Transcriptional profiles and corresponding clinical data for human gliomas were retrieved from three publicly available databases: The Cancer Genome Atlas (TCGA, https://www.cancer.gov/about-nci/organization/ccg/research/structural-genomics/tcga), the Chinese Glioma Genome Atlas (CGGA, http://www.cgga.org.cn/), as well as the REMBRANDT (https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE108474) dataset. The final analytical dataset comprised 702 glioma samples from TCGA, 693 from CGGA, and 475 from REMBRANDT. Additionally, RNA sequencing data from 15 patients with recurrent glioblastoma (rGBM) who received neoadjuvant anti-PD-1 immunotherapy (GSE121810, https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE121810) were retrieved from the Gene Expression Omnibus (GEO) database for external evaluation of the immunotherapy response prediction model. Immunohistochemical staining data for normal brain tissue and glioma specimens were acquired from the Human Protein Atlas (https://www.proteinatlas.org/). Glioma subtypes were defined according to the World Health Organization (WHO) Classification of Tumors of the Central Nervous System (4th edition).

The Kaplan-Meier approach was applied to generate survival estimates, while the log-rank test served to evaluate prognostic disparities among the subgroups.

Somatic mutation data for gliomas were retrieved from TCGA and analyzed using the R package maftools (version 2.20)^[7]. The mutational landscape of the 20 most frequently altered genes, together with IL-32, was profiled and compared between gliomas stratified by IL-32 expression levels.

Stromal and immune scores were computed for glioma samples utilizing the R package ESTIMATE (version 1.0.13) and correlated with IL-32 expression to characterize the TME. Immune cell composition was inferred from TCGA, CGGA, and REMBRANDT datasets using CIBERSORTx^[8], which deconvolutes bulk transcriptomic data based on a reference of 22 human immune cell types. This approach enabled quantitative assessment of immune cell proportions in individual gliomas, helping to dissect the link between IL-32 expression and immune infiltration patterns.

Glioma samples were stratified into two groups according to IL-32 expression: high and low. Differentially expressed genes (DEGs) between the two groups were identified utilizing the R package limma (version 3.60.6), with thresholds of |log₂ fold change| > 1 and adjusted P-value < 0.05. A total of 51 DEGs were consistently detected across TCGA, CGGA, and REMBRANDT cohorts, which were then analyzed for their biological significance through functional enrichment.

Functional enrichment of the identified DEGs was performed using Gene Ontology (GO) and Kyoto Encyclopedia of Genes and Genomes (KEGG) pathway analyses via the R package clusterProfiler (version 4.12.6). Significant enrichment in GO and KEGG analyses was determined by a corrected P-value < 0.05.

R package ConsensusClusterPlus (version 1.68.0) was applied to the 51-DEG expression matrix, assigning each glioma specimen to one of two newly defined molecular subgroups. Cluster stability and the optimal number of clusters (K) were determined via the cumulative distribution function (CDF) plot. Principal component analysis (PCA) was subsequently performed using the R package FactoMineR (version 2.11) to visualize variance captured by the principal components and validate subtype separation. Differential expression patterns across subtypes were illustrated with a heatmap of the 51 DEGs, confirming their biological distinctiveness. Overall survival (OS) analyses were then conducted to determine whether these subtypes hold distinct clinical significance.

We developed a prognostic risk signature based on the Least Absolute Shrinkage and Selection Operator (LASSO) Cox regression algorithm via the R package glmnet (version 4.1.8). LASSO identified ten genes with non-zero coefficients from the 51 DEGs, which were used to compute a risk score for each sample as follows: Risk score = $$ {\textstyle \sum_{\mathrm{i}}^{\mathrm{n}}} \mathrm{Coef \ i} × \mathrm{Expression \ i} $$. Using the median risk score as the cutoff, the TCGA glioma cohort was split into HR and low-risk (LR) groups. The independent prognostic value of the risk model was further evaluated using multivariate Cox proportional hazards regression incorporating established clinical and molecular prognostic factors, including age, WHO grade, isocitrate dehydrogenase (IDH) mutation status, 1p/19q co-deletion status, and O6-methylguanine-DNA methyltransferase (MGMT) promoter methylation status when available. The incremental predictive performance of the risk score was assessed by comparing time-dependent receiver operating characteristic (ROC) curves between the risk model and conventional clinical variables. Furthermore, decision curve analysis (DCA) was performed using the R package rmda (version 1.6) to evaluate the net clinical benefit of applying the risk model compared with traditional prognostic factors across a range of threshold probabilities. For external validation, the model was applied to the CGGA and REMBRANDT cohorts, demonstrating its robustness and clinical applicability.

Immune escape potential in glioma was evaluated using the tumor immune dysfunction and exclusion (TIDE) algorithm^[9], which predicts tumor immune evasion based on gene expression profiles. Glioma transcriptomic data were uploaded to the TIDE web portal (http://tide.dfci.harvard.edu/) to calculate immune-related parameters (microsatellite instability scores, PD-L1 expression, as well as myeloid-derived suppressor cell levels), enabling the estimation of immune checkpoint inhibitor (ICI) responsiveness. In addition, TIDE scores were compared between HR and LR groups using unpaired t-tests.

Next, we examined the risk model’s relationship to immunotherapy benefit by profiling RNA-seq count data from 15 recurrent GBM patients who received neoadjuvant anti-PD-1 therapy^[10]. Patients were classified as responders [overall survival (OS) > 200 days] or non-responders (OS < 200 days), and risk scores were compared between subgroups utilizing unpaired t-tests (P < 0.05).

Drug sensitivity for individual glioma patients was predicted using the calcPhenotype function from the R package oncoPredict (version 1.2)^[11], which implements a ridge regression model trained on pharmacogenomic datasets from the Genomics of Drug Sensitivity in Cancer (GDSC1, GDSC2) and the Cancer Therapeutics Response Portal (CTRP v2). The Pearson correlation between each drug’s half-maximal inhibitory concentration (IC50) and the calculated risk score was computed, with |r| > 0.50 considered indicative of a strong association. To ensure robustness, drugs demonstrating significant correlations across the TCGA, CGGA, and REMBRANDT glioma cohorts were retained as a consensus set of candidate therapeutics.

The U87 and LN229 GBM cell lines, sourced from Procell (Wuhan, China), were maintained in Dulbecco’s Modified Eagle Medium (DMEM, Gibco, USA) supplemented with 10% fetal bovine serum (FBS, ScienCell, USA) at 37 °C under humidified conditions with 5% CO₂. U87 and LN229 cells were maintained at passages 5-10, respectively. Both cell lines were authenticated by short tandem repeat (STR) profiling (Procell, Wuhan, China) and tested negative for mycoplasma using Mycoplasma Detection Kit (Beyotime, China).

Transfections were performed using Lipofectamine 3000 (Invitrogen, USA) following the manufacturer’s protocols. The reagents used in this study include siRNA (s17656, s17657, AM4611, Thermo Fisher Scientific, USA). The compounds used for IL-32 overexpression were purchased from OriGene (OR, USA). The lyophilized protein was reconstituted in sterile phosphate-buffered saline (PBS) containing 0.1% bovine serum albumin (BSA) at a concentration of 100 μg/mL as per the manufacturer’s instructions, and aliquots were stored at -80 °C until use.

The BeyoClick EdU kit was applied to assess cellular proliferation (Cat. C0075L, Beyotime, Shanghai, China) following the manufacturer’s protocol. A 20 mM working solution of Alexa Fluor 555-EdU was prepared, and cells were incubated with EdU for 1 h. Cells were then fixed with 4% paraformaldehyde for 20 min and washed three times with PBS containing 3% BSA. Permeabilization was performed with 0.1% Triton X-100 (Cat. P0096-100mL, Beyotime, Shanghai, China) for 20 min, followed by incubation with the Click-iT EdU reaction mixture for 30 min. Nuclei were counterstained with Hoechst 33342 for 5 min in the dark. Finally, images were captured using a fluorescence microscope (Olympus, Japan). The proliferation index was calculated as (EdU⁺ nuclei/total nuclei) × 100%. Three independent experiments were performed on different days with different cell passages (passages 5-10).

Transwell migration assays were performed by placing 3 × 10⁴ cells (in serum-free DMEM) into the upper insert, using 600 μL of complete medium in the lower well as the migratory stimulus. After 48 h of incubation, non-migratory cells were removed, and the inserts were washed, fixed with methanol for 20 min, and stained with 1% crystal violet for 30 min. For invasion assays, the upper chambers were pre-coated with Matrigel (Cat. 356255, BD, USA) to mimic the extracellular matrix. Migrated or invaded cells were visualized under a light microscope (Olympus, Japan), and three random fields per insert were counted using ImageJ software.

All data analyses were carried out utilizing R (version 4.4.1) and appropriate R packages as described above. Statistical significance was defined as follows: * P < 0.05, ** P < 0.01, *** P < 0.001, and **** P < 0.0001. All analyses adhered to established guidelines to ensure rigor, reproducibility, and reliability.

This study was based on publicly available and de-identified datasets from the TCGA, CGGA, REMBRANDT, and GEO databases, including GSE121810. No new human participants were recruited and no additional human specimens were collected for this study. The original studies from which these datasets were derived obtained relevant ethical approvals and informed consent. Therefore, no additional ethical approval or informed consent was required for the present secondary analysis.

To explore IL-32’s function in glioma, we initially compared its mRNA levels between tumor and normal brain tissues. Relative to normal brain, IL-32 was significantly upregulated in both low-grade gliomas and GBM [Figure 1A]. Besides, its expression progressively increased with glioma WHO grade across the TCGA, CGGA, and REMBRANDT cohorts [Figure 1B]. At the protein level, immunohistochemistry confirmed the highest IL-32 expression in GBM and minimal expression in normal cerebral cortex [Figure 1C], suggesting a correlation between IL-32 overexpression and tumor malignancy. Survival analyses demonstrated that elevated IL-32 expression was consistently linked to a worse prognosis across the three glioma cohorts [Figure 1D], supporting its potential as a prognostic biomarker. Mutation profiling of the 20 most frequently altered genes and IL-32 revealed no detectable somatic mutations in IL-32 across the TCGA glioma cohort [Figure 1E]. This finding suggests that the clinical and biological significance of IL-32 in glioma is more likely driven by transcriptional dysregulation than by recurrent genomic alterations. Together with the consistent overexpression of IL-32 across glioma grades and its association with poor prognosis, these results indicate that aberrant IL-32 expression, rather than mutation, may underlie its contribution to glioma progression. The mutation profiles are summarized in Supplementary Figure 1.

To explore the biological role of IL-32 in glioma, we examined its expression across GBM subtypes in three independent cohorts. The mesenchymal subtype exhibited the highest IL-32 expression [Figure 2A], a subtype characterized by elevated tumor-associated macrophage (TAM) infiltration and an immunosuppressive TME^[12]. Consistently, higher IL-32 expression correlated positively with increased stromal and immune scores across all cohorts [Figure 2B], indicating a strong association with the immune modulation within TME. Using CIBERSORTx to deconvolute 22 immune cell types, IL-32 expression showed the strongest positive correlation with M2-type macrophages [Figure 2C], which are implicated in tumor progression and immune evasion. Gene expression profiling revealed that positively regulated DEGs in IL-32-high tumors were enriched in immune-related pathways [Figure 3A-C and Supplementary Table 1]. Based on the above results, it could be inferred that IL-32 is critical for shaping the glioma immune microenvironment, particularly through its association with M2-type TAMs and immune-related pathways, potentially contributing to immune evasion and tumor progression.

To investigate the immunological role of IL-32 in glioma, we examined its correlation with IC genes. IL-32 was strongly and positively associated with major immune checkpoints (ICs): CD274 (PD-L1), PDCD1 (PD-1), as well as PDCD1LG2 (PD-L2) (P < 0.05; Figure 4). These results indicate that IL-32 plays a role in the immune modulation of the glioma TME through interactions with checkpoint-associated pathways, highlighting its potential as a regulator of immune evasion and a candidate therapeutic target.

To establish a survival risk model, we first assessed the prognostic significance of the 51 DEGs consistently identified across TCGA, CGGA, and REMBRANDT cohorts [Supplementary Figure 2A]. Consensus clustering based on these DEGs divided glioma patients into two distinct molecular subtypes (K = 2) in all three datasets [Figure 5A], with cluster stability confirmed by CDF analysis [Supplementary Figure 2B]. PCA further validated the clustering, with the first principal component (PC1) effectively separating subtype 1 from subtype 2 [Figure 5B]. A heatmap of clinical characteristics revealed that subtype 2 exhibited higher expression of the 51 DEGs [Figure 5C] and was associated with significantly shorter OS compared with subtype 1 [Figure 5D]. These results were consistent across all cohorts, demonstrating the strong prognostic value of the 51 DEGs and establishing a robust basis for constructing a survival risk model to inform personalized therapeutic strategies in glioma.

To evaluate the biological effects of IL-32 over a concentration-dependent range, U87 and LN229 cells were treated with recombinant human IL-32 at 0, 20, 40, and 60 ng/mL. The selected concentrations were based on previously published studies investigating IL-32-mediated regulation of macrophage function, in which recombinant IL-32 was commonly used at concentrations up to 40 ng/mL^[13,14]. An additional 60 ng/mL group was included to determine whether the biological effects of IL-32 on glioma cells remained concentration-dependent over a broader range. EdU assays demonstrated that IL-32 promoted cell proliferation in a dose-dependent manner [Figure 6A-D]. Transwell migration and invasion assays further showed that exogenous IL-32 markedly enhanced glioma cell motility and invasiveness [Figure 6E-H]. To validate these effects, siRNA-mediated knockdown of IL-32 significantly reduced proliferation, migration, and invasion compared with controls, whereas IL-32 overexpression produced the opposite effects [Figure 7]. These results collectively identify IL-32 as a pivotal regulator of glioma progression and support its potential as a therapeutic target.

To develop a precise prognostic tool for glioma, LASSO-penalized Cox regression was applied to the core set of DEGs [Figure 8A and B]. Ten genes, namely IL13RA2, FCER1G, S100A8, GPNMB, LUM, ANXA1, EMP3, CHI3L1, PTX3 and IGFBP2, were selected to construct the risk model. Each patient’s risk scores were computed by multiplying LASSO-derived coefficients of each gene by its corresponding mRNA expression level, then summing the product: Risk score = 0.0240873736114369 * IL13RA2 + 0.0639986787396445 * FCER1G + 0.012863666565142 * S100A8 + 0.114384024092816 * GPNMB - 0.064292476977219 * LUM + 0.0877923634931701 * ANXA1 + 0.000136148895132336 * EMP3 + 0.0220760633277022 * CHI3L1 + 0.0696953333682331 * PTX3 + 0.297974768973838 * IGFBP2.

The TCGA cohort served as the training set, with CGGA and REMBRANDT cohorts used for validation. Heatmaps of clinical characteristics revealed that the 10 genes were upregulated in the HR group [Figure 8C]. Patients were stratified into HR and LR groups based on median risk scores (TCGA: 5.232; CGGA: 2.898; REMBRANDT: 5.863) [Figure 8D]. HR patients exhibited a greater likelihood of mortality [Figure 8D], and OS was significantly shorter compared with LR patients [Figure 8E]. Univariate and multivariate Cox regression models were applied to determine whether the IL-32-associated risk score serves as an independent prognostic factor [Supplementary Figure 3]. In the multivariate Cox models, established prognostic factors, including age, WHO grade, IDH mutation status, 1p/19q co-deletion status, and MGMT promoter methylation status when available, were incorporated as covariates. The risk score remained independently associated with overall survival across the TCGA, CGGA, and REMBRANDT cohorts, with hazard ratios of 2.26 (95%CI: 1.67-3.00, P < 0.001), 1.40 (95%CI: 1.22-1.60, P < 0.001), and 1.50 (95%CI: 1.06-2.00, P < 0.05), respectively [Figure 8F], indicating that the model provided prognostic information beyond established clinicopathological and molecular features. Time-dependent ROC analysis further demonstrated that the 10-gene risk model achieved improved predictive discrimination compared with conventional prognostic variables, including WHO grade, age, IDH mutation status, 1p/19q co-deletion status, and MGMT promoter methylation [Supplementary Figure 3]. Moreover, DCA showed that the risk model provided greater net clinical benefit than individual traditional prognostic factors across a range of clinically relevant threshold probabilities [Supplementary Figure 4]. Together, these findings support the robustness of the IL-32-related risk model and suggest its potential value for individualized prognostic stratification in glioma patients.

To evaluate the model’s predictive power for immunotherapy, the TIDE algorithm was applied to transcriptomic data from glioma cohorts. In TCGA, LR patients had a higher proportion of predicted responders (61%) compared with HR patients (25%) (Fisher’s exact test, P < 0.0001; Figure 9A). Similar trends were observed in CGGA and REMBRANDT cohorts [Figure 9B and C], confirming the model’s robustness across independent datasets. Further validation using the independent GSE121810 cohort, which included 15 recurrent GBM patients treated with neoadjuvant anti-PD-1 therapy, showed that responders and non-responders formed distinct clusters by PCA [Figure 9D]. In contrast with responders, non-responders showed higher risk scores [Figure 9E]. These results indicate that the 10-gene risk model can stratify glioma patients by prognosis and predict their likelihood of responding to immunotherapy.

To investigate possible therapeutic strategies for HR glioma patients, drug sensitivity was estimated utilizing the oncoPredict R package^[11], with IC50 values representing sensitivity (higher IC50 indicates greater resistance). Pearson correlations between risk scores and IC50 values were calculated, focusing on drugs with r < -0.50 [Figure 10A]. Six compounds, namely lovastatin, TGX-221, ifosfamide, IC-87114, PLX-4720, and fluvastatin, showed consistent sensitivity across TCGA, CGGA, and REMBRANDT cohorts [Figure 10B]. These results highlight potential therapeutic options for personalized treatment in HR glioma patients.