We extracted the RNA expression profiles, gene mutations, and relevant clinical data for LGG from the databases of TCGA. The TCGA datasets were used as the training set to construct the model. To validate externally, RNA transcriptomics data from two glioma cohorts were obtained from the CGGA databases (13, 14). The training dataset aided in the creation of a predictive model, whereas the external validation dataset evaluated the model’s strength and accuracy in prediction. For analysis purposes, the data were presented in FPKM format, which stands for Fragments Per Kilobase of transcript per Million mapped reads, and then transformed using a logarithmic function. Furthermore, the RNA-seq information of healthy brain tissue was obtained from the GTEx database (https://commonfund.nih.gov/GTEx). The ‘sva’ package in R corrected the batch effect on the TCGA, GTEx, and CGGA datasets.

To calculate the disulfidptosis scores for individual LGG samples, the Gene Set Enrichment Analysis (GSEA) utilized the ‘GSVA’ and ‘GSEABase’ packages in R. The gene set associated with disulfidptosis was carefully selected based on an extensive review of relevant literature (7). Supplementary Table 1 contains the exhaustive compilation of these genes.

The TCGA-LGG dataset was used to construct gene co-expression networks through the execution of the WGCNA using the ‘WGCNA’ package in R. By examining the connectivity within each set of genes and its correlation with phenotypic traits, WGCNA allows for the identification of highly co-expressed gene modules as well as potential biomarker genes or therapeutic targets. In particular, the WGCNA method was utilized to detect gene modules that are associated with disulfidptosis scores in LGG patients and to describe these genes within the identified modules. Additionally, in order to ensure WGCNA’s rationality, we performed a hierarchical cluster analysis to exclude outliers.

Differential expression analysis was conducted using control samples of healthy brain tissue obtained from the GTEx database. To identify differentially expressed genes (DEGs) (15), the R package ‘limma’ was utilized. After applying the Benjamini-Hochberg method for multiple hypothesis testing adjustments, DEGs were chosen by considering an absolute log2 fold change (|log2FC|) greater than 1.585 and an adjusted p-value lower than 0.05, ensuring a false discovery rate of less than 0.05. The ‘ggplot2’ R package was used to create a volcano map for the visualization of DEGs. To facilitate further analyses, a Venn diagram was utilized to determine the overlap between important module hub genes obtained from WGCNA and the DEGs.

To begin with, a univariate Cox regression analysis was conducted to identify genes with prognostic importance by utilizing the hub genes associated with disulfideptosis that were differentially expressed. Furthermore, a comprehensive examination of two machine learning algorithms was employed to screen these predictive genes. The LASSO algorithm, which utilized penalty parameter tuning through 10-fold cross-validation, and the SVM-RFE algorithm, which selected the optimal variables based on the minimum 10-fold cross-validation error value, were the two algorithms used. A prognostic signature was developed by performing multivariate Cox survival analysis on the extracted and selected overlapping genes using the LASSO and SVM-RFE algorithms. To calculate prognostic signature risk scores, the following formula was utilized: R i s k   s c o r e s = ∑ i = 1 n ( C o e f i * E x p i ) , Where ‘i’ denoted the serial number of gene, ‘Coef’ represented the coefficient value, and ‘Exp’ represented the expression value of gene.

It is crucial that the prognostic signature is validated by another independent cohort. As a result, the strength of our predictive pattern was confirmed by validating it on two separate glioma datasets (mRNAseq_325 and mRNAseq_693) obtained from the CGGA database. The TCGA- LGG cohort and two CGGA cohorts were divided into high-risk and low-risk groups based on the median risk score obtained from the TCGA cohort. Following assessments involved the utilization of Kaplan-Meier survival analysis and Receiver Operating Characteristic (ROC) curve analysis to evaluate the variations in prognosis between the high-risk and low-risk subcategories, as well as the precision of the risk model’s predictions.

The Cox model was used to assess independent predictors of OS. The variables that showed a strong correlation with prognosis in the univariate Cox analysis were later included in the multivariate Cox analysis. Furthermore, the Cox proportional hazards models were used to calculate hazard ratios (HRs) and their corresponding 95% confidence intervals (CIs).

To predict the chances of 3-, 5-, and 7-year OS in patients with LGG, a nomogram was developed. The ‘rms’ and ‘regplot’ packages in R were used to create this model, which combines the recognized independent prognostic clinical factors and the risk score signature. To evaluate the reliability and precision of the developed nomogram, calibration curves and decision curve analysis (DCA) were utilized. Dynamic nomograms were created using the “DynNom” package, and an interactive web-based application was created using the Shiny platform (https://yudong-cao.shinyapps.io/dynamic-Nomo/) (16).

The R package ‘GSVA’ was used to perform gene set variation analysis with the ‘c2.cp.kegg.v7.5.1.symbols.gmt’ dataset obtained from the Molecular Signatures Database. The R package ‘heatmap’ was used to generate a heat map for visualizing the enrichment results. Statistical significance was determined at an adjusted p-value of less than 0.05 using the “limma” R package.

To determine the makeup of 22 different types of immune cells in humans, the CIBERSORT algorithm was employed by analyzing gene expression profiles (17–19). For successful deconvolution, a P-value of less than 0.05 was used to determine the statistical significance after executing 1,000 permutations for each sample.

The ESTIMATE algorithm is specifically created to measure the amount of infiltrating immune and stromal cells present in tumor tissue, while also providing an estimation of tumor purity based on gene expression data (20). Using ssGSEA, ESTIMATE calculates scores for immune, stromal, and tumor purity. The application of this tool was used to assess the tumor purity, immune score, and stromal score of every LGG sample. Furthermore, the ssGSEA was employed to calculate numerical values for every individual stage in a seven-stage cancer immunity cycle, which acts as a structure for assessing anti-cancer immune responses (21).

The TIDE algorithm is an online tool created to measure response metrics of immunotherapy, acting as a predictive tool to evaluate the efficacy of immune checkpoint inhibitors, and performing thorough analyses on various tumor expression profiles. By employing this tool, we computed several immune scores, such as immune malfunction, immune segregation, cancer-related fibroblast (CAF), and the comprehensive TIDE score. These scores were derived from standardized expression profiles obtained from the TCGA repository (22). To assess disparities in TIDE scores between subgroups at high and low risk, a Wilcoxon test was conducted.

The gene mutation patterns of patients with LGG, acquired from the TCGA database, were produced utilizing the ‘maftools’ application. Afterwards, the comprehensive gene mutation records were combined with the risk assessment. Additionally, we utilized the R software ‘oncoPredict’ to calculate the IC50 values, representing the concentrations at which the inhibitory effect of commonly used chemotherapeutic drugs is reduced by half. The assessment enabled the examination of the correlation between the risk rating and the responsiveness to medication. Wilcoxon signed-rank tests were conducted to compare the IC50 values between the two risk groups.

The Tumor Immune Single-Cell Hub (TISCH) was employed to examine scRNA-seq data of the GSE148842 dataset. TISCH functions as a dedicated repository for scRNA-seq, specifically targeting the TME [15]. By providing comprehensive annotations of cell types at the individual cell level, it enables the exploration of the TME in different types of cancers (23).

The glioma cell lines SHG44 and U251 were utilized in the experiments and were cultured in DMEM (Dulbecco’s Modified Eagle Medium) with the addition of 10% fetal bovine serum. The cells were kept in a typical sterile CO2 incubator at a temperature of 37°C, with a CO2 concentration of 5%. Every day, the cells were rinsed thrice using PBS (Phosphate-Buffered Saline), and the culture medium was substituted. When the cell confluency reached 80%-95%, subculturing was carried out with a split ratio of 1:2.

The Ethics Committee of Xiangya Hospital, Central South University granted ethical approval for this study (Ethical Approval Code 202309185). Informed consent was obtained from all patients who participated. All participants were included in the study. During the time frame from January 2021 to August 2022, the Department of Neurosurgery at Xiangya Hospital, Central South University, obtained a total of 70 glioma samples and 12 normal tissues from the surrounding area. Following previous protocols, the specimens were processed into tissue microarrays to investigate the expression of ABI3 (24).

To remove paraffin and rehydrate, tissue microarrays embedded in paraffin were baked at a temperature of 65°C for one hour. The endogenous peroxidase activity was eliminated by incubating in methanol with 0.3% hydrogen peroxide for 30 minutes. Afterwards, the sections were obstructed for 30 minutes using a solution of 2% BSA in PBS. The anti-ABI3 antibody (diluted 1 250; SANTA, USA) was incubated overnight at 4°C.The MaxVision HRP-polymer IHC Kit Detection System (peroxidase/DAB, rabbit/mouse) from MaxVision, Fuzhou, China, was utilized to visualize immunocomplexes, following the guidelines provided by the manufacturer. Hematoxylin (Beyotime Biotechnology) was employed for nuclear counterstaining. Evaluation of staining was conducted using an optical microscope with a magnification of 200×.We conducted a semi-quantitative evaluation of ABI3 protein expression using methodologies that were previously established in our research (25).

The cells were placed in a 6-well dish with a seeding density of 2.5×105 cells per well. The cells were transfected with 5μl of ABI3 SiRNA (The siRNAs targeting the ABI3 gene were synthesized by RiboBo Corporation (Guangzhou, China)) using Lipofectamine 3000 Reagent (Invitrogen, Carlsbad, CA, USA) following the manufacturer’s instructions when they reached a confluency of 60% to 80%.

The Trizol lysis method was used to extract total RNA from SHG44 and U251 cells that were treated. The Thermo Scientific RevertAid First Strand cDNA Synthesis Kit (Thermo Scientific, Waltham, MA) was utilized for the synthesis of cDNA. RNA levels of ABI3 andβ-Actin were analyzed by the 2−ΔΔCt method. The primers were created by The Beijing Genomics Institute (BGI) and the sequences were designed as follows: for ABI3, the forward primer was 5’-CAGGTGGAAGCCCGTGTAAG-3’ and the reverse primer was 5’-AGTGGCTAAGGTGCCGATCTC-3’. The forward primer for β-Actin was 5’-CATGTACGTTGCTATCCAGGC-3’ and the reverse primer was 5’-CTCCTTAATGTCACGCACGAT-3’.

The Western blotting assay was performed in accordance with the instructions provided by the manufacturer. Western blotting utilized antibodies targeting ABI3 (sc-376982, SANTA), β-Actin (20536-1-AP, Proteintech), vimentin (10366-1-AP, Proteintech), and ZO-1 (21773-1-AP, Proteintech).

Prior to conducting the cell migration assay, U251 and SHG44 cells underwent interference treatment. When the cell density reached 80%-90%, a 100 µl pipette tip was utilized to generate four intersecting lines in every well, producing a crisscross pattern that resembled a grid. After the line was created, the culture medium was substituted with either a serum-free or low-serum medium. Subsequently, 8-10 positions were selected and recorded under an imaging microscope. Thereafter, at 24-hour intervals, the same positions were imaged using the imaging microscope. The experiment involved capturing multiple time points (2, 3). The distances of migration were measured using suitable software for statistical analysis based on the captured images.

Migration and invasion experiments were conducted in a 24-well plate using 8.0 μm pore inserts (Millipore, Bedford, MA, USA).2 million cells were placed in the upper chamber of the Transwell insert for the migration experiment. To perform the invasion test, Matrigel-coated filters (BD Biosciences, Franklin Lakes, NJ, USA) were utilized. For migration and invasion assays, cells were incubated for 24 and 48 hours, respectively. After incubation, the cells that migrated and invaded were fixed and stained using 0.1% (w/v) crystal violet. Using high-power microscopy, the number of cells that were chosen randomly, migrated, and invaded were counted.

1000 cells were cultured in 96-well plates, each sample was repeated 5 times, and the number of cells was calculated daily using the CCK-8 assay (A311-02, China Nanjing Novizan Biotechnology Co., Ltd.) for 5 days. For colony formation assay, cells were cultured in 6-well plates at a density of 500 cells per well for 2 weeks. The medium was changed every 3 days. After 2 weeks, the medium was removed and the cell colonies were stained with crystal violet (0.1%, 20% methanol) for 10 minutes.

The U251 and SHG44 glioma cell lines were taken out of the culture medium and rinsed twice with PBS. Afterwards, they were subjected to 3.7% formaldehyde treatment for a duration of 20 minutes, followed by 2-4 washes using PBS containing 0.1% Triton X-100. Each wash lasted for 5 minutes. The Actin-Tracker Red (C2205s, Beyotime) was thinned in PBS with 1-5% BSA and 0.1% Triton X-100 at a proportion of 1: 40-200. Following that, the cells were placed in a dark room at room temperature and incubated for 60 minutes. DAPI (Sigma, USA) was used to stain the cell nuclei. Observations were made using a laser confocal microscope.

GraphPad Prism software (version 9.0) was utilized to perform statistical analysis on the experimental data. The data were obtained from three separate experiments and are presented as the mean plus or minus the standard deviation (SD). Student’s t-tests were conducted for intergroup comparisons, with significance levels indicated as *P< 0.05, **P< 0.01, and ***P< 0.001.

Figure 1 displayed the schematic diagram of the research process. WGCNA was applied to systematically identify gene modules that exhibit covariance with disulfidptosis. As evidenced by Figure 2A , the adoption of a soft-thresholding value of 4 results in data that more closely adhere to a power-law distribution, while also stabilizing average network connectivity. This renders the dataset amenable to subsequent analyses. As depicted in Figure 2B , twelve distinct non-gray modules were generated through the amalgamation of modules exhibiting similarity coefficients below 0.15, with a minimum module size established at 50 genes and a deepSplit parameter set to 2. As discerned from Figures 2C, D , the brown and blue modules exhibit the strongest correlation with disulfidptosis, with correlation coefficients of -0.67 and 0.66, respectively. Figure 2D delineates the statistical significance of gene constituents within the brown and blue modules in relation to their correlation with disulfidptosis scores. Furthermore, WGCNA was also employed to pinpoint hub genes within brown and blue modules, resulting in the identification of 181 key hub genes.

Initially, an integrated analysis of TCGA and GTEx databases was performed, yielding 2093 DEGs specific to LGG ( Figures 3A , Supplementary Table 2 ). A Venn diagram was employed to intersect the DEGs with the hub genes identified via WGCNA, identifying 39 vital regulatory genes ( Figures 3B , Supplementary Table 3 ). Following the execution of univariate Cox regression analysis, a subset of 34 genes was retained, each demonstrating a significant correlation with prognosis ( Supplementary Table 4 ). Subsequently, upon isolating 34 prognostic genes, the LASSO method was applied to refine this list to 19 essential genes (as shown in Figures 3C, D ), and the SVM-RFE algorithm was employed to narrow down the list to 16 essential genes ( Figures 3E, F ). Upon intersecting the critical genes identified through both LASSO and SVM-RFE algorithms, a set of 13 potentially important genes emerged ( Figure 3G ). These genes were subsequently subjected to multivariate Cox regression analysis for further refinement and the construction of 9-gene prognostic signature using the following equation: Risk Score= 1.656 × expABI3 + 1.041 × exp APOBEC3C + 0.949 × expCD53 + 0.851 × expRNASE6 + 0.276 × expOLFML3 - 0.222 × expHLA-DRB1 - 0.942 × expTYROBP - 1.201 × expSIGLEC10 - 1.763 × expTNFAIP8L2 ( Figures 3H ).

In the training cohort, samples were segregated into low-risk and high-risk categories, as illustrated in Figure 4A . The incidence of mortality was significantly lower in the low-risk set than in the high-risk set ( Figure 4A ).

The heatmap delineates that the expression levels of the nine model genes were markedly elevated in the high-risk subgroup compared to the low-risk subgroup within the TCGA cohort ( Figure 4A ). These findings were corroborated through analysis of two independent databases, namely the CGGA_325 and CGGA_693 datasets ( Figures 4B, C ).

As depicted in Figures 4D-F , the prognosis for patients categorized into the high-risk group was significantly poorer than those in the low-risk group across both the TCGA training cohort and the two CGGA validation cohorts (P< 0.001). This evidence substantiates that the disulfidptosis-related prognostic model exhibits high accuracy in forecasting patient outcomes across both the TCGA and CGGA cohorts. ROC curve analysis was conducted in the training set and the two test sets to further evaluate the predictive accuracy of this signature. The areas under the curve for the test group at 3, 5, and 7 years were observed to be 0.87, 0.79, and 0.80, respectively ( Figure 4G ). The Area Under the Curve (AUC) exceeded 0.72 for both the two CGGA test sets, indicating robust prognostic accuracy ( Figures 4H, I ).

Clinical circle plots were constructed to elucidate disparities in clinicopathological features between the low-risk and high-risk groups within the context of LGG, and the analysis revealed an asymmetric distribution across all examined features with the exception of gender ( Figure 5A ). Patients presenting with higher pathological grade, advanced age, wild-type IDH1, absence of 1p/19q co-deletion, and un-methylated MGMT promoter were primarily allocated to the high-risk category ( Figures 5A ).

To rigorously validate the predictive efficacy of the disulfidptosis-related risk model in conjunction with other clinical features, both univariate and multivariate Cox regression analyses were conducted within the TCGA training cohort. Univariate Cox regression analysis revealed that both the risk scores and specific clinical parameters—namely age, tumor grade, IDH1 mutation status, 1p/19q co-deletion status, and MGMT promoter methylation status—exhibited significant associations with the prognosis of LGG, as depicted in Figure 5B . Multivariate Cox regression analysis substantiated that the risk score persisted as an independent prognostic factor (HR = 1.99, 95% CI 1.594-2.485, P< 0.001), as illustrated in Figure 5C . Conventional clinical attributes, such as age (HR = 1.041, 95% CI 1.024-1.058, P< 0.001) and 1p/19q co-deletion status (HR = 2.214, 95% CI 1.294-3.789, P = 0.004), were also verified as independent determinants of OS ( Figure 5C ).

Further investigation was conducted to elucidate the disparities in drug resistance among the two risk groups, as depicted in Figures 5D-G . According to our analysis, Gemcitabine, Dasatinib, Selumetinib, and PD0325901 have shown promise as potential treatment options for individuals classified as high-risk.

Incorporating the risk score along with all independent prognostic clinical variables, a nomogram was constructed to quantify better the risk associated with LGG patients ( Figure 6A ). A nomogram was employed to enhance the accuracy of patient risk determination, thereby informing subsequent treatment strategies. Furthermore, the clinical prediction models were assessed and optimized through decision and calibration curve analyses, offering a comprehensive evaluation of their value in clinical decision-making. The findings demonstrated that the performance of the nomogram surpassed other clinical markers, indicating its robust capability in prognostic prediction ( Figures 6B, C ). Consequently, it holds the potential for utilization as a clinical decision-support instrument. Besides, the nomogram was further developed into an online, interactive tool to streamline the risk assessment process and facilitate user engagement and clinical decision-making (https://yudong-cao.shinyapps.io/dynamic-Nomo/) ( Figure 6D ). The web server is designed to generate estimated survival rates and Kaplan-Meier curves upon entry of the relevant covariates.

The “c2.cp.kegg.v7.5.1.symbols” file was acquired from the Molecular Signature Database, and subsequent GSVA analysis was conducted utilizing this gene set to gain initial insights into the potential mechanisms underlying the impact of the risk score on prognosis in LGG patients. The GSVA results indicated that the high-risk group exhibited significant enrichment in biological processes encompassing signaling pathways related to DNA replication and damage repair, immune response, apoptosis, cell signaling, cell-cell interactions, and cell-matrix interactions, such as notch signal pathway, adherens junction, pathways in cancer, DNA replication, non-homologous end joining, focal adhesions, jak stat signaling pathway, p53 signaling pathway, MARK signaling pathway, and regulation of the actin cytoskeleton ( Figure 7 ). Biochemical and metabolic pathways were predominantly enriched in the low-risk group. The aforementioned findings reaffirmed that the disulfidptosis-related genes utilized for risk score calculation play crucial roles in malignant progression, metabolism, DNA damage repair, immune responses, cell communication, and cytoskeleton regulation of glioma.

The CIBERSORT technique assessed the level of immune cell penetration in each sample, providing insights into the arrangement and interrelation of the relative ratios of 22 immune cells that infiltrated tumors within the TCGA-LGG group. Figure 8A provides an overview of the landscape depicting the infiltration patterns of 22 distinct immune cell types. In particular, patients classified in the high-risk group demonstrated elevated levels of regulatory T cells (T-cell regulatory cells and T-cell follicular helper cells) and quiescent immune cells (resting CD4+memory T cells and resting NK cells). Conversely, the proportion of T-cell CD4+ primary cells and activated natural killer cells was notably lower in this group ( Figure 8A ). The TME plays a pivotal role in modulating the immune response against malignant cells. We assessed the composition of the TME between different risk subgroups within the TCGA cohort. Briefly, elevated stromal and immune scores were significantly associated with the high-risk subset compared to the low-risk subgroup. Conversely, tumor purity was markedly lower in the high-risk subgroup ( Figure 8B ). This indicates that high-risk scores are concomitant with increased levels of immune cell infiltration within the TME. Furthermore, the high-risk subgroup exhibited elevated expression levels of most immune checkpoints, while the low-risk subgroup displayed the opposite pattern, as depicted in Figure 8C . These data implied that the upregulation of immunosuppressive cells, the inactivation of NK cells, and the expression of immune checkpoints may contribute to establishing an immunosuppressive TME in patients classified as high-risk group.

Finally, the TIDE analysis was conducted to assess the variability in immunotherapeutic responsiveness among patients stratified by distinct risk profiles. Based on the study findings, it was observed that patients categorized in the low-risk group displayed a more favorable response to immunotherapy. This observation can be attributed to their comparatively lower TIDE, Dysfunction, and Exclusion scores, as depicted in Figure 8D .

The cancer immunotherapy field is guided by the conceptual framework of the 7-step cancer-immunity cycle, which is currently driving modern research (21). Using TIP, a web tool for evaluating tumor immunophenotypes (26), we examined the immune response against cancer in various risk subcategories of LGG patients by analyzing the seven-step cancer-immunity cycle. In patients belonging to the high-risk subgroup, there was notable heightened activity observed at steps 1 (pertaining to the release of tumor antigens), 2 (related to cancer antigen presentation), 4 (involving T-cell migration to the tumor site), and 5 (linked to immune cell penetration into the TME). Conversely, steps 3 (characterizing initiation and activation processes of immune cells), 6 (involving T-cell recognition of tumor cells), and 7 (concerning the killing of tumor cell) exhibited distinctive inhibitory characteristics ( Figure 9A ). Contrarily, patients categorized within the low-risk subgroup displayed augmented activity levels at steps 3, 6, and 7, while showing discernible inhibition at steps 1, 2, 4, and 5, as depicted in Figure 9A . These results suggested ameliorating the immunosuppressive conditions within the high-risk subgroup of patients, as well as enhancing immune cell infiltration among individuals within the low-risk subgroup, could potentially be conducive to achieving favorable clinical outcomes in patients with LGG.

Moreover, it is essential to note that the majority of genes implicated in the negative regulation of immune processes exhibited upregulation within the high-risk subgroup of LGG patients, as visually represented in Figure 9B . Finally, we conducted an in-depth examination of potential pathways linked to risk subgroups through GSEA enrichment analysis. The outcomes unveiled a significant enrichment of negative regulation of immune pathways within the high-risk subgroups, encompassing pathways such as B-cell-mediated humoral immunity, cellular immunity, adaptive immune response, and immunoglobulin-mediated immune response ( Figure 9C ).

Distinct patterns of somatic mutations were identified between the high- and low-risk subgroups. Despite IDH mutations being the most prevalent, the relative frequency of IDH mutations exhibited variations across subtypes. Among patients within the low-risk subgroup, IDH1 mutations were notably more frequent, constituting 92% of the total mutations. In contrast, patients within the high-risk subgroup displayed a lower proportion, with IDH mutations accounting for only 59% of the total mutations, as illustrated in Figures 10A, B . Furthermore, it is noteworthy that the highest incidence of TP53 mutations was observed within the low-risk subgroup, accounting for 49%, followed closely by the high-risk subgroup, where TP53 mutations constituted 43% of the mutations. Tumor Mutation Burden (TMB) levels exhibited significant disparities between the two risk subgroups, with a noteworthy positive correlation observed between risk scores and TMB values, as depicted in Figures 10C, D .

Subsequent investigations involved the assessment of mutation frequencies within some major oncogenic pathways among the two risk subgroups. The outcomes revealed that the RTK-RAS, Hippo, and TP53 pathways were predominantly identified in the high-risk subgroup, as illustrated in Figure 10E . Conversely, the low-risk subgroup showed a predominant detection of the Notch and Cell Cycle pathways, as shown in Figure 10F , whereas the other pathways did not display notable disparities between the two subgroups.

To obtain further validation on the particular cell types that express the genes in the risk model within the TME, scRNA-seq analysis was performed using the TISCH online tool on the GSE148842 dataset. A total of six distinct cell clusters were identified through UMAP analysis, which included the following categories: AC-like Malignant cells, CD8Tex cells, Malignant cells, Monocytes/macrophages, Oligodendrocyte, and other types of cells ( Figure 11A ). The findings revealed distinct patterns of gene expression within the identified cell clusters. Specifically, ABI3 exhibited predominant expression in malignant cells and monocytes/macrophages, while APOBEC3C was primarily expressed in CD8Tex cells. CD53, TYROBP, and HLA-DRB1 showed predominant expression in monocytes/macrophages, CD8Tex cells, and malignant cells. In contrast, SIGLEC10, TNFAIP8L2, OLFML3, and RNASE6 displayed low expression levels in non-tumorigenic and tumorigenic cells ( Figures 11B, C ).

Utilizing the GEPIA2 online platform (27), we assessed the expression profile of ABI3 in glioma and normal tissue samples. Significantly, ABI3 expression in glioma tissues was notably higher when compared to their normal counterparts ( Figure 12A ). Furthermore, analysis of the survival curve indicated that increased expression was linked to an adverse prognosis in patients with glioma ( Figure 12B ). In order to validate ABI3 expression levels in glioma tissues, we conducted immunohistochemistry experiments on tissue microarrays made from glioma specimens collected from Xiangya Hospital (containing 11 paraneoplastic tissues and 84 diffuse glioma samples). The experimental data demonstrated a significant diminution in ABI3 protein expression in paraneoplastic tissues relative to tumor tissues ( Figure 12C ). Moreover, there was a significant increase in protein expression of ABI3 in WHO grade III (P< 0.05) and grade IV gliomas (P< 0.001) compared to WHO grade II gliomas ( Figure 12D ). Nevertheless, no significant statistical difference in protein expression levels was observed between gliomas of WHO grade III and grade IV. Immunohistochemical data pertaining to ABI3 expression are presented in Figure 12E .

To elucidate the biological role of ABI3 in glioma, we initially assessed the mRNA expression levels of ABI3 across six glioma cell lines, namely HGS683, SHG44, U251, LN229, U87 and A172 ( Figure 12F ). Among them, the GBM cell lines exhibited relatively higher ABI3 mRNA levels than the LGG cell lines. We chose the LGG cell line SHG44 and the GBM cell line U251 for subsequent cell experiments. We validated the effective downregulation of ABI3 expression in SHG44 using siRNA-1 and siRNA-2 through qPCR and Western blot analysis ( Figures 12G, H ).

Further loss-of-function experiments were conducted on SHG44 and U251 cells. LGG cell line SHG44 and the GBM cell line U251 were segregated into a negative control group (NC) and an ABI3-silenced group (wherein ABI3 expression was suppressed in both SHG44 and U251 cell lines). Subsequent scratch assays and Transwell matrigel invasion assays revealed that ABI3 downregulation notably attenuated the migratory and invasive capabilities of SHG44 and U251 glioma cells ( Figures 13A–C ). Subsequently, we performed a correlation analysis between ABI3 expression and molecular markers indicative of epithelial-mesenchymal transition using the TCGA database. The analysis disclosed a significant negative association between ABI3 expression and the epithelial marker ZO-1 ( Figure 13D ), as well as a pronounced positive correlation with the mesenchymal marker vimentin ( Figure 13E ). To corroborate these observations, we quantified the expression levels of ZO-1 and vimentin via Western blot analysis. The experimental data revealed an upregulation of ZO-1 expression and a downregulation of vimentin expression following ABI3 interference in both SHG44 and U251 cell lines ( Figures 13F, G ). These findings suggest that overexpression of ABI3 may serve as a contributory factor in the facilitation of glioma cell invasion and metastasis.

Regarding cellular proliferation, CCK8 assay results indicated a moderate reduction in the proliferative capacity of SHG44 cells ( Supplementary Figure 1A ) following the interference with ABI3 expression. However, the proliferation of the U251 cell line ( Supplementary Figure 1B ) remained unaffected. Colony formation assay results demonstrated that interference with ABI3 expression in SHG44 and U251 cells did not yield significant differences in the number and size of cell colonies compared to the control group ( Supplementary Figures 2C, D ). These findings suggested that the knockdown of the ABI3 gene exerted a minimal impact on the proliferation of glioma cells.

Using the expression profile data from TCGA and GTEx, we analyzed the mRNA levels of SLC7A11 in patients with glioma. The findings indicated a significant upregulation of SLC7A11 in the tumor samples, as depicted in Figure 14A . Moreover, studies have shown that increased SLC7A11 levels contributed to the malignant progression and unfavorable prognosis of glioma (8, 28). As a result, we substituted the DMEM medium with the low-glucose 1640 medium for both the control group (NC) and the experimental group. Following 48 hours, we proceeded to stain the cytoskeleton using phalloidin. During the examination under confocal microscopy, it was observed that the suppression of ABI3 gene expression led to the contraction of the cytoskeleton and the formation of lamellipodia in SHG44 ( Figure 14B ) and U251 cells ( Figure 14C ). The findings suggested that ABI3 may act as a potential target in disulfidptosis.