A fibroblast-specific gene signature as a therapeutic target for glioblastoma developed based on the characteristics of tumor microenvironment Authors: Nan Liu, Mingyue Zhao, Yeting Cui, Jiaxuan Zhao, Yanyang Tu, Tongcun Zhang, Xiaofei Hu Published: 2025 ABSTRACT BackgroundThis study identified fibroblast-specific genes to develop a RiskScore model to improve prognostic accuracy and guide personalized treatment in glioblastoma (GBM).MethodsWe analyzed fibroblast-specific signatures in the GSE273274 cohort using “Seurat” R package for scRNA-seq data processing. Fibroblast-related gene modules were identified via WGCNA, and functional enrichment was assessed with “clusterProfiler” package. A RiskScore model was established using univariate, Lasso Cox regression analysis, and “survival” package, validated by “timeROC” for receiver operator characteristic (ROC) curve. Finally, immune infiltration and drug sensitivity was evaluated applying “ESTIMATE,” “TIMER,” “MCPcounter,” and “pRRophetic” packages. Experimental validation included qPCR for gene expression detection, and CCK-8, wound healing, and Transwell assays for functional measurement.ResultsThe scRNA-seq analysis identified nine cell types of cells, with fibroblasts elevated in the GBM group. Fibroblast signatures were linked to tumorigenesis, cytoskeleton remodeling, and regulation of neuronal development process that affected GBM invasion. A 6-gene RiskScore divided GBM patients into high- and low-risk groups in training and validation sets, with high-risk patients exhibiting poorer survival, elevated StromalScore, and negative correlations with the infiltration of neutrophils and B_cells. Moreover, high-risk patients demonstrated heightened sensitivity to Cisplatin, MG-132, AZ628, Dasatinib, CGP-60474, A-770041, TGX221, and Bortezomib. Finally, qPCR showed that the VWA1 was upregulated in GBM cells, while knock-down of VWA1 inhibited the cell proliferation, migration, and invasion activity.ConclusionWe constructed a RiskScore model for predicting the survival outcomes based on fibroblasts-related genes. These findings highlighted the role of fibroblasts in GBM development and offered six potential therapeutic targets (VWA1, DUSP6, LOXL1, IGFBP4, CYGB, and ZIC3) for GBM treatment. Additionally, immune infiltration analysis and drug sensitivity prediction further supported the model’s utility in guiding personalized treatment of GBM.Supplementary InformationThe online version contains supplementary material available at 10.1186/s40001-025-03528-w. Introduction Glioblastoma (GBM) is the most common primary malignant brain tumor, originating from progenitor cells and neuroglial stem cells that accumulate specific mutations [1] and are destined to astrocyte differentiation [2]. Approximately 90% cases are isocitrate dehydrogenase (IDH) wild type [3], while other subtypes include IDH-mutant astrocytoma and oligodendroglioma [4, 5]. In 2020, the United States reported approximately 1,603,844 new cases and 602,347 deaths related to GBM [6]. GBM exhibits aggressive growth and invasive characteristics, with a median survival of 15 months and a 5-year survival of 10%, and it is more frequently detected in males [7]. Surgical resection followed by radiation and chemotherapy are standard treatments for GBM, yet the median survival of GBM remains 18 months [8], and most patients experience recurrence within two years [9]. Aside from rare genetic predispositions (e.g., neurofibromatosis 1, Li-Fraumeni syndrome and turcot syndrome) or prior radiation exposure [10], no established risk factors have been confirmed for GBM [11, 12]. As marked heterogeneity and complexity of GBM often impede therapeutic advances, exploring the molecular mechanism and microenvironmental regulation of GBM is highly significant for understanding GBM pathogenesis and identifying new treatment targets. Recent evidence suggested that targeting both malignant cells and their surrounding ecosystem simultaneously may help overcome drug resistance and immune evasion [13]. Therefore, elucidating the composition of the tumor stroma and its role in modulating the tumor microenvironment (TME) is crucial for improving treatment efficacy [14]. Cancer-associated fibroblasts (CAFs) are a core component of the TME. In GBM, fibroblasts are activated by fibroblast activation protein (FAP) to affect tumor progression and chemoresistance [15, 16]. Previous study has identified CAFs as a major of GBM stromal-like cells [17] and demonstrated their ability to promote TGF-β secretion, which enhance tumor growth together with other immunosuppressive cytokines such as IL-10 and TGF-β [18, 19]. These findings indicated that CAFs features may be a reliable phenotype for GBM biomarker development. However, currently limited evidence of the presence of CAFs in GBM [20] required further studies. In this research, we analyzed the scRNA-seq data of GSE273274 from Gene Expression Omnibus (GEO) database. The analysis identified fibroblast populations based on specific gene marker expression. Prognosis-related genes were screened through scRNA-seq analysis, WGCNA, univariate analysis, and Lasso regression. A RiskScore model was then established to classify GBM patients into different risk groups for survival evaluation. In addition, we performed immune infiltration analysis and drug sensitivity analysis between high- and low-risk groups, offering insights for personalized treatment of GBM. Materials and methods Data acquisition and preprocessing The clinical information and RNA-seq data of GBM patients were derived from The Cancer Genome Atlas (TCGA) database. A total of 153 tumor samples with a survival time > 0 day were included for subsequent study after eliminating those with incomplete survival time or status. The RNA-seq expression profiles were converted to TPM format and log₂-transformed. Additionally, “CGGA_325” dataset (a validation set) was acquired from the Chinese Glioma Genome Atlas (http://www.cgga.org.cn) [21]. Single-cell RNA-seq analysis The scRNA-seq dataset (GSE273274) encompassing 3 tumor samples and 3 paired adjacent non-tumor tissue samples was collected from GEO database. In this study, scRNA-seq analysis was carried out utilizing “Seurat” R package [21, 22]. Following initial filtering with the CreateSeuratObject function to create a Seurat object, data were normalized by the SCTransform function. Principal component analysis (PCA) was conducted with the RunPCA function to reduce dimensionality, and batch effects were corrected by the RunHarmony function in the harmony package. The first 50 PCs were analyzed by ElbowPlot function. Then, the FindNeighbors and FindClusters functions, and the RunUMAP function were used for UMAP dimensionality reduction and visualization of cell population distribution. Eventually, cell types were annotated according to marker genes from the CellMarker2.0 database. Fibroblast-specific signature genes were identified from the DEGs between fibroblasts and other cell subpopulations applying FindAllMarkers function, followed by functional enrichment analysis. Function enrichment analysis DEGs across cell types were screened by the FindAllMarkers function. Fibroblast-specific DEGs were subsequently subjected to Gene Ontology (GO) and Kyoto Encyclopedia of Genes and Genomes (KEGG) analysis using “clusterProfiler” R package [23, 24]. WGCNA Fibroblast activity scores for TCGA-GBM samples were computed using “GSVA” R package. Gene modules correlated with the fibroblast signature was clustered by “WGCNA” R package. Briefly, the optimal soft-thresholding power β was determined by the pickSoftThreshold function. Hierarchical clustering was performed under the criterion of minModuleSize = 60 to section gene modules. Finally, the module showing the strongest correlation with the fibroblast signature score was selected [25, 26]. Hub genes with |Module Membership (MM)|> 0.7 and |Gene Significance (GS)|> 0.3 were selected to intersect with the DEGs to obtain key genes for GBM. Development and validation of a riskscore model Subsequently, univariate Cox regression analysis was performed on the key genes using the "survival" R package, retaining those with p < 0.01. Lasso regression analysis was then performed via the “glmnet” R package, followed by multivariate analysis to identify prognostic genes and calculate their coefficients. A RiskScore model established based on these genes were formulated as follows: RiskScore = \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\sum (\beta {\text{i}} \times {\text{Expi)}}$$\end{document}∑(βi×Expi), where β represents the coefficient, i is the gene expression. The optimal RiskScore cutoff was determined with the “survminer” R package to split GBM patients into high-risk and low-risk groups. KM survival analysis was performed using “survival” R package, and the receiver operating characteristic (ROC) curves were generated with “timeROC” R package [23]. Tumor microenvironment (TME) analysis We employed the ESTIMATE algorithm to evaluate the overall immune microenvironment infiltration score. The “MCPcounter” R package was used to quantify the abundance of 10 immune cell types, while TIMER online tool (http://cistrome.org/TIMER) was employed to estimate the infiltration another 6 types of immune cells in TCGA cohort [27]. Drug sensitivity analysis Drug sensitivity was evaluated using the “pRRophetic” R package on the TCGA-GBM dataset. The half-maximal inhibitory concentration (IC₅₀) of commonly used chemotherapeutic agents was estimated, and its correlation with the RiskScore was subsequently analyzed. Cell culture and transfection Dulbecco's Modified Eagle Medium (DMEM) with 1% penicillin–streptomycin (PS) and 10% fetal bovine serum (FBS) was prepared to cultivate human Glioma Cell line HEB (IM-H209) and GBM cell lines U251 (IM-H328) and U87 (IM-H211) at 37 °C in 5% CO₂ and 95% air. All cell lines were purchased from the Immocell Biotechnology (Xiamen, Fujian, China). The effect of VWA1 on tumor cells was examined using siRNA: si-VWA1#1 (SS: UUAAAGUCUGGGAUGUUAGUG, AS: CUAACAUCCCAGACUUUAAAA), si-VWA1#2 (SS: UGUUUGUGGAGGAGUAAAGCA, AS: CUUUACUCCUCCACAAACAGC) from Sangon Biotech (Shanghai, China). Following the manual, 100 nM siRNA was transfected into U251 and U87 cells by 100 nM Lipofectamine 2000 per well (11,668,027, Invitrogen, Carlsbad, CA, USA). The cells were stored at 37 °C for 48 h post-transfection. Quantitative PCR (qPCR) Total RNA was isolated using the commercial RNA extractor kit (R0017S, Beyotime, Shanghai, China). Cells were lysed in TRIzol, followed by chloroform separation [28]. After centrifugation (12,000 g, 15 min), RNA was precipitated with isopropanol, washed with ethanol, dried, and dissolved in DEPC water. Genomic DNA was eliminated utilizing RNase-free DNase I following the specification. cDNA was synthesized using the BeyoRT™ II kit (D7168S, Beyotime) with oligo(dT), dNTPs, buffer, and reverse transcriptase at 42 °C for 60 min, at 80 °C for 10 min. RT-qPCR was performed in 20 μL reactions containing cDNA, primers, and BeyoFast™ SYBR Green Mix (D7260, Beyotime) on a standard qPCR system: 95 °C for 2 min; 40 cycles of 95 °C for 15 s and 60 °C for 15 s (with fluorescence detection). Gene expression from three biological and technical replicates was calculated using 2−ΔΔCt method. Primer sequences are shown in Table 1. Table 1The specific primer for qPCRGeneForward (5′-3′)Reverse (5′-3′)VWA1GAGAAGCACCTGCACTTTGTGGTGGACGTGATCTCCGTGGCATGDUSP6CTCGGATCACTGGAGCCAAAACGTCACAGTGACTGAGCGGCTAALOXL1ACAGCACCTGTGACTTCGGCAACGGTTATGTCGATCCACTGGCAIGFBP4ACCCACGAGGACCTCTACATCACACACCAGCACTTGCCACGCTCYGBACCGCTGCCTACAAGGAAGTGGAAGGAGGGTCTTCAGAACTCGGZIC3TGCATACCTCGGACAAGCCCTAGATTCATAGCCTGAACTGGCAGCGAPDHGTCTCCTCTGACTTCAACAGCGACCACCCTGTTGCTGTAGCCAA Table 1The specific primer for qPCRGeneForward (5′-3′)Reverse (5′-3′)VWA1GAGAAGCACCTGCACTTTGTGGTGGACGTGATCTCCGTGGCATGDUSP6CTCGGATCACTGGAGCCAAAACGTCACAGTGACTGAGCGGCTAALOXL1ACAGCACCTGTGACTTCGGCAACGGTTATGTCGATCCACTGGCAIGFBP4ACCCACGAGGACCTCTACATCACACACCAGCACTTGCCACGCTCYGBACCGCTGCCTACAAGGAAGTGGAAGGAGGGTCTTCAGAACTCGGZIC3TGCATACCTCGGACAAGCCCTAGATTCATAGCCTGAACTGGCAGCGAPDHGTCTCCTCTGACTTCAACAGCGACCACCCTGTTGCTGTAGCCAA The specific primer for qPCR Cell proliferation, migration, and invasion assays For CCK‑8 assay, U251 and U87 cells were seeded at 5 × 103 cells/well into 96-well plates containing 100 μL medium. After 24, 48, and 72 h of culture, 10 μL of CCK-8 reagent (C0037, Beyotime) was added into each well at 37 °C for 30 min. Absorbance was detected at 450 nm [29]. For the migration assay, U251 and U87 cells (3 × 104 cells/well) were seeded into 6-well plates and cultured at 37 °C with 5% CO₂ until 90% confluence. A sterile pipette tip was used to create a scratch, and after washing with PBS, cells were imaged at 0 and 48 h to monitor migration [30]. For invasion assays, upper Transwell chamber was coated with Matrigel and placed at 37 °C for 30 min and then added with cell suspension (100 μL, 3 × 104 cells), while the lower chamber contained serum-containing medium. After 48 h, invaded cells were fixed by 4% paraformaldehyde for 15 min, stained by crystal violet for 10 min, and counted with a microscope from at least five random fields [31]. Statistical analysis Statistical analyses on bioinformatic data were conducted using R language (version 3.6.0) and Sangerbox 2 Platform [32]. Wilcoxon rank-sum test was employed to compare two-group differences of continuous variables, whereas log-rank test was utilized to compare the survival among different patient groups [23]. For the experimental data, GraphPad Prism software (version 8.0.2) was applied, and the comparison on the data was accompolished via unpaired t test and one-way/two-way analysis of variance. A p < 0.05 stood for statistical significance. Results Single-cell landscape of the cell types in GBM samples The scRNA-seq analysis of the GSE273274 cohort revealed nine distinct cell subpopulations: astrocytes, microglial cells, neurons, macrophages, oligodendrocytes, progenitor cells, fibroblasts, endothelial cells, and T cells (Fig. 1A). Key marker genes for each type were identified (Fig. 1B), including high-expressed CA12, GPC5, ACSBG1 and ALDH1L1 in astrocytes, and FTL and APOE in microglial cells. Cell proportion analysis showed that fibroblasts were more abundant in the tumor group (Fig. 1C) and that astrocytes represented the most abundant population among all cell types (Fig. 1D). Further characterization showed fibroblasts did not express typical CAF marker genes (e.g., ACTA2, PDGFRB, FAP) but expressed several cancer stem cell-related marker genes (e.g., ITGA6 and VIM) (Figure S1). This suggested that these cells may possess certain tumor-associated properties rather than representing normal stromal fibroblasts, and may be more relevant to the microenvironment of central nervous system.Fig. 1Single-cell landscape of the GBM and control samples. A UMAP plot of the identified cell types. B Bubble plot of marker genes expression in each cell type. C The proportion of cell types in tumor and adjacent non-tumor tissue. D The cell proportion in all samples Single-cell landscape of the GBM and control samples. A UMAP plot of the identified cell types. B Bubble plot of marker genes expression in each cell type. C The proportion of cell types in tumor and adjacent non-tumor tissue. D The cell proportion in all samples Specific signature analysis of fibroblasts Differential gene analysis revealed distinct gene expression patterns in each cell type, with neurons and oligodendrocytes showing a higher number of significantly upregulated genes (Fig. 2A). Further KEGG pathway analysis demonstrated that fibroblast-specific DEGs were chiefly implicated in pathways related to tumor occurrence and cytoskeletal remodeling, such as proteoglycans in cancer and regulation of actin cytoskeleton (Fig. 2B). In GO enrichment analysis, these genes were mainly enriched in BP term including regulation of protein complex assembly, response to steroid hormone, and positive regulation of neurogenesis. In CC term, these genes were primarily enriched in structures including presynapse, neuronal cell body, and synaptic membrane, and enriched MF term included enzyme activator activity, ubiquitin protein ligase binding, and ubiquitin-like protein ligase binding (Fig. 2B).Fig. 2Specific marker genes analysis of Fibroblast. A Distribution of differentially expressed genes of each cell type. B Function enrichment analysis of the fibroblast marker genes Specific marker genes analysis of Fibroblast. A Distribution of differentially expressed genes of each cell type. B Function enrichment analysis of the fibroblast marker genes WGCNA for fibroblast-related gene module Fibroblast score for each patient was computed applying WGCNA to identify co‑expression modules associated with this score. The soft threshold β = 10 was selected to ensure a scale-free network topology (Fig. 3A). Hierarchical clustering classified 18 co-expression modules after merging, with the gray module containing unassignable genes (Fig. 3B). Correlation analysis showed that the cyan module was notably positively linked to the fibroblast score (Fig. 3C, Fig. 3D), and that the genes in this module were also positively correlated with the fibroblast score of GBM patients (Fig. 3E), suggesting this module may play an imperative role in regulating fibroblast cell function.Fig. 3WGCNA for the fibroblast-related gene module. A Scale-free fitting index analysis of soft-threshold power (β). B Gene tree plot based on dissimilarity measure (1-TOM) clustering. C Correlation between the module feature of each module and Fibroblast. D The number of genes in each module. E Correlation between cyan Module gene Membership and Fibroblast WGCNA for the fibroblast-related gene module. A Scale-free fitting index analysis of soft-threshold power (β). B Gene tree plot based on dissimilarity measure (1-TOM) clustering. C Correlation between the module feature of each module and Fibroblast. D The number of genes in each module. E Correlation between cyan Module gene Membership and Fibroblast Establishment and validation of a prognostic riskscore Significant prognostic genes from the cyan module were selected using univariate Cox regression analysis followed by LASSO regression. As the penalty parameter λ increased, the coefficients of most genes shrank toward 0. Ten‑fold cross‑validation indicated optimal predictive performance at λ = 0.0085 (lambda.min), which retained six non-zero feature genes to construct the RiskScore model (Fig. 4A, B). Multivariate Cox regression analysis was applied to formulate the RiskScore =\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text{0.085*VWA1+ 0.131*DUSP6+ 0.124*LOXL1+ 0.119*IGFBP4+ 0.1*CYGB}+ \left(\text{-0.392}\right)\text{*ZIC3}$$\end{document}0.085*VWA1+ 0.131*DUSP6+ 0.124*LOXL1+ 0.119*IGFBP4+ 0.1*CYGB+-0.392*ZIC3 (Fig. 4C). Using the optimal RiskScore cutoff, patients were stratified into high- and low-risk groups. The model showed strong discriminative ability for 1- and 2-year survival prediction (AUC = 0.72 and 0.69, respectively; Fig. 4D). High-risk patients exhibited significantly poorer outcomes (Fig. 4E). Except for ZIC3, other five genes were upregulated in the high-risk group in TCGA-GBM cohort (Fig. 4F). External validation in the CGGA_325 cohort applying the same method also confirmed the model’s robustness, showing consistent predictive performance in both the training set (TCGA-GBM) and validation set (CGGA_325) (Fig. 4G) and notably worse survival in high-risk group (Fig. 4H, I).Fig. 4Establishment and validation of RiskScore model. A–B LASSO coefficient path and LASSO coefficient path plot. C Coefficients distribution of model genes. D ROC analysis of 1- and 2-year survival rate in TCGA-GBM cohort. E KM survival analysis of patients in TCGA-GBM cohort. F The expression difference of model gene in high- and low-risk groups. G–H ROC analysis and KM survival analysis in CGGA_325 cohort. I The survival status of the high and low-risk groups in CGGA_325 cohort. (**p < 0.01, ****p < 0.0001) Establishment and validation of RiskScore model. A–B LASSO coefficient path and LASSO coefficient path plot. C Coefficients distribution of model genes. D ROC analysis of 1- and 2-year survival rate in TCGA-GBM cohort. E KM survival analysis of patients in TCGA-GBM cohort. F The expression difference of model gene in high- and low-risk groups. G–H ROC analysis and KM survival analysis in CGGA_325 cohort. I The survival status of the high and low-risk groups in CGGA_325 cohort. (**p < 0.01, ****p < 0.0001) The relationship between the riskscore and immune infiltration To investigate the relationship between the RiskScore and immune cell infiltration, we employed multiple computational approaches. First, ESTIMATE algorithm showed that high-risk group had markedly higher StromalScore (Fig. 5A). Using the TIMER algorithm, the RiskScore and the scores of neutrophils and B_cells were found to be negatively correlated (Fig. 5B), indicating lower immune infiltration of neutrophils and B_cells in high-risk group. Finally, the MCPcounter analysis further revealed notable positive associations between the RiskScore and the scores of myeloid dendritic cells and fibroblasts (Fig. 5C).Fig. 5Difference analysis of immune infiltration in high and low-risk groups. A ESTIMATE algorithm for StromalScore, ImmuneScore, ESTIMATEScore difference in TCGA-GBM cohort. B Correlation heatmap between RiskScore and 6 types of immune cell infiltration score. C Correlation between RiskScore and MCPcounter of immune infiltration score (nsp > 0.05, *p < 0.05, **p < 0.01, ***p < 0.001, ****p < 0.0001) Difference analysis of immune infiltration in high and low-risk groups. A ESTIMATE algorithm for StromalScore, ImmuneScore, ESTIMATEScore difference in TCGA-GBM cohort. B Correlation heatmap between RiskScore and 6 types of immune cell infiltration score. C Correlation between RiskScore and MCPcounter of immune infiltration score (nsp > 0.05, *p < 0.05, **p < 0.01, ***p < 0.001, ****p < 0.0001) Drug sensitivity analysis for different risk groups Drug sensitivity prediction showed that high-risk group had significantly increased sensitivity to Cisplatin, MG-132, AZ628, Dasatinib, CGP-60474, A-770041, TGX221, and Bortezomib (p-value < 0.05), suggesting potential therapeutic effects of these drugs for high-risk GBM patients (Fig. 6). In contrast, low‑risk patients were more responsive to VX-680, CMK, KIN001-135, and QS11 (Fig. 6), reflecting underlying molecular and immune microenvironment differences. These findings from drug sensitivity predictions may contribute to personalized chemotherapy regimens for GBM patients in different risk groups and improvement of their survival.Fig. 6Drug sensitivity to IC50 in high- and low-risk groups. (*p < 0.05, ** p < 0.01, *** p < 0.001) Drug sensitivity to IC50 in high- and low-risk groups. (*p < 0.05, ** p < 0.01, *** p < 0.001) VWA1 promoted migration and proliferation of GBM cells The qPCR showed that the mRNA levels of VWA1, DUSP6, LOXL1, IGFBP4, and CYGB were significantly higher in the GBM cell lines (U251 and U87) than in HEB cells, while ZIC3 was downregulated in the GBM cell lines (Fig. 7A). We further constructed the knock-down plasmids of VWA1 in GBM cell line. It was found that both si-VWA1#1 and si-VWA1#2 significantly lowered mRNA expression of VWA1, with si-VWA1#1 showing a higher efficiency (Fig. 7B, C). Next, CCK-8 assay showed that the proliferation of both U251 and U87 cells treated with si-VWA1#1 was significantly inhibited when compared to the negative control (si-NC) across all the time points measured (24 h, 48 h, and 72 h) (Fig. 7D, E). In addition, the knock-down of VWA1 significantly reduced the migration and proliferation in U251 and U87 GBM cells. In the wound healing assay, si-VWA1#1 cell showed a lower wound closure than the si-NC group at 48 h (Fig. 7F, G). Similarly, the Transwell assay also demonstrated suppressed invasion of si-VWA1#1-transfected GBM cells, indicating reduced invasion (Fig. 7H, I). These results suggested that VWA1 promoted GBM cell migration and proliferation, while its suppression can effectively inhibit these processes in GBM cells.Fig. 7Gene expression and function detection. A qPCR for the gene expression in GBM cells (U251 and U87) and HEB cells. B–C qPCR for the expression of VWA1 in cell with gene knock-down. D–E CCK-8 for the cell proliferation detection. F–I Wound healing and Transwell tests for the cell migration and invasion. (*p < 0.05, **p < 0.01, ***p < 0.001, ****p < 0.0001) Gene expression and function detection. A qPCR for the gene expression in GBM cells (U251 and U87) and HEB cells. B–C qPCR for the expression of VWA1 in cell with gene knock-down. D–E CCK-8 for the cell proliferation detection. F–I Wound healing and Transwell tests for the cell migration and invasion. (*p < 0.05, **p < 0.01, ***p < 0.001, ****p < 0.0001) Discussion GBM is a highly aggressive and heterogeneous primary brain tumor with a poor prognosis [33]. Despite current standard treatments such as surgery and chemotherapy, tumor recurrence remains common [34], underscoring the need to better understand the molecular mechanism that drivers GBM progression and treatment resistance [35]. In GBM, CAFs and CAF‑like cells promote tumor aggressiveness by forming perivascular niches and supporting stem cell proliferation [36]. Through secretion of production of growth factors and extracellular matrix (ECM) proteins, CAFs play a key role in GBM progression [37]. Thus, a deeper understanding of CAF biology may therefore help elucidate the mechanisms underlying GBM pathogenesis and chemoresistance, ultimately advancing personalized therapy. This study characterized the single-cell landscape of GBM and developed a fibroblast-related RiskScore model to improve the personalized treatment of GBM. The single-cell landscape showed that astrocytes had the highest proportion in the tumor tissue and the fibroblast proportion was increased in the GBM group. Previous studies demonstrated that astrocytes, which account for 50% of all brain cells, can direct contact with and activated by GBM cells to form tumor-associated astrocytes (TAAs), which promote GBM cell growth and invasion in healthy tissue and limit the radiation or chemotherapy response [38]. Tumor-associated mesenchymal stem/stromal cells (TA-MSCs) can be co‑opted by GBM for fibroblast differentiation, enhancing the tumor growth and invasion [39]. Functional enrichment analysis of fibroblast‑specific markers indicated the involvement of these cells in regulation of actin cytoskeleton, positive regulation of neurogenesis, neuronal cell body, and ubiquitin protein ligase binding process. The regulation of actin cytoskeleton in fibroblasts are crucial for cell migration [40]. Fibroblasts producing fibroblast growth factor-2 (FGF-2) not only stimulate neural cell growth but also promote the development of GBM stem cells (GBM-SCs) [41]. These findings underscored the imperative role of fibroblasts in GBM tumorigenesis and migration, supporting their significance as a potential biomarker for GBM. In the RiskScore model, Zinc finger of the cerebellum 3 (ZIC3) was identified as a protective factor, while others genes contributed to a higher risk of GBM. This result was consistent with the expression trends of the six genes detected by qRT-PCR. ZIC3, a C2H2 zinc finger transcription factor, is involved in the formation of early vertebrate embryonic morphology [42] and can induce the transformation of human fibroblasts into stable neural progenitor-like cells [43]. Previous study also reported the correlation between high-expressed ZIC3 and a favorable prognosis in GBM [44]. As a member of the Gli (glioma-associated oncogene homolog) superfamily, ZIC3 can influence the Shh signaling pathway through regulating Gli3 activity, which may also be involved in the pathogenesis of GBM [45], though further research is still needed. Von Willebrand factor A domain containing 1 (VWA1), an ECM protein expressed in muscle and peripheral nerves [46], has been associated with aggravated neuromyopathy progress upon mutation [47] and contributes to the structural stability and integrity of the ECM [48]. Zuchegna et al., revealed that the overexpression of dual-specificity phosphatase (DUSP6) promoted invasion of GBM in vitro [49]. Lysyl oxidase enzymes-like 1 (LOXL1) is a crucial regulator for ECM in remodeling TME, the knock-down of LOXL1 can suppress cell proliferation and invasion in GBM by downregulating EMT pathway [50]. Meanwhile, upregulated Insulin-like growth factor binding protein 4 (IGFBP4) is associated with the OS of GBM patients [51]. Cytoglobin (CYGB) plays an important role in ECM remodeling, fibrosis, and tissue damage repair [52], and its expression can serve as an independent prognostic factor for GBM, showing a correlation with patient survival [53]. In conclusion, all the six genes were involved in the pathogenesis of GBM. The involvement of multiple genes in the ECM and their association with CAF‑related functions suggested that their coordinated action might be related to ECM regulation, though the specific regulatory mechanisms still demanded further experimental validation. Furthermore, ESTIMATE analysis revealed significantly high StromalScore in high-risk patients, which may be attributed to the upregulation of most of the model genes involved in ECM remodeling. We also found that the RiskScore was significantly negatively correlated with the neutrophil cells and B_cell infiltration score. Given that neutrophil cells can efficiently traverse physiological barriers for anti-tumor response [54], the result suggested that the TME of high-risk GBM patients was markedly immunosuppressive, potentially facilitating immune escape and limiting immunotherapy response. However, fibroblasts were closely positively correlated with the RiskScore, suggesting the critical role of these genes in fibroblasts in regulating the invasion and proliferation of GBM cells. Studies have found that CAFs are closely involved in intercellular interactions within the TME and can promote TME remodeling into an ECM rich in type I collagen, which is closely associated with brain metastases [55]. The development of GBM induces profound changes in ECM components, resulting in elevated levels of collagen type I [56]. As a crucial immunomodulatory component of the TME, the ECM can be harnessed to enhance the efficacy of immunotherapy [57]. Combined with the close correlation between the six genes and the ECM, these results suggested that the RiskScore model may capture key mechanisms mediated by ECM underlying GBM progression. Finally, Cisplatin, MG-132, AZ628, Dasatinib, CGP-60474, A-770041, TGX221, and Bortezomib were predicted as potential therapeutic agents for high‑risk patients, potentially improving the personalized treatment and prognosis of GBM patients. Micellar formulations of Dasatinib have demonstrated promising anti‑tumor activity in GBM models in vitro and in vivo [58]. Although combined therapy of Temozolomide and Cisplatin has exhibited therapeutic potential in clinical trials, its clinical efficacy is limited by blood–brain barrier (BBB) [59]. Studies reported that pretreatment of GBM with Bortezomib can enhance the drug sensitivity of GBM to Temozolomide [60]. In summary, though several of the drugs have exhibited clinical promise in GBM treatment, challenges such as poor BBB penetration still remain. Further clinical studies are needed to evaluate their combined use with immunotherapy and to validate their utility using clinical trials. Several limitations of this study should be equally recognized. Although we characterized the single-cell landscape of GBM, limited sample size and technical constraints may lead to under‑representation of certain cell types or states. Expanding the sample size to include multi-center, multi-subtype samples is necessary to enhance the reliability of the results. Furthermore, in identifying fibroblast signature genes, it is essential to consider the functional heterogeneity of fibroblasts and how the identified genes reflected specific fibroblast subpopulations. The biological roles of the model genes also required further verification from in vitro and in vivo assays, for example, establishing GBM mouse models and performing experiments such as Western blotting (WB). Additionally, clinical trials are also need to evaluate the clinical translation and non‑invasive monitoring potential of these biomarkers. Such efforts could include enrolling larger patient cohorts, detecting biomarker expression in peripheral blood, and combining imaging technologies to further investigate their clinical reliability. Conclusion This study integrated scRNA-seq analysis and WGCNA to identify six fibroblast-specific genes (ZIC3, VWA1, DUSP6, LOXL1, IGFBP4, and CYGB) for GBM and developed a RiskScore model capable of distinguishing high- and low-risk GBM patients. The RiskScore can effectively predict the survival outcomes and uncover differences in immune cell infiltration and chemotherapeutic sensitivity between the two risk groups. Given the close association of these core genes with the ECM, together with the results of immune analysis that the ECM functioned as a key immunomodulatory component of the TME, this study further emphasized the clinical translational potential of the RiskScore for GBM treatment and provided a theoretical basis for the development of combined therapeutic strategies targeting ECM-related pathways in GBM. Supplementary Information Additional file 1. Figure S1 The expression profiles of ITGA6 and VIM in fibroblastsAdditional file 2 Additional file 1. Figure S1 The expression profiles of ITGA6 and VIM in fibroblasts Additional file 2