2.1 Data Collection and Processing

Details regarding data acquisition and preprocessing are provided in Data S1.

2.2 Identification of CHI3L1-Related Vascular Signature Genes

Glioma vascular genes were identified based on the differential expression patterns observed between the vasculature of gliomas and normal blood vessels [21, 22]. Patients with gliomas were divided into two groups based on the median value of CHI3L1 expression. Differentially expressed genes (DEGs) between the two groups were analyzed using the DESeq2 R package. The cut-off values for DEGs were set at a log₂ fold change greater than two and a false discovery rate (p adj) < 0.05. The intersection of DEGs and vascular genes was used to perform a prognostic analysis using univariate Cox regression analysis. Robustness of the identified genes was evaluated using bootstrapping. Subsequently, two machine learning algorithms, the least absolute shrinkage and selection operator (LASSO) algorithm and random survival forest (RSF) algorithm, were used to further identify robust prognostic signatures. The RSF analysis was repeated 1000 times independently, and genes with a relative importance greater than 0.75 were selected using the “randomForestSRC” R package [23]. The LASSO algorithm was conducted with 10-fold cross-validation using the glmnet R package [24]. Genes identified by both the RSF and LASSO algorithms were considered to constitute the CHI3L1-associated vascular signature (CAVS).

2.3 Establishment of Vascular-Related Risk (VR) Score

A VR score scheme was constructed based on CAVS to quantify the vascular-related phenotypes of individual patients, in accordance with the methodology outlined in a previous study [25]. Initially, an unsupervised clustering approach was used to categorize the patients into distinct clusters. Subsequently, CAVS expression levels, which showed positive or negative correlations with these clusters (Clusters A, B, and C), were categorized as vascular gene signatures A and B, respectively. The Boruta algorithm was employed to reduce the dimensionality of vascular gene signatures A and B. Subsequently, the first principal component (PC1) was extracted using PCA and served as the signature score. The VR score was defined as follows: VR score = ∑PC1A–∑PC1B.

2.4 Statistical Analysis

Statistical significance was defined as a two-sided p value of less than 0.05. All data processing and analyses were conducted using R software (version 4.3.1).

See (Data S1) for detailed bioinformatics methods and experimental methods descriptions.

3.1 Identification of the CHI3L1-Associated Vascular Signature Genes

We evaluated the relationship between CHI3L1 expression and glioma progression in the TCGA (n = 647) and CGGA (n = 970) cohorts and found that increased CHI3L1 levels were correlated with higher glioma grades (Figure S1A). Survival analysis showed that patients with high CHI3L1 expression had significantly poorer OS compared to those with low expression (p < 0.001, Figure S1B). Multivariate Cox regression identified CHI3L1 as an independent predictor of poor survival outcomes (Figure 1A). Consistent with the enrichment of the angiogenesis pathway (Figure S1C), patients with high CHI3L1 expression had a greater abundance of endothelial cells (Figure 1B). Furthermore, high CHI3L1 expression was significantly associated with key angiogenic signatures, including VEGFA, PIGF, ANGPT1, ANGPT2, and FGF1 (Figure S1D). Immunohistochemistry (IHC) and H&E staining further confirmed that elevated CHI3L1 expression was associated with increased tumor vascular density in glioma tissues (Figures 1C and S1E).

To identify CHI3L1-associated angiogenesis genes, we analyzed 798 differentially expressed genes (DEGs) between high- and low-CHI3L1 expression groups (Table S1). From the literature [21, 22], we compiled 456 glioma vascular genes (Table S2) and identified 30 overlapping genes between the two datasets (Figure S1F). Among these, 23 genes were significantly associated with prognosis (p < 0.001, Figure S1G). Using LASSO and RSF algorithms, we further refined this set, with LASSO identifying 20 genes (Figures 1D and S1H) and RSF highlighting genes with a relative importance score > 0.75 (Figure 1E). Both analyses converged on three key genes: TAGLN2, GBP1, and TIMP1, forming the CHI3L1-associated vascular signature (CAVS). Correlation analysis revealed a strong association between these CAVS genes and CHI3L1 expression (R > 0.7, Figure S1I). Furthermore, in vitro studies demonstrated that CHI3L1 knockdown significantly reduced the expression of TIMP1, TAGLN2, and GBP1 in U251 glioma cell lines (Figure S1J). Immunohistochemical validation confirmed that TAGLN2, GBP1, and TIMP1 protein levels were significantly higher in high-grade glioma tissues compared to low-grade gliomas (Figure 1F).

3.2 Evaluation of Correlation Between CAVS and Single-Cell Characteristics in Glioma

We analyzed CAVS expression at the single-cell transcriptome level to explore its role in different cellular contexts. A total of 24,800 cells and 23 cell clusters were identified from seven cell types in single-cell RNA sequencing profiles (Figure 2A). Malignant glioma cells were detected using inferCNV analysis to calculate CNV scores (Figure S2A). Gene expression profiles revealed specific marker genes in each cell cluster (Figure S2B), with CAVS expression patterns shown across different cell types (Figure 2B). Notably, fibroblasts exhibited significant expression of α-SMA (ACTA2), a hallmark of cancer-associated fibroblasts (CAFs), indicating their transition into a more aggressive phenotype within the TME. Using the AUCell R package [26], we estimated CAVS expression levels, finding high-AUC cells predominantly in MES glioma cells (Clusters 3 and 8), tumor-associated macrophages (TAMs, Clusters 6 and 11), astrocytes (Cluster 19), CAFs (Cluster 18), and endothelial cells (ECs, Cluster 22) (Figure 2C).

To further investigate CAVS function, we examined the biological processes of high- and low-CAVS clusters. GO and KEGG analyses revealed that TAMs with high CAVS expression had diminished immune functions, including antigen processing and presentation, compared to low-CAVS TAMs (Figure 2D). Phagocytic gene signature analysis showed that low-CAVS TAMs had higher phagocytosis scores than high-CAVS TAMs (Figure 2D). Pathway analysis indicated that high-CAVS glioma cells, TAMs, fibroblasts, and ECs displayed significant tumor-promoting activity, including tumor angiogenesis (Figures 2E and S2C). CellChat analysis was used to explore cell–cell interactions mediated by ligand-receptor signaling, revealing that glioma cells, TAMs, and fibroblasts regulate ECs through angiogenesis-related pathways, including VEGF, NOTCH, FGF, ANGPT, CD99, and CALCR. The high-CAVS group exhibited significantly stronger interactions with ECs compared to the low-CAVS group (Figures 2F and S2D,E).

3.3 Identification of Vascular-Related Molecular Phenotypes at the Bulk Transcriptome Level

To identify vascular-related phenotypes, unsupervised clustering was performed based on CAVS expression, resulting in the classification of patients into three clusters (A, B, and C). t-SNE analysis confirmed the effective separation of these subtypes (Figure 3A). Survival analysis revealed significant prognostic differences among the clusters, with Cluster A having the worst overall survival (OS) and Cluster C showing the best prognosis (p < 0.001; Figure 3B).

The correlation between the expression of CAVS and clinicopathological characteristics was also examined. Cluster A was characterized by a high prevalence of IDH1 wild-type status, elevated CHI3L1 expression, and older patient age (Figure 3C). Additionally, Cluster A exhibited an angiogenic phenotype, marked by angiogenic sprouting (Figure 3D), endothelial cell formation, and endothelial tip cell differentiation (Figure 3E). In contrast, Cluster C displayed a quiescent endothelial profile, featuring low-permeability vasculature and enrichment of blood–brain barrier (BBB) components [27] (Figure 3F).

3.4 Characteristics of Tumor Immune Microenvironment (TIME) in Distinct Phenotypes

We investigated the underlying biological behaviors of the three clusters. Cluster A showed significantly elevated activity in stromal and tumorigenic pathways, including epithelial-mesenchymal transition (EMT), angiogenesis, TNF-α, EGFR, MAPK, and hypoxia signaling. Additionally, the infiltration levels of 24 microenvironment cell types were analyzed. Cluster A had markedly increased infiltration of both pro-tumor and anti-tumor cells, such as neutrophils, fibroblasts, CD8+ T cells, DCs, and NK cells, while these levels progressively decreased in Clusters B and C (Figure 4A). Cluster A also exhibited significantly higher immune checkpoint expression compared to Clusters B and C. This immune profile was further validated in meta-cohorts (Figure S3A,B).

To assess anti-cancer immune responses, we analyzed the cancer immunity cycle [28]. Cluster A demonstrated increased activity in Step 1 (release of cancer cell antigens) and Step 4 (trafficking of immune cells to tumors) (Figure 4B). Cluster B showed the highest immune cell infiltration (Step 5), while Cluster C had the greatest capacity for antigen presentation (Step 2), priming and activation (Step 3), and cancer cell death (Step 7). The relative abundance of tumor-infiltrating immune cells was also examined, with DCs being most abundant in Cluster A, T cells in Cluster B, and B cells in Cluster C (Figures 4C,D and S3C).

Solid tumors are classified into three immunological phenotypes based on the status of tumor-infiltrating T cells: immune-excluded, immune-inflamed, and immune-desert. The immune-excluded phenotype, characterized by abundant immune cells located in the stroma rather than the tumor core [29], is observed in Cluster A, which exhibits significant stromal activation, high immune cell infiltration, and low TILs. Cluster B, distinguished by a considerable presence of TILs, likely represents an immune-inflamed phenotype. Cluster C, with minimal immune cell infiltration, was hypothesized to represent an immune-desert phenotype. Diagnostic slides from TCGA samples were analyzed, and three representative whole slide images (WSI) were displayed (Figure 4E). Cluster A patients showed low TILs, Cluster B patients had a high density of TILs, and Cluster C patients exhibited high tumor purity with low TME cell infiltration. These immunological phenotypes for Clusters A, B, and C were consistent with the pathological features observed through multi-color immunofluorescence analysis (Figure 4F).

3.5 Metabolic, Ferroptosis, Senescence, and Genome Alterations Characteristics Among Vascular-Related Phenotypes

To explore the metabolic reprogramming of vascular-related phenotypes within the TME, we assessed enrichment scores for metabolic pathways across phenotypes. Cluster A was positively associated with metabolic pathways promoting tumor growth, including kynurenine and tryptophan metabolism, arachidonic acid metabolism, and cofactor and vitamin metabolism. In contrast, Cluster C showed upregulation of tumor-suppressive pathways, such as butanoate, taurine, and hypotaurine metabolism (Figure 5A). Correlation analyses revealed that the pathways enriched in Cluster A were linked to angiogenesis, immune checkpoint expression, immunosuppressive activity, and poor prognosis (Figure 5B). Cluster A also demonstrated significantly elevated senescence and FPI scores compared to Clusters B and C, suggesting increased levels of senescence and ferroptosis (p < 0.05, Figure 5C,D).

We also examined genomic differences across phenotypes, focusing on mutation signatures and single nucleotide variants (SNVs). Cluster A exhibited a higher tumor mutational burden (TMB) than Clusters B and C (p < 0.001; Figure 5E) and showed increased mutation rates in the age-related signature (SBS1) and base excision repair deficiency signature (SBS30). Among the top 20 mutated genes, IDH1, ATRX, CIC, and NOTCH1 mutations were prevalent in Cluster C, while PTEN mutations were more common in Cluster A (Figure 5F). Notably, Cluster A exhibited amplification of 7p11.2 (EGFR), 12q14.1 (CDK4), deletion of 9p21.3 (CDKN2A/B, MTAP), and PTEN loss (Figure 5F).

3.6 Construction of Vascular-Related Score and Exploration of Its Clinical Relevance

The VR score, derived from three hub genes, was developed to evaluate vascular-related phenotypes in individual patients. In the TCGA cohort, the Kruskal–Wallis test revealed that Cluster A had the highest VR score, followed by Clusters B and C (Figure S4A, p < 0.001). Survival analysis across multiple cohorts demonstrated that patients with higher VR scores had significantly worse prognoses, with subgroup analysis validating its prognostic value across different glioma grades (p < 0.001, Figure 6A). Univariate Cox analysis also indicated that patients with higher VR scores had poorer overall survival (OS) compared to those with elevated VEGFA levels (Table S3). Additionally, the VR score showed a strong correlation with angiogenesis (R > 0.92, Figure S4B).

We validated the VR score's utility by integrating clinical and molecular characteristics. The AUC values for the VR score at 1, 3, and 5 years for OS were 0.87, 0.906, and 0.817, respectively, in the TCGA cohort, with similar results in the CGGA and Rembrandt cohorts (Figures 6B and S4C). Given the critical role of molecular markers such as IDH1 mutation, CDKN2A/B loss, PTEN loss, and EGFR gain in glioma classification, we further evaluated the VR score's predictive ability for these markers. The VR score demonstrated a significant association with these markers, along with high AUC values ranging from 0.874 to 0.967 (Figures 6C,D and S4D–F, p < 0.001). In multiple cohorts, it also showed high accuracy in predicting GBM status (AUC = 0.978, 0.964, and 0.854, Figures 6D and S4G). Multivariate analysis identified the VR score as an independent prognostic factor (Figure 6E, p < 0.001). The nomogram integrating the VR score and clinical factors demonstrated strong predictive accuracy and potential for clinical application (Figures 6F and S4H,I).

3.7 The Correlation Between the VR Score and Immune Status and ICIs Therapy Response

To explore the molecular mechanisms linking the VR score to glioma prognosis, GSVA analysis revealed that the VR score was associated with immune response and angiogenesis pathways, including interferon-alpha response, EMT, and angiogenesis (Figure S5A). Analysis of the cancer-immunity cycle showed significant immune suppression in key processes (steps 2, 3, 5, and 7) within the high-VR-score group (Figure 7A). Correlation analysis further indicated a strong link between the VR score and pro-tumor factors, such as oncogenic pathways, T cell exhaustion, and immune checkpoint expression, suggesting an immunosuppressive status in the high-VR-score group (Figure S5B).

Given the VR score's association with immune response, we investigated its predictive value for ICI therapy. Patients with high VR scores had elevated TIDE scores and a lower likelihood of responding to ICIs (Figures 7B and S5C, p < 0.001). Additionally, the VR score was positively correlated with TIDE scores (Figure 7C; R > 0.3, p < 0.001). Across six immunotherapy cohorts, the VR score's predictive capacity for ICI response was further confirmed. PCA analysis illustrated baseline gene expression profiles before and after batch effect correction (Figure 7D). Patients with low VR scores exhibited superior responses to ICI therapy, demonstrating both enhanced responsiveness (Figure 7E, p = 0.003) and improved survival benefits (Figure 7F, p < 0.001) compared to those with high VR scores. These findings suggest that the VR score may serve as a valuable predictor of immunotherapy response in glioma patients.

3.8 Identification of Potential Therapeutic Agents for High-VR-Score Patients

To identify potential drugs for highly vascularized tumors, we analyzed IC50 values from a pan-cancer cell line dataset (GDSC, 749 samples) to explore the correlation between the VR score and sensitivity to antiangiogenic agents. Cancer cell lines with elevated VR scores showed lower IC50 values for Sorafenib, Lapatinib, Foretinib, and Cediranib, indicating enhanced sensitivity to these drugs (Figure 8A). Moreover, glioma patients with high VR scores demonstrated significantly greater responsiveness to Foretinib compared to those with low VR scores (Figure 8B).

To identify potential therapeutic agents for high-VR-score patients, we performed XSum analysis based on transcription factors (TFs). The top 10 TFs, primarily involved in angiogenesis and EMT pathways, showed differential activity between high- and low-VR-score groups (Table S4, Figure 8C). Further screening of prognosis-related TFs (p < 0.001) identified arachidonyltrifluoromethane, camptothecin, PHA.00816795, tanespimycin, and X5707885 as the top five drugs with the lowest CMap scores, suggesting their potential efficacy in high-VR-score patients (Figure 8D).