain poorly defined.

In this study, we aimed to dissect the immune-regulatory transcriptional landscape of GBM from a tumor cell–intrinsic perspective. By integrating bulk and single-cell transcriptomic data, immune stratification, machine learning–based prognostic modeling, and transcription factor network inference, we identified distinct tumor-driven immune subtypes and candidate oncogenic regulators associated with poor prognosis. Furthermore, we employed computational drug screening to propose potential small molecules capable of modulating these regulatory axes.

Together, our findings provide new insight into how GBM tumor cells exploit immune-related pathways to maintain plasticity and resist treatment, and offer novel targets for therapeutic intervention in this intractable malignancy.

Materials and methods

Immune profiling

Based on the expression levels of immune-related genes defined by Jia et al. [7], hierarchical clustering of samples in the TCGA glioblastoma cohort was performed using Ward’s minimum variance method (Ward.D) and Euclidean distance to measure similarity between samples. The specific gene set used for clustering is provided in Supplementary Table 1. The Chi-Square test was used to assess the correlation between clustering groups and GBM molecular subtypes. Kaplan–Meier survival analysis was performed to compare overall survival (OS) differences among patients across the various clustering groups. Single-sample Gene Set Enrichment Analysis (ssGSEA) was used to assess the level of immune infiltration in the samples. A p-value of less than 0.05 was considered statistically significant.

Identification of characteristic genes

Differentially expressed genes between the two immune clustering groups were identified using the limma package [8]. Prognostic-related genes were screened using univariate Cox regression and Least Absolute Shrinkage and Selection Operator (LASSO) regression analysis. Based on risk scores (median cutoff value), the samples were divided into low-risk and high-risk groups. The dataset was randomly split into training and validation cohorts for LASSO regression [9], using a 1000-fold tenfold cross-validation method to minimize instability in results. The optimal penalty parameter (λ) was chosen as the λ with minimal bias, and genes with non-zero coefficients were considered the best variables for subsequent analysis. We stratified samples into high and low expression groups based on the median expression of the target gene, performed differential expression analysis using limma, extracted log fold change (logFC) to construct a ranked gene list, and conducted GSEA using the gseKEGG function from the clusterProfiler package to identify pathways enriched in the high and low expression groups. Pathways with NES > 0 were considered enriched in the high-expression group, while those with NES < 0 were enriched in the low-expression group.

Analysis of transcriptional expression levels and prognosis of characteristic genes

The unified, standardized pan-cancer dataset the cancer genome atlas (TCGA) Pan-Cancer was downloaded from the University of California Santa Cruz (UCSC) database (https://xenabrowser.net/) and expression data were extracted. We analyzed the correlation between the transcriptional expression levels of the characteristic genes and patient survival using datasets from the Chinese Glioma Genome Atlas (CGGA) [10–13], TCGA, and Rembrandt databases [14]. To further validate the prognostic value of these genes, we performed additional survival analyses in three independent GBM cohorts: Ducray, Gravendeel, and Nutt [15–17]. The CPTAC (Clinical Proteomic Tumor Analysis Consortium) proteomic data were analyzed using the UALCAN online platform (https://ualcan.path.uab.edu/index.html).

Mutation landscape, methylation, and copy number variation of characteristic genes

Genetic variation analysis of characteristic genes in glioblastoma (GBM) was conducted using the Gene Set Cancer Analysis (GSCA) database (http://bioinfo.life.hust.edu.cn/GSCA/#/). The RSEM-normalized mRNA expression data, corrected for batch effects, were downloaded from UCSC Xena (http://xena.ucsc.edu/). Single nucleotide variation (SNV) data were obtained from the Synapse database (ID: syn7824274), and only deleterious mutations were included in the analysis, including missense mutations, nonsense mutations, frameshift insertions and deletions, splice site alterations, and in-frame insertions or deletions. These deleterious alterations were defined as mutations in the SNV-related modules. Copy number variation (CNV) data were downloaded from the TCGA database and processed using GISTIC2.0 to identify significant genomic amplifications and deletions.

For DNA methylation analysis, Illumina HumanMethylation450K level-3 data were used. For each gene, Pearson correlation analysis was performed between gene expression and methylation probes to identify the most negatively correlated CpG site, which was used to represent the methylation level of the gene.

scRNA-seq data processing and analysis

The Seurat package in R 4.3.0 was used to explore scRNA-seq data from GSE138794, GSE141383, GSE173278, and syn22257780 [18]. Batch effects were eliminated using the “Harmony” R package. After quality control by calculating the mitochondrial gene ratio and excluding low-quality cells, the top 2000 highly variable genes were selected using the “FindVariableFeatures” function. Principal component analysis (PCA) was performed for dimensionality reduction (npcs = 20). Subsequently, the uniform manifold approximation and projection (UMAP) algorithm (dims = 1:17) was used for further analysis. Cells were annotated as 9 main cell types based on typical marker genes using the “FindNeighbors” and “FindClusters” functions for subsequent analysis.

Downstream single-cell analysis

To distinguish malignant and normal cells, the “infercnv” package was used to estimate CNV. CNV scores were estimated based on gene annotations and a 100-gene moving average. The K-Nearest Neighbors (KNN) algorithm, a supervised learning method, was used to classify scRNA-seq data based on the distance between bulk-seq and scRNA-seq data. Metabolic activity of tumor cell subtypes was assessed using marker gene sets corresponding to MES-like, AC-like, NPC-like, and OPC-like cells, as defined by Aizhen Xiong et al. Feature scoring was performed using the “AddModuleScore” function in Seurat [19]. The full list of marker genes is provided in Supplementary Table 2. A phylogenetic tree of subtypes was constructed using the “clustree” package, and gene ontology (GO) analysis was performed using the “clusterProfiler” package [20]. CIBERSORTx (https://cibersortx.stanford.edu/) was used for deconvolution analysis of CGGA mRNA_693 data. Gene set scoring was performed using the “GSVA” package after downloading hallmark gene sets from msigdb. Single-cell Weighted Gene Co-expression Network Analysis (WGCNA) was performed using the “hdWGCNA” package [21], followed by transcription factor (TF) motif scanning from JASPAR database [22], and XGBoost modeling to identify relationships between gene expression and TF expression in modules.

Drug prediction, molecular docking, and simulation

Drug sensitivity analysis of GBM tumor cells was performed by combining large-scale pharmacogenomics data with scRNA-seq transcriptomic data. The GSCA online tool was used for drug sensitivity analysis of characteristic genes, leading to the identification of intersecting drugs. Three-dimensional protein structures were downloaded from the Protein Data Bank (PDB, https://www.rcsb.org/) and AlphaFold (https://alphafold.com/), and saved in PDB format as protein receptors. Two-dimensional small molecule structures were downloaded from the PubChem database, saved as “SDF” format, and converted to mol2 format using OpenBabel (version 3.1.1) as small molecule ligands. Water molecules and original ligands were removed from protein structures using PyMOL. The PDB and mol2 files were converted to pdbqt format and hydrogenated using AutoDockTools (version 1.5.7). The X, Y, Z centers were adjusted to the original ligands on different receptors, with the center coordinates for AEBP1 and EFEMP2 being (0.155, 1.834, − 18.766) and (− 0.889, − 0.25, − 2.639), respectively. Finally, molecular docking was performed using AutoDock Vina (version 1.1.2), and the binding energy was evaluated, with lower binding energy indicating stronger binding affinity and stability. The results with the lowest receptor-ligand binding energy were visualized using PyMol (version 3.1.3). Molecular dynamics simulations were carried out using the iMOD server (http://imods.chaconlab.org) to assess the stability and structural dynamics of the protein–ligand complex [23–25].

Results

Immune subtype analysis, and key gene identification in GBM

To explore the immune microenvironment heterogeneity of GBM, we performed hierarchical clustering of GBM samples based on the transcriptional levels of classical immune-related genes, resulting in the identification of two immune subtypes (c1 and c2) (Fig. 1A, B). Further analysis revealed a significant correlation between immune subtypes and GBM molecular subtypes (proneural, classical, and mesenchymal). The c2 subtype was predominantly associated with mesenchymal GBM, while the c1 subtype was primarily linked to classical and proneural GBM (Fig. 1C). Survival analysis showed that patients with the c2 subtype had significantly lower 1-year and 3-year survival rates compared to those with the c1 subtype (Fig. 1D). Differential gene expression analysis revealed several significantly upregulated immune-related genes in the c2 subtype, such as IL2RA, MARCO, and CXCL6, suggesting their involvement in the immunosuppressive microenvironment [26–29] (Fig. 1E, Supplementary Table 3).

Fig. 1.

Immune subtype analysis, and key gene identification in GBM. A Overview of the study workflow, including immune subtype clustering, survival analysis, and key gene selection. B Dendrogram showing hierarchical clustering of samples based on immune-related gene expression using Euclidean distance and Ward’s linkage method. C Heatmap illustrating the correlation between identified immune subtypes and established GBM molecular subtypes, calculated by Pearson correlation. D Dot plot showing 1- and 3-year survival rates across immune subtypes. Survival analysis was performed using the Kaplan–Meier method on TCGA data. E Volcano plot of differentially expressed genes (DEGs) between immune subtypes (c2 vs. c1), identified using the limma package with thresholds of |log2 fold change|> 1 and adjusted p-value < 0.05. Key upregulated genes are labeled. F LASSO regression performed with the glmnet package to select an optimal prognostic gene signature. G Box plot depicting risk scores from the LASSO model stratified by patient survival status. H Receiver operating characteristic (ROC) curve evaluating the predictive performance of the risk model, with an area under the curve (AUC) of 0.6608. I Forest plot showing hazard ratios (HR) of key prognostic genes from univariate Cox regression analysis. J Correlation analysis between key survival-associated genes (RPL39L, AEBP1, EFEMP2, GALNT12) and immune cell infiltration levels estimated by ssGSEA

To further identify key survival-predictive genes, we performed univariate Cox analysis to assign a risk score to each sample and combined it with LASSO regression for feature selection, ultimately determining an optimal gene set (Fig. 1F, G). The predictive performance of the survival model constructed based on these genes was evaluated using ROC analysis, yielding an AUC of 0.6608, indicating moderate predictive accuracy (Fig. 1H). Four genes—RPL39L, AEBP1, EFEMP2, and GALNT12—were significantly associated with patient survival (Fig. 1I). Correlation analysis revealed significant associations between these genes and immune cell infiltration levels, suggesting their potential roles in modulating the immune microenvironment in GBM (Fig. 1J).

Subsequently, we conducted single-gene GSEA analyses on these four genes to investigate their functional implications in GBM (Figure S1A). The results demonstrated that high expression of these genes was predominantly associated with enrichment of pathways involved in cell proliferation, metabolic activity, and stemness features, which may contribute to tumor growth advantages and therapeutic resistance, potentially explaining their link to poor prognosis. Conversely, low expression levels correlated with pathways related to immune activation and inflammatory responses, indicative of a more active immune microenvironment. These findings suggest that RPL39L, AEBP1, EFEMP2, and GALNT12 may influence GBM progression and patient outcomes through dual mechanisms involving intrinsic tumor cell phenotypes and regulation of the tumor immune microenvironment.

Expression and survival analysis of key genes across different datasets

To systematically evaluate the expression levels of RPL39L, AEBP1, EFEMP2, and GALNT12 in GBM and normal tissues, we first analyzed their pan-cancer expression profiles in the TCGA database (Fig. 2A). The results showed that AEBP1 and EFEMP2 were significantly upregulated in GBM tissues (p < 0.05), whereas RPL39L and GALNT12 exhibited an increasing trend but did not reach statistical significance (p > 0.05).

Fig. 2.

Expression and survival analysis of key genes across different datasets. A Violin plots display the expression levels of RPL39L, AEBP1, EFEMP2, and GALNT12 in tumor and normal tissues from the TCGA database. B Box plots illustrate the expression patterns of AEBP1 across different WHO grades of gliomas (II, III, IV) and GBM molecular subtypes (Proneural, Classical, Mesenchymal) in the CGGA, TCGA, and Rembrandt databases. Kaplan–Meier survival curves compare the overall survival between patients with high and low AEBP1 expression. C Box plots illustrate the expression patterns of EFEMP2 across different WHO grades of gliomas (II, III, IV) and GBM molecular subtypes (Proneural, Classical, Mesenchymal) in the CGGA, TCGA, and Rembrandt databases. Kaplan–Meier survival curves compare the overall survival between patients with high and low EFEMP2 expression

Furthermore, we analyzed the expression patterns of AEBP1 and EFEMP2 across different glioma grades (II, III, IV) and GBM molecular subtypes (Proneural, Classical, Mesenchymal) in the CGGA, TCGA, and Rembrandt databases, and assessed their impact on patient survival outcomes. The results indicated that AEBP1 and EFEMP2 expression levels increased with glioma grade and were highest in Mesenchymal-type GBM (Fig. 2B, C). Survival analysis revealed that patients with high AEBP1 and EFEMP2 expression had significantly worse overall survival compared to those with low expression (p < 0.05), suggesting their potential role as unfavorable prognostic biomarkers (Fig. 2B, C, Figure S2). In addition, subtype-stratified survival analyses in the TCGA_GBM, CGGA_GBM, and Rembrandt_GBM cohorts revealed that high expression of AEBP1 and EFEMP2 was significantly associated with poorer overall survival in most cases across the Mesenchymal, Proneural, and Classical GBM subtypes (Figure S1B, C), further supporting their potential as adverse prognostic biomarkers across molecular subtypes. To complement these findings, we additionally analyzed the CPTAC proteomics dataset and observed consistent upregulation of AEBP1 and EFEMP2 at the protein level in GBM samples (Figure S1D), providing preliminary proteomic validation of their dysregulated expression.

Somatic mutation, DNA methylation, CNV analysis of AEBP1 and EFEMP2 in GBM

To systematically evaluate the genetic variation characteristics of AEBP1 and EFEMP2 in GBM, we first analyzed their SNV in the TCGA-GBM database. The results showed that the most common mutation type in GBM was missense mutation, with C>T and C>A being the predominant SNV categories (Fig. 3A). The mutation frequency of AEBP1 in GBM was 0.99%, with mutations predominantly occurring in the FASC, M14_CPX_like, and MMP14-HPX-C_like domains, primarily in the form of missense mutations (Fig. 3B). The main SNV types were C>T (3 cases) and C>A (1 case), indicating a specific nucleotide substitution pattern of AEBP1 in GBM somatic mutations (Fig. 3C). No mutations were detected in EFEMP2.

Fig. 3.

Somatic Mutations, Copy Number Variations (CNV), and Prognostic Analysis of AEBP1 and EFEMP2 in GBM. A Overview of somatic mutations in GBM samples, including mutation classification (Variant Classification), mutation types (Variant Type), single nucleotide variation classes (SNV Class), number of mutations per sample (Variants per Sample), a summary of mutation categories, and the top 10 most frequently mutated genes. Mutation data were obtained from the Synapse database (ID: syn7824274), and only deleterious alterations—missense, nonsense, frameshift insertions/deletions, splice site changes, and in-frame indels—were included in the analysis. B Lollipop plot showing the distribution of AEBP1 mutations across the protein sequence and the corresponding mutation types. C Distribution of SNV types in AEBP1 among GBM samples. D Kaplan–Meier survival curve assessing the association between AEBP1 somatic mutations and overall survival (OS) in GBM patients. Mutation-based survival data were derived from the GSCA database. E Association between AEBP1 methylation and survival outcomes in GBM. Methylation data were based on the Illumina HumanMethylation450K platform. For each gene, Pearson correlation was used to identify the most negatively correlated CpG probe with gene expression, which was then used to represent the gene’s methylation level. Survival analyses included disease-specific survival (DSS), OS, and progression-free survival (PFS). F Distribution of copy number variation (CNV) events for AEBP1 and EFEMP2 in TCGA GBM samples. G Spearman correlation analysis between CNV status and mRNA expression levels of AEBP1 and EFEMP2. RSEM-normalized expression data were downloaded from UCSC Xena and batch-corrected prior to analysis. H Kaplan–Meier survival analysis evaluating the prognostic impact of CNV alterations in AEBP1 and EFEMP2 on overall survival in GBM patients

Further analysis of the impact of AEBP1 somatic mutations on GBM prognosis using Kaplan–Meier survival analysis showed that AEBP1 mutations were not significantly associated with OS (p = 0.85) (Fig. 3D). Additionally, we investigated the potential association between AEBP1 and EFEMP2 methylation levels and patient survival, including disease-specific survival (DSS), overall survival (OS), and progression-free survival (PFS). The results showed no statistically significant correlations (p > 0.05) for either gene, although some trends were observed (Fig. 3E).

In terms of CNV analysis, AEBP1 was predominantly affected by heterozygous amplification (Hete. Amp.) in GBM tissues, whereas EFEMP2 mainly exhibited heterozygous deletion (Hete. Del.), suggesting distinct regulatory patterns of these two genes in genomic alterations (Fig. 3F). Additionally, we further analyzed the impact of CNV on mRNA expression, and Spearman correlation analysis revealed a significant positive correlation between AEBP1 CNV and its mRNA expression (ρ = 0.36, FDR = 1.9 × 10⁻⁴), whereas EFEMP2 CNV was not significantly correlated with mRNA expression (ρ = 0.04, FDR = 0.84). These findings suggest that AEBP1 expression may be regulated by CNV changes, whereas EFEMP2 expression might be influenced by other regulatory mechanisms (Fig. 3G).

Finally, we evaluated the impact of AEBP1 and EFEMP2 CNV alterations on patient survival outcomes. Survival analysis indicated that patients with AEBP1 amplification had significantly worse OS, PFS, and DSS (p < 0.001), suggesting that AEBP1 amplification may serve as an unfavorable prognostic factor in GBM. In contrast, EFEMP2 CNV alterations had no significant impact on OS, PFS, or DSS (p > 0.05), implying that EFEMP2 CNV may not directly affect GBM prognosis (Fig. 3H).

Single-cell heterogeneity analysis and key gene expression characteristics in GBM tumor cells

To investigate cellular heterogeneity in glioblastoma (GBM), we constructed a single-cell transcriptomic atlas. Using UMAP dimensionality reduction, we identified nine major cell types. Further analysis of AEBP1 and EFEMP2 expression patterns in brain tissue revealed that both genes are significantly upregulated in GBM tumor cells (Figure S3A). The tumor cells were further subdivided into five subtypes: neural crest-like, olfactory ensheathing cell-like, neurogenic progenitor-like, anaplastic astrocyte-like, and an undefined subtype (Figs. 4A and D). The accuracy of cell type classification was validated by examining the expression patterns of known marker genes (Fig. 4B). InferCNV analysis demonstrated prominent chromosomal abnormalities in GBM tumor cells, notably amplification of chromosome 7 and deletion of chromosome 10 [30] (Fig. 4C).

Fig. 4.

Single-cell sequencing-based analysis of GBM tumor cell heterogeneity and key gene expression characteristics. A UMAP clustering of single-cell RNA-seq data delineates the cellular landscape of GBM tumor cells. B Dot plot showing the expression patterns of canonical marker genes across distinct GBM cell types. C Inference of copy number variations using inferCNV, revealing characteristic chromosome 7 amplifications and chromosome 10 deletions in malignant GBM cell populations (AC-like, MES-like, NPC-like, OPC-like, and unknown) compared to non-malignant reference cells. D Re-clustering and refined annotation of malignant GBM tumor cells into distinct subtypes based on transcriptional signatures. E Projection of classical GBM molecular subtypes (Proneural, Classical, Mesenchymal) from bulk TCGA RNA-seq data onto the single-cell map to assess concordance at the cellular level. F Mapping of immune subtypes (c1 vs. c2) and Cox-based prognostic risk groups (high vs. low) onto the single-cell atlas, linking tumor-intrinsic transcriptional programs with immune context and survival risk. G Functional enrichment analysis (GO and KEGG) of significantly differentially expressed genes among the five GBM tumor cell subtypes. H Box plots showing the proportions of each tumor cell subtype across individual GBM patients, highlighting inter-patient heterogeneity. I Kaplan–Meier survival curves comparing patient outcomes across tumor cell subtype-enriched samples, based on bulk expression deconvolution. J Gene set variation analysis (GSVA) of hallmark signaling pathways among tumor cell subtypes, identifying subtype-specific pathway enrichment. K Violin plots showing the expression distributions of AEBP1 and EFEMP2 across GBM tumor cell subtypes

To further characterize the identity of the unknown subpopulation in GBM, we extracted and re-clustered tumor cells. Compared to other subtypes, the unknown subtype exhibited the highest proportion of cells in the G2/M and S phases, along with significantly elevated metabolic activity (Figure S3B, C). After removing cell cycle-related effects, the unknown subtype still clustered into classical GBM classification (AC, MES, NPC, and OPC) (Figure S3D). Phylogenetic analysis further illustrated clear lineage relationships among these GBM subtypes (Figure S3E).

Furthermore, by mapping classical bulk RNA-seq GBM subtypes (Proneural, Classical, and Mesenchymal) from the TCGA database, as well as previously identified immune and Cox risk subtypes, onto the single-cell atlas, we confirmed the consistency between bulk RNA-seq and single-cell-based classifications (Fig. 4E, F).

We further identified the differential characteristic genes of distinct GBM subtypes (Figure S3F). Functional enrichment analysis revealed that subtype-specific differentially expressed genes (DEGs) in GBM were associated with distinct biological pathways. Specifically, the NPC-like subtype was enriched in neuroprogenitor-related pathways, including intracellular vesicle transport, axon development, and regulation of neuron projection development; the OPC-like subtype was enriched in oligodendrocyte precursor cell-associated pathways such as synapse organization, glial cell differentiation, and axon development; the AC-like subtype was significantly enriched in immune-related pathways, particularly endogenous antigen processing and MHC class I-mediated antigen presentation, suggesting involvement in tumor antigen presentation processes; the MES-like subtype was enriched in malignant progression-related pathways including hypoxia response, coagulation, and apoptosis signaling, indicative of higher malignancy and metastatic potential; and the unknown subtype exhibited enrichment in cell cycle-related pathways, including chromosome segregation and mitotic nuclear division, highlighting its high proliferative potential (Fig. 4G, Supplementary Table 4).

Comparative analysis of subtype proportions among different patients showed that the AC-like subtype was the most abundant (Fig. 4H). Survival analysis indicated that patients with the MES-like subtype exhibited the poorest prognosis (Fig. 4I). Furthermore, GSVA pathway analysis demonstrated that the MES-like subtype was significantly associated with hypoxia, epithelial–mesenchymal transition (EMT), and TNF-α signaling pathways, further supporting its aggressive characteristics (Fig. 4J). Finally, we evaluated the expression of key genes, AEBP1 and EFEMP2, and observed their highest expression levels in the MES-like subtype, suggesting their potential roles in driving GBM malignancy progression (Fig. 4K, Figure S3G).

Transcription factor regulatory network analysis reveals key regulators of AEBP1 and EFEMP2

To further clarify the function of core genes in the MES-like subtype, we applied the hdWGCNA algorithm to identify genes co-expressed with AEBP1 and EFEMP2 in the MES-like subtype. By selecting a flexible threshold of 8, genes were divided into 24 functional modules (Fig. 5A–C). The results show that AEBP1 and EFEMP2 are located in the MES-M21 and MES-M17 modules, respectively (Fig. 5D, E). Enrichment analysis indicates that genes in the MES-M21 module are mainly involved in neural system development, amyloid protein metabolism, cholesterol transport, ion regulation, and mechanisms related to the tumor microenvironment and neurodegeneration (Fig. 5F). In contrast, genes in the MES-M17 module are closely related to inflammation regulation, blood vessel development, and glial cell differentiation, which may affect the GBM microenvironment and neuroadaptation (Fig. 5G).

Fig. 5.

Transcription factor regulatory network analysis shows key regulatory factors of AEBP1 and EFEMP2. A Soft threshold selection plot used to determine the optimal threshold. B Dendrogram showing the co-expression module MES_like identified by hdWGCNA. C–E Identification of specific gene modules. F, G GO enrichment analysis of genes in the MES-M21 and MES-M17 modules. H XGBoost-based regulatory importance scores display the key transcription factors for AEBP1 and EFEMP2. I Regulatory network of NFIA and BATF3. Nodes represent individual genes, and they are colored based on their relationship with NR4A2. Directed edges represent TF regulatory relationships, colored according to the correlation of TF gene expression. J Bar chart showing the top predicted positively correlated target genes of NFIA and BATF3. K, L Selected pathway enrichment results of NFIA and BATF3 target genes. The results show the pathway enrichment for target genes with negative (left) and positive (right) gene expression correlations with NFIA and BATF3. M Heatmap summarizing the TF-mediated regulatory relationships between each pair of co-expression modules. Orange indicates stronger positive regulation than negative regulation, while blue indicates stronger negative regulation. N, O Network diagrams showing the regulatory networks of MES-M21 and MES-M17 modules as regulatory sources (source) and regulatory targets (target)

Next, we scanned the promoter regions of 660 TFs in the JASPAR TF motif database to detect the presence of TF binding motifs and identify potential regulatory targets for each TF. Then, we applied an ensemble learning algorithm (extreme gradient boosting trees, XGBoost) to model the expression of target genes based on the expression of potential TF regulators, calculating TF regulatory scores to quantify the importance of specific TFs in gene expression and identify the most likely regulatory factors. The results revealed that the key positive regulatory factors for AEBP1 and EFEMP2 are NFIA and BATF3, respectively (Fig. 5H). Further analysis of the regulatory networks formed by the primary and secondary target genes of NFIA and BATF3 (Fig. 5I, J), followed by pathway enrichment analysis in their major target gene sets, categorized according to the activation or inhibition of target gene effects. The results showed that NFIA and BATF3 target genes are widely involved in immune cell differentiation, angiogenesis, neural precursor cell proliferation, and neuronal differentiation pathways, which may play important roles in GBM development (Fig. 5K, L).

Additionally, based on the potential regulatory relationships between TFs and predicted target genes, we further analyzed the interactions between co-expression modules. The results showed that the MES-M21 module has a positive regulatory effect on four modules and a negative regulatory effect on three modules, while the MES-M17 module positively regulates five modules and negatively regulates two modules (Fig. 5M). Finally, we constructed network diagrams of the MES-M21 and MES-M17 modules as regulatory sources (source) and regulatory targets (target), further revealing their potential roles in GBM microenvironment regulation (Fig. 5N, 5).

Immune escape mechanisms and potential therapeutic drug prediction for GBM

To further explore the immune escape features and potential therapeutic strategies for different GBM subtypes, we conducted immune escape scoring and drug sensitivity prediction analyses. Firstly, the TIDE algorithm was utilized to comprehensively evaluate immune escape potential across three classical GBM molecular subtypes (Proneural, Classical, and Mesenchymal). The results indicated that the Mesenchymal subtype exhibited significantly higher dysfunction and microsatellite instability (MSI) scores (p < 0.05), suggesting a stronger immune escape capability (Fig. 6A).

Fig. 6.

Immune escape mechanism analysis and prediction of potential therapeutic drugs for GBM. A Violin plots depicting differences in TIDE, immune dysfunction, immune exclusion, and microsatellite instability (MSI) scores among three classical GBM molecular subtypes (Proneural, Classical, Mesenchymal). B Bar graph showing the proportion of resistant drugs across GBM single-cell subtypes (NPC-like, OPC-like, AC-like, MES-like). C Top 30 potential therapeutic drugs for GBM tumor cells predicted by the scPharm database. D Correlation analysis between drug sensitivity and expression levels of key genes AEBP1 and EFEMP2 based on the GDSC databases. E The Venn diagram shows the overlap between drugs predicted by scPharm and those predicted by the GDSC databases. F The lollipop chart illustrates the shared potential drugs for AEBP1 and EFEMP2. G Molecular docking simulation of AEBP1 with Bosutinib, Methotrexate, and Ruxolitinib, and EFEMP2 with XAV939. H, I Eigenvalue and variance plots [Colored bars show the individual (purpule) and cumulative (green) variances]. J Elastic network model

Further analysis based on single-cell subtypes (NPC-like, OPC-like, AC-like, MES-like) revealed that the MES-like subtype displayed the highest proportion of resistant drugs, implying greater therapeutic difficulty (Fig. 6B). Using the scPharm computational framework, we integrated the single-cell transcriptomic profiles of GBM tumor cells with drug genomic profiles to predict the pharmacological phenotypes of MES-like cells. As a result, we identified 30 potential drugs that may be effective against GBM tumor cells, including RVX-208, IWP-2, and Bosutinib (Fig. 6C). Concurrently, drug sensitivity analysis using GDSC databases identified several drugs significantly correlated with the expression of key genes AEBP1 and EFEMP2 (Fig. 6D, Supplementary Table 5).

Intersection analysis of scPharm-predicted drugs and those identified from the GDSC databases revealed five drugs targeting AEBP1 (Bosutinib, Methotrexate, Ruxolitinib, Cisplatin, and Neratinib) and two drugs targeting EFEMP2 (XAV939 and Cisplatin) (Fig. 6E, F). To further validate the binding potential between these drugs and their target proteins, molecular docking analysis was performed. The results showed that AEBP1 successfully docked with Bosutinib, Methotrexate, and Ruxolitinib, while EFEMP2 successfully docked with XAV939, with binding affinity scores of − 6.1, − 6.7, − 8.0, and − 5.6 kcal/mol, respectively (Fig. 6G). Generally, a binding affinity score lower than − 5 kcal/mol suggests that the complex is likely to form stably. These findings suggest that the aforementioned drugs may exert therapeutic effects by directly targeting AEBP1 and EFEMP2.

To further evaluate the stability and flexibility of the docking complexes, molecular dynamics (MD) and normal mode analysis (NMA) simulations were performed using the iMod server. The eigenvalue plots indicate the overall rigidity of the protein–ligand complexes, where a lower eigenvalue suggests greater flexibility (Fig. 6H). The variance plots display the percentage of motion variance across different modes, with higher variance indicating regions of greater flexibility (Fig. 6I). Additionally, the elastic network diagrams illustrate the atomic interactions within the docked proteins, where darker gray regions represent more rigid structural areas, suggesting stronger constraints in molecular motion (Fig. 6J). These findings provide insights into the conformational dynamics and structural adaptability of AEBP1-Ruxolitinib and EFEMP2-XAV939 complexes.

Discussion

Glioblastoma (GBM) remains one of the most treatment-resistant malignancies, with its aggressive progression driven by extensive cellular heterogeneity and immune evasion [31]. In this study, we applied machine learning-based immune profiling and single-cell transcriptomics to characterize GBM immune subtypes, malignant cell states, and potential therapeutic targets. Our findings characterize interactions between tumor and immune microenvironments, suggesting possible implications for personalized therapeutic strategies in GBM.

We identified two distinct immune subtypes, with c2 being associated with the mesenchymal (MES) molecular subtype. Survival analysis further confirmed the prognostic significance of these immune subtypes, with the c2 subtype showing a lower survival rate. This aligns with reports showing that MES-like GBM exhibits increased hypoxia, epithelial-to-mesenchymal transition (EMT), and immune checkpoint activation, all contributing to immune escape [32, 33]. Our study further investigated these findings by integrating immune-related gene expression clustering with transcriptomic subtypes, identifying an association between tumor plasticity and immune suppression mechanisms.

Through machine learning-based feature selection, we identified four immune-associated prognostic genes (RPL39L, AEBP1, EFEMP2, GALNT12), with AEBP1 and EFEMP2 being the most strongly associated with GBM progression. AEBP1 was found to be regulated by copy number variation (CNV) and showed a significant positive correlation with poor survival outcomes, consistent with its previously reported role in cancer-associated fibroblast activation and extracellular matrix remodeling [34, 35]. Similarly, EFEMP2 was highly expressed in the MES-like subtype, a finding supported by previous studies linking its function to tumor hypoxia adaptation and invasion [36].

Single-cell RNA sequencing further revealed that AEBP1 and EFEMP2 were predominantly expressed in the MES-like GBM cell population. The MES-like state, characterized by high plasticity and therapy resistance, has been implicated in tumor recurrence and progression. Our analysis identifies these genes as candidate biomarkers associated with high-risk GBM subtypes, and their molecular roles warrant further investigation as possible therapeutic targets against mesenchymal transformation.

Using hdWGCNA analysis, we identified that AEBP1 and EFEMP2 are localized within distinct mesenchymal (MES)-specific co-expression modules, suggesting their involvement in different transcriptional programs within the MES state. Specifically, the MES-M21 module is primarily enriched for pathways related to neural development and neurodegeneration, whereas the MES-M17 module is significantly associated with inflammatory responses, angiogenesis, and extracellular matrix remodeling. Based on these findings, we propose that AEBP1 may characterize an inflammatory/angiogenic MES subtype, potentially promoting tumor progression through modulation of the tumor microenvironment; in contrast, EFEMP2 may define a neurodevelopmental MES state, which may be linked to tumor cell plasticity and neural lineage characteristics. Although further validation is needed, this functional stratification of MES subtypes enhances our understanding of glioblastoma heterogeneity and may provide valuable insights for the development of subtype-specific therapeutic strategies.

By constructing transcriptional regulatory networks, we identified NFIA and BATF3 as upstream regulators of AEBP1 and EFEMP2, respectively. NFIA has been reported to play a role in glioma development and glial cell differentiation [37, 38], whereas BATF3 is involved in immune regulation and dendritic cell differentiation [39–41], suggesting a potential link between tumor-intrinsic pathways and immune modulation. These findings shed light on potential mechanisms by which GBM tumors may foster an immunosuppressive microenvironment alongside sustained malignant plasticity.

Given the clinical relevance of AEBP1 and EFEMP2, we further assessed their druggability based on structural and localization characteristics. AEBP1 is primarily localized in the cytoplasm and nucleus, and prior studies have suggested its potential as a target for small-molecule inhibitors [42]. EFEMP2, an extracellular matrix glycoprotein enriched with EGF-like domains, has been investigated as an antibody target in other tumor types [43]. These features suggest that both genes may be accessible to therapeutic interventions via small molecules or antibody-based strategies. To explore therapeutic vulnerabilities, we performed drug response prediction and molecular docking analyses, identifying Bosutinib and XAV939 as candidate drugs targeting AEBP1 and EFEMP2 (Supplementary Table 3). Bosutinib, a Src/Abl kinase inhibitor, has demonstrated efficacy in preclinical glioblastoma models by disrupting key oncogenic signaling pathways [44]. XAV939, a tankyrase inhibitor, is known to modulate Wnt signaling and may have implications for suppressing mesenchymal transition in GBM [45]. Molecular dynamics simulations further supported their stable binding potential, suggesting that these compounds warrant further investigation in preclinical GBM models.

Our study suggests that precision medicine approaches may be beneficial in GBM treatment, especially for patients with mesenchymal-like, immunosuppressive tumors. The characterization of immune subtypes and key prognostic genes could facilitate biomarker-guided patient stratification, potentially enhancing therapeutic outcomes with targeted and immunomodulatory agents.

Furthermore, our data indicate that combinatorial therapeutic approaches may help address the interconnected challenges of tumor plasticity and immune evasion. Based on the observed association between immune infiltration and the MES-like state, future investigations could evaluate potential synergies between immune checkpoint inhibitors and agents targeting AEBP1 and EFEMP2 pathways.

Conclusion and limitation

This study combines immune profiling, single-cell transcriptomics, regulatory network modeling, and drug response prediction to identify candidate prognostic markers and therapeutic targets in GBM. Our analysis suggests an association between immune infiltration, tumor plasticity, and treatment resistance, nominating AEBP1 and EFEMP2 as potential targets for further investigation. These findings may inform future precision medicine approaches to address immunosuppression in GBM.

While our integrated approach provides important insights, several limitations should be noted: (1) Although our transcriptomic analyses implicate AEBP1 and EFEMP2 in GBM progression and immune modulation, their mechanistic roles require experimental validation through functional studies; (2) while molecular docking and drug prediction analyses identify potential therapeutic compounds, their efficacy needs to be rigorously tested in GBM preclinical models; (3) given the marked heterogeneity of GBM, future investigations incorporating spatial transcriptomics and single-cell proteomics will be essential to fully elucidate tumor-immune microenvironment interactions.

Abbreviations

GBM

Glioblastoma

EMT

Epithelial–mesenchymal transition

scRNA-seq

Single-cell RNA sequencing

NPC-like

Neural-progenitor-like

OPC-like

Oligodendrocyte-progenitor-like

AC-like

Astrocyte-like

MES-like

Mesenchymal-like

Ward.D

Ward’s minimum variance method

OS

Overall survival

ssGSEA

Single-sample gene set enrichment analysis

LASSO

Least absolute shrinkage and selection operator

TCGA

The cancer genome atlas

UCSC

University of California Santa Cruz

CGGA

Chinese Glioma Genome Atlas

SNV

Single nucleotide variations

CNV

Copy number variations

GSCA

Gene Set Cancer Analysis

PCA

Principal component analysis

UMAP

Uniform manifold approximation and projection

KNN

K-nearest neighbors

GO

Gene ontology

WGCNA

Weighted Gene Co-expression Network Analysis

TF

Transcription factor

DSS

Disease-specific survival

PFS

Progression-free survival

MD

Molecular dynamics

NMA

Normal mode analysis

Author contributions

Y.Y. wrote the original draft, conducted the investigation and formal analysis, and contributed to the conceptualization. X.F. contributed to the investigation. S.G. performed validation, developed methodology, and curated data. Y.X. participated in validation and investigation. W.W. contributed to methodology development and data curation. Z.L. was responsible for project administration and conceptualization. S.W. performed validation and contributed to methodology. Z.G. and K.H. reviewed and edited the manuscript. C.L. and H.W. contributed to writing—review & editing, validation, project administration, and conceptualization. All authors reviewed the manuscript.

Funding

This work was supported by the Special Project for Clinical Research of the Shanghai Municipal Health Commission (grant number 202140403), the National Natural Science Foundation of China (grant numbers 82371839, 32370895 and 32070862), and the Natural Science Foundation of Ningxia (grant numbers 2024AAC05084).

Data availability

The datasets analysed during the current study are available in the following public repositories: RNA sequencing and microarray data of GBM samples (TCGA, CGGA, REMBRANDT, Ducray, Gravendeel, Nutt via Gliovis, https://gliovis.bioinfo.cnio.es/); proteomics data (CPTAC, https://gdc.cancer.gov/about-gdc/contributed-genomic-data-cancer-research/clinical-proteomic-tumor-analysis-consortium-cptac); mutation, DNA methylation, and CNV data (TCGA); single-cell RNA sequencing data (GEO: GSE138794, GSE141383, GSE173278 https://www.ncbi.nlm.nih.gov/geo/; Synapse: syn22257780, https://accounts.synapse.org/); three-dimensional protein structures (PDB: https://www.rcsb.org/; AlphaFold: https://alphafold.com/); and small-molecule drug structures (PubChem: https://pubchem.ncbi.nlm.nih.gov/). Other data supporting the findings of this study are available from the corresponding author upon reasonable request.

Code availability

All analysis scripts used in this study are publicly available at https://github.com/yinyiyy/Glioblastoma_Bioinformatics_Analysis.

Declarations

Ethics approval and consent to participate

This study was entirely based on publicly available datasets (e.g., TCGA and CGGA) and did not involve any experiments on human participants or animals conducted by the authors. Therefore, ethical approval was not required. Not applicable. All data used in this study were obtained from public databases with informed consent obtained by the original data providers.

Consent for publication

Not applicable. This study does not include any personal identifying information or new individual data.

Competing interests

The authors declare no competing interests.