Integrated proteomic and single-nucleus transcriptomic profiling identifies prognostic markers and therapeutic targets in glioblastoma Authors: Chengyong Yang, Junchi Luo, Sha Yang, Mei Deng, Tao Xiong, Junheng Zhang, Wenjing Yang, Jiqin Zhang, Xin Xiang, Liangzhao Chu, Ying Tan Published: 2026 ABSTRACT Glioblastoma (GBM) is the most aggressive primary malignant brain tumor in adults and remains associated with marked heterogeneity and poor prognosis. To identify prognostic biomarkers and characterize their cellular contexts, we integrated public proteomic and single-nucleus transcriptomic datasets. Protein expression data from the TCPA GBM cohort were used to construct a prognostic model by univariate and multivariable Cox regression analyses, followed by Kaplan–Meier survival, time-dependent ROC, and nomogram analyses. Single-nucleus RNA-seq data from GSE138794 were used to define the cellular distribution of model proteins and to perform marker-centered functional analyses. A four-protein signature consisting of BCL2, CTNNB1, CHEK2, and NDRG1_pT346 was established. The derived risk score significantly stratified patients into high- and low-risk groups with distinct overall survival and progression-free survival outcomes, and remained independently associated with prognosis after adjustment for clinical covariates. In the single-cell atlas, inferCNV-guided refinement identified malignant glioma cells and further resolved them into Glioma_MES, Glioma_PN, and Glioma_CL states. The four prognostic markers showed distinct cellular preferences: BCL2 and NDRG1 were mainly enriched in oligodendroglial-related populations and selected glioma states, whereas CHEK2 and CTNNB1 were more highly expressed in myeloid-associated populations. Among the four markers, only CHEK2- and NDRG1-defined malignant glioma subgroups showed sufficient transcriptional divergence for downstream analysis. CHEK2-positive glioma cells were enriched for DNA replication, chromosome segregation, DNA damage response, and checkpoint-related programs, whereas NDRG1-positive glioma cells were characterized by hypoxia response, glycolysis, and metabolic adaptation. CellChat analysis further revealed distinct communication patterns for CHEK2- and NDRG1-associated malignant glioma cells.Supplementary InformationThe online version contains supplementary material available at 10.1007/s12672-026-05218-3. Background Glioblastoma (GBM) is the most common and aggressive primary malignant tumor of the central nervous system in adults, characterized by diffuse infiltration, high recurrence rates, and resistance to standard therapies. Despite advances in surgical techniques, radiotherapy, and chemotherapeutic regimens such as temozolomide, the prognosis of GBM patients remains dismal, with a median survival of approximately 15 months and a 5-year survival rate of less than 10% [1, 2]. The inherent inter- and intra-tumoral heterogeneity of GBM, coupled with its highly immunosuppressive microenvironment, poses substantial challenges to effective treatment and biomarker-guided patient stratification [3]. Although conventional molecular biomarkers such as IDH1 mutation, MGMT promoter methylation, and TERT promoter mutation have been incorporated into clinical classification systems, these markers offer limited value in accurately predicting therapeutic response or capturing dynamic biological processes [4]. In recent years, the application of multi-omics technologies has significantly broadened the scope of GBM research. Among these, proteomics plays a critical role in uncovering functional alterations at the translational level, offering insights into post-transcriptional regulation and protein-mediated signaling cascades that are not detectable through transcriptomic profiling alone [5]. Because proteins more directly reflect downstream biological activity than transcripts alone, proteomics-based analyses may provide complementary prognostic information and help prioritize biologically relevant candidate markers for further investigation. Integrating proteomic data with clinical variables has thus emerged as a promising strategy for identifying robust prognostic indicators and candidate therapeutic targets. In parallel, single-nucleus RNA sequencing (snRNA-seq) has enabled high-resolution mapping of gene expression at the single-cell level, facilitating the characterization of tumor heterogeneity, lineage differentiation, and cell-type-specific regulatory programs [6]. This technique is particularly valuable in the context of GBM, where complex cellular ecosystems—comprising malignant cells, astrocytes, microglia, and infiltrating immune populations—coexist and interact through diverse ligand–receptor signaling pathways. The development of computational tools such as CellChat has further enabled the systematic inference of intercellular communication networks, offering a framework to explore how signaling dynamics may be associated with tumor progression and therapeutic resistance [7]. Given these advances, the integration of proteomic and single-cell transcriptomic data provides a powerful framework for dissecting the molecular landscape of GBM and identifying functionally relevant biomarkers. However, few studies to date have combined these two omics layers to prioritize proteomics-derived prognostic markers and to examine their potential cellular context within GBM. Accordingly, in the present study, we first identified prognostically relevant proteins from TCPA proteomic data and constructed a protein-based risk signature for GBM. We then used public single-nucleus transcriptomic data to investigate the cellular distribution and potential microenvironmental associations of the selected markers, followed by drug–gene interaction analysis to prioritize candidate compounds for future validation. This integrative design was intended to provide a proteomics-guided and cell-contextualized framework for biomarker prioritization in GBM, rather than to establish definitive molecular mechanisms. Methods and materials Data collection Protein expression data for GBM were obtained from The Cancer Proteome Atlas (TCPA) portal (https://www.tcpaportal.org/tcpa/, accessed on July 18, 2025) [8]. A total of 205 GBM samples with expression data for 223 proteins were initially retrieved from the TCPA portal. Corresponding clinical information, including overall survival time, survival status, Age, Gender, and Race, was also collected. Samples lacking survival information or key clinical variables were excluded, resulting in a final cohort of 188 GBM cases for downstream prognostic analyses. The public single-nucleus RNA-seq dataset GSE138794 was downloaded from GEO to provide cellular context for the proteomics-derived prognostic markers [9]. To match the GBM-focused proteomic cohort and minimize heterogeneity, only 10 primary GBM snRNA-seq samples (GSM4119521–GSM4119530) were included. Samples derived from scRNA-seq or snATAC-seq, as well as non-GBM glioma samples, were excluded. Raw count matrices and sample metadata were used for downstream analyses. This dataset was used for cell-resolved characterization of marker expression rather than as an independent validation cohort. The study flow is shown in Supplementary Fig. 1. Construction of the protein-based prognostic signature To identify prognostic proteins associated with the overall survival (OS) of GBM patients, univariate Cox proportional hazards regression analysis was first performed on the protein expression dataset. Proteins showing statistical significance in the univariate analysis were retained for subsequent multivariate Cox regression analysis to construct the final prognostic protein signature. A prognostic risk score for each patient was calculated as a linear combination of the expression levels of the selected proteins weighted by their corresponding multivariate Cox regression coefficients. Patients were then stratified into high-risk and low-risk groups according to the median risk score. Kaplan–Meier survival analysis was conducted using the survival R package to compare OS between the two groups. Time-dependent receiver operating characteristic (ROC) curves and area under the curve (AUC) values were generated using the timeROC R package to evaluate the predictive performance of the protein-based prognostic signature [10]. The predictive performance of the signature was assessed at 1-, 3-, and 5-year time points. To assess the validity of the multivariable Cox model, the proportional hazards assumption was examined using Schoenfeld residual-based tests implemented by the cox.zph function in the survival package. Independent prognostic analysis of the protein-based signature To determine whether the protein-based risk score could serve as an independent prognostic factor, both univariate and multivariate Cox proportional hazards regression analyses were performed, incorporating clinical variables such as Age, Gender, and Race along with the risk score. Variables with a p-value < 0.05 in the univariate analysis were included in the multivariate model. Hazard ratios (HRs), 95% confidence intervals (CIs), and corresponding p-values were calculated to evaluate the prognostic significance of each variable. After the protein-based risk score had been calculated for each patient, it was integrated with the available clinical variables for independent prognostic analysis. The protein-based risk score was considered an independent prognostic factor if it remained statistically significant in the multivariate Cox regression model after adjustment for the available clinical covariates. A forest plot was generated using the forestplot R package to visualize the HRs, CIs, and p-values for each variable. The proportional hazards assumption of the multivariable model incorporating Age, Gender, Race, and the protein-based risk score was further evaluated using Schoenfeld residual-based tests implemented by the cox.zph function in the survival package. Nomogram construction and evaluation of predictive accuracy To provide a clinically applicable tool for individualized survival prediction, a nomogram was constructed based on the multivariable prognostic model incorporating the protein-based risk score together with the available clinical variables (Age, Gender, and Race). The nomogram was developed using the rms package in R and was designed to estimate the probability of 1-, 3-, and 5-year OS for GBM patients [11]. Additionally, calibration curves were plotted to evaluate the agreement between predicted survival probabilities and observed outcomes. These curves were generated for 1-, 3-, and 5-year OS using 1,000 bootstrap resamples. Calibration analysis was used as an internal assessment of the agreement between nomogram-predicted survival probabilities and the observed outcomes. Single-cell analysis strategy Single-nucleus transcriptomic analysis was performed using the Seurat and Harmony R packages [12, 13]. A total of 16,046 nuclei from nine samples were initially included, and 13,515 high-quality nuclei were retained after quality control, with 2,531 nuclei (15.8%) removed. Ambient RNA contamination was estimated using decontX (mean contamination proportion = 0.054), and doublets were identified using scDblFinder, yielding 689 doublets and 12,655 singlets. A total of 3,000 highly variable genes were selected for downstream analysis. The filtered nuclei were processed in Seurat for normalization, variable feature selection, scaling, and principal component analysis [14]. To reduce inter-sample technical variation, batch correction and sample integration were performed using Harmony with sample identity as the batch variable. Graph-based clustering and UMAP visualization were then conducted based on the Harmony-corrected embeddings, and resolution optimization supported the use of resolution 0.5, resulting in 12 clusters. Cell populations were preliminarily annotated according to cluster-specific marker genes and canonical lineage markers, including AQP4, ALDH1L1, and SLC1A3 for astrocytes; MBP, PLP1, and MOG for oligodendrocytes; RBFOX3, SYT1, and SNAP25 for neurons; CX3CR1, P2RY12, and TMEM119 for microglia; and CD68, CD163, and C1QA for tumor-associated macrophages. To distinguish malignant glioma cells from non-malignant cells, inferCNV analysis was performed using neurons-like cells and oligodendrocytes as reference populations. Cell populations showing broad and recurrent copy-number variation patterns were defined as malignant, whereas those lacking such alterations were considered non-malignant. During annotation refinement, two populations with partial similarity to canonical oligodendrocytes or neurons were conservatively designated as oligodendrocytes-like cells and neurons-like cells, respectively. InferCNV-supported malignant clusters were further classified into glioma transcriptional states using curated marker programs, including CD44, VIM, CHI3L1, and GFAP for MES-like cells; OLIG2, SOX2, NES, FABP7, and SOX4 for PN-like cells; and EGFR, PDGFRA, CA9, SLC2A1, VEGFA, and ANGPTL4 for CL-like cells. Marker-centered single-cell stratification analysis To further evaluate the cellular contexts of the four prognostic markers, we first identified the top three cell populations with the highest expression for each marker based on average expression and expressing-cell proportion. Within each of these major expressing populations, cells were stratified into marker-positive and marker-negative groups according to whether the raw expression count of the target gene was greater than 0 or equal to 0, respectively. Only cell populations containing at least 10 cells per group were retained for downstream comparison. Differential expression analysis was performed using a pseudobulk DESeq2 strategy. Briefly, cells from the positive and negative groups were randomly partitioned into up to 10 pseudobulk samples per group, with each pseudobulk sample containing at least 10 cells. Gene-level counts were summed within each pseudobulk sample to generate pseudobulk count matrices. Genes with low expression were filtered before model fitting, and differential expression analysis was then conducted using the DESeq2 R package. Significant differentially expressed genes were defined as those with adjusted P < 0.05 and |log2 fold change| > 0.5. Downstream functional analyses Only comparisons yielding more than 10 significant differentially expressed genes were considered sufficiently informative for downstream functional interpretation. For these comparisons, GO and KEGG enrichment analyses were conducted using the clusterProfiler R package to identify overrepresented biological processes, cellular components, molecular functions, and signaling pathways among upregulated and downregulated genes [15]. Terms with adjusted P values below the predefined significance threshold were considered significantly enriched. GSVA was performed to evaluate pathway-level transcriptional differences between marker-positive and marker-negative glioma cell groups. Enrichment scores were calculated using the GSVA R package based on curated gene sets, and significantly altered pathways were identified by comparing pathway activity scores between groups [16]. CellChat analysis was performed to infer intercellular communication patterns associated with CHEK2- and NDRG1-defined malignant glioma subgroups [17]. Incoming and outgoing signaling patterns, as well as pathway-level communication probabilities, were compared between marker-positive and marker-negative glioma cells within the broader tumor microenvironment. Drug–gene interaction analysis using CTDbase To identify compounds reported to modulate the expression of the prognostic markers in glioblastoma (GBM), we queried the Comparative Toxicogenomics Database (CTDbase, https://ctdbase.org/, accessed on March 21, 2026) [18]. The four prognostic proteins from the multivariable Cox model—BCL2, CHEK2, CTNNB1, and NDRG1—were used as input genes. The disease term “glioblastoma” was applied as a context filter to retrieve GBM-related chemical–gene interaction records. For each gene, we extracted manually curated chemical–gene interactions with documented evidence of increased or decreased gene expression. Only compounds with directionally annotated expression effects in a GBM-related context were retained. The direction of regulation (upregulation or downregulation) was recorded for each gene–compound pair, and the resulting compound lists were summarized for downstream interpretation. Results Construction and validation of the protein-based prognostic signature To develop a protein-based prognostic model, we first performed univariate Cox regression analysis in the TCPA GBM cohort, which comprised 204 GBM samples and 223 quantified proteins. Among these, 28 proteins were significantly associated with overall survival (OS) at P < 0.05 (Supplementary Table 1, Fig. 1A). These candidate proteins were subsequently entered into a multivariable Cox regression model, which identified four prognostic proteins—BCL2, BETACATENIN, CHK2, and NDRG1_pT346—for construction of the final risk-score formula (Supplementary Table 2). Fig. 1Construction and prognostic evaluation of the four-protein risk signature in the TCPA GBM cohort. A Univariate Cox regression analysis of 223 quantified proteins in the TCPA GBM cohort (n = 204) identified 28 proteins significantly associated with overall survival (OS) (P < 0.05). The x-axis represents log2(hazard ratio), and the y-axis represents –log10(P value). Proteins significantly associated with poor prognosis are highlighted in red, and those associated with favorable prognosis are highlighted in green. The four proteins retained in the final multivariable model (BCL2, BETACATENIN, CHK2, and NDRG1_pT346) are indicated. B Kaplan–Meier analysis of OS according to the four-protein risk score. Patients were stratified into high-risk and low-risk groups using the median risk score as the cutoff. The high-risk group showed significantly poorer OS than the low-risk group. (C–F) Kaplan–Meier survival analyses for the individual model proteins: C BCL2, D BETACATENIN, E NDRG1_pT346, and F CHK2. Patients were dichotomized into high- and low-expression groups, and the observed survival trends were consistent with the directions of the coefficients in the multivariable Cox model. G Kaplan–Meier analysis of progression-free survival (PFS) according to the four-protein risk score, showing significantly shorter PFS in the high-risk group. H Time-dependent receiver operating characteristic (ROC) analysis of the four-protein risk model for predicting 1-, 3-, and 5-year OS, with area under the curve (AUC) values of 0.681, 0.809, and 0.916, respectively Construction and prognostic evaluation of the four-protein risk signature in the TCPA GBM cohort. A Univariate Cox regression analysis of 223 quantified proteins in the TCPA GBM cohort (n = 204) identified 28 proteins significantly associated with overall survival (OS) (P < 0.05). The x-axis represents log2(hazard ratio), and the y-axis represents –log10(P value). Proteins significantly associated with poor prognosis are highlighted in red, and those associated with favorable prognosis are highlighted in green. The four proteins retained in the final multivariable model (BCL2, BETACATENIN, CHK2, and NDRG1_pT346) are indicated. B Kaplan–Meier analysis of OS according to the four-protein risk score. Patients were stratified into high-risk and low-risk groups using the median risk score as the cutoff. The high-risk group showed significantly poorer OS than the low-risk group. (C–F) Kaplan–Meier survival analyses for the individual model proteins: C BCL2, D BETACATENIN, E NDRG1_pT346, and F CHK2. Patients were dichotomized into high- and low-expression groups, and the observed survival trends were consistent with the directions of the coefficients in the multivariable Cox model. G Kaplan–Meier analysis of progression-free survival (PFS) according to the four-protein risk score, showing significantly shorter PFS in the high-risk group. H Time-dependent receiver operating characteristic (ROC) analysis of the four-protein risk model for predicting 1-, 3-, and 5-year OS, with area under the curve (AUC) values of 0.681, 0.809, and 0.916, respectively The risk score for each patient was calculated as a weighted linear combination of the four protein expression values using the corresponding regression coefficients. Patients were then stratified into high-risk and low-risk groups according to the median risk score. Kaplan–Meier survival analysis showed that patients in the high-risk group had significantly poorer OS than those in the low-risk group (Fig. 1B). Consistently, the heatmap of the four model proteins demonstrated distinct expression patterns between the two risk groups (Supplementary Fig. 2A), while the risk score distribution and survival status plots showed that patients with higher risk scores were predominantly clustered in the high-risk group and generally exhibited shorter survival times (Supplementary Fig. 2B, C). We further examined the prognostic relevance of each individual model protein. Survival analyses for single proteins showed expression-associated survival trends that were consistent with the directions of the coefficients in the multivariable Cox model, supporting the internal consistency of the four-protein signature (Fig. 1C–F). In addition, progression-free survival (PFS) analysis demonstrated that the high-risk group also had significantly shorter PFS than the low-risk group (Fig. 1G). Time-dependent ROC analysis yielded AUC values of 0.681, 0.809, and 0.916 for predicting 1-, 3-, and 5-year OS, respectively, indicating favorable predictive performance of the model within this cohort (Fig. 1H). Independent prognostic value of the protein-based risk score To determine whether the protein-based risk score was independently associated with prognosis, univariate and multivariable Cox regression analyses were performed incorporating Age, Gender, Race, and risk score (Fig. 2A, B). In the multivariable model, Age remained significantly associated with overall survival (OS) (HR = 1.036, 95% CI: 1.021–1.051, P = 2.64 × 10⁻⁶), and the risk score was also identified as an independent prognostic factor (HR = 2.009, 95% CI: 1.527–2.644, P = 6.28 × 10⁻⁷), whereas Gender and Race were not statistically significant (Fig. 2B). These results indicate that the protein-based risk score retained prognostic significance after adjustment for the available clinical covariates. Fig. 2Independent prognostic value of the protein-based risk score and construction of the prognostic nomogram in the TCPA GBM cohort. A Forest plot of univariate Cox regression analysis incorporating Age, Gender, Race, and risk score in the TCPA GBM cohort. Hazard ratios (HRs) with 95% confidence intervals (CIs) and corresponding P values are shown for each variable. B Forest plot of multivariable Cox regression analysis including Age, Gender, Race, and risk score. In the adjusted model, Age and the protein-based risk score remained significantly associated with overall survival (OS), whereas Gender and Race were not statistically significant. C Prognostic nomogram constructed from the multivariable Cox regression model incorporating risk score, Age, Gender, and Race for individualized prediction of 1-, 3-, and 5-year OS probabilities in patients with glioblastoma. D Calibration curves for the nomogram at 1, 3, and 5 years, showing the agreement between nomogram-predicted survival probabilities and observed OS. The diagonal gray line indicates the ideal calibration Independent prognostic value of the protein-based risk score and construction of the prognostic nomogram in the TCPA GBM cohort. A Forest plot of univariate Cox regression analysis incorporating Age, Gender, Race, and risk score in the TCPA GBM cohort. Hazard ratios (HRs) with 95% confidence intervals (CIs) and corresponding P values are shown for each variable. B Forest plot of multivariable Cox regression analysis including Age, Gender, Race, and risk score. In the adjusted model, Age and the protein-based risk score remained significantly associated with overall survival (OS), whereas Gender and Race were not statistically significant. C Prognostic nomogram constructed from the multivariable Cox regression model incorporating risk score, Age, Gender, and Race for individualized prediction of 1-, 3-, and 5-year OS probabilities in patients with glioblastoma. D Calibration curves for the nomogram at 1, 3, and 5 years, showing the agreement between nomogram-predicted survival probabilities and observed OS. The diagonal gray line indicates the ideal calibration Schoenfeld residual-based tests were subsequently performed to assess the proportional hazards assumption of the multivariable model (Supplementary Fig. 3A-D). The results showed that the risk score did not significantly violate the proportional hazards assumption (P = 0.165), whereas Age showed evidence of non-proportionality (P = 0.0010). The global test was also significant (GLOBAL P = 0.0096), indicating that the multivariable clinical model should be interpreted with caution. Construction and evaluation of a prognostic nomogram To provide individualized survival estimation, a prognostic nomogram was constructed based on the multivariable model incorporating the protein-based risk score together with the available clinical variables (Age, Gender, and Race) (Fig. 2C). The nomogram was used to estimate 1-, 3-, and 5-year OS probabilities in patients with GBM. Calibration plots for 1-, 3-, and 5-year OS showed an overall agreement between nomogram-predicted and observed survival probabilities (Fig. 2D). These findings suggest that the nomogram may provide supportive individualized prognostic estimation within this cohort. However, because Schoenfeld residual-based testing indicated non-proportionality for Age and a significant global test in the multivariable clinical model, the nomogram-based predictions should be interpreted cautiously (Supplementary Fig. 3E-H). snRNA-seq data quality assessment and clustering After quality control, 13,515 high-quality nuclei from nine samples were retained for downstream analysis, and 3,000 highly variable genes were selected (Supplementary Figs. 4 and 5). Following Harmony integration, inter-sample batch effects were effectively mitigated while the overall data structure was preserved, leading to the identification of 12 clearly separated clusters in UMAP space (Supplementary Fig. 6). These findings demonstrated the robustness of the processed snRNA-seq dataset for subsequent cellular characterization. Cell-type annotation of the integrated snRNA-seq atlas Using cluster-specific marker genes, the integrated snRNA-seq dataset was first annotated into major cellular populations, including astrocytes, oligodendrocytes, oligodendrocytes-like cells, neurons-like cells, microglia, NK cells, TAMs, and a glioma-like population (Fig. 3A). At this stage, glioma cells were treated as a single major compartment rather than being further subdivided. These populations showed clear spatial separation on the UMAP embedding. Oligodendrocyte-lineage populations, including oligodendrocytes and oligodendrocytes-like cells, were mainly located on the left side of the embedding, whereas the glioma-like compartment occupied a broad central continuum. Microglia and TAMs formed a distinct myeloid compartment, while neurons-like cells and NK cells were distributed in separate peripheral regions (Fig. 3A). Fig. 3Cell-type annotation and inferCNV-guided refinement of the integrated snRNA-seq atlas. A UMAP embedding of the integrated snRNA-seq dataset showing major annotated cell populations, including astrocytes, oligodendrocytes, oligodendrocytes-like cells, neurons-like cells, microglia, NK cells, TAMs, and a glioma-like compartment. Glioma cells were initially treated as a single major population. Oligodendrocyte-lineage populations were predominantly located on the left side of the embedding, whereas the glioma-like compartment occupied a broad central continuum. Microglia and TAMs formed a distinct myeloid compartment, and neurons-like cells and NK cells were distributed in separate peripheral regions. B InferCNV analysis using neurons-like cells and oligodendrocytes as references highlighted broad and recurrent large-scale copy-number variation patterns in the glioma-like compartment, supporting its assignment to the malignant glioma population. Non-malignant microenvironmental populations lacked comparable CNV patterns. C Subdivision of the malignant glioma compartment into three transcriptional states—Glioma_MES, Glioma_PN, and Glioma_CL—based on inferCNV-supported clustering and curated marker programs. Glioma_CL cells were distributed along the central elongated glioma trajectory, Glioma_PN cells were enriched in the upper region, and Glioma_MES cells were located in the lower-central area adjacent to astrocyte-rich regions Cell-type annotation and inferCNV-guided refinement of the integrated snRNA-seq atlas. A UMAP embedding of the integrated snRNA-seq dataset showing major annotated cell populations, including astrocytes, oligodendrocytes, oligodendrocytes-like cells, neurons-like cells, microglia, NK cells, TAMs, and a glioma-like compartment. Glioma cells were initially treated as a single major population. Oligodendrocyte-lineage populations were predominantly located on the left side of the embedding, whereas the glioma-like compartment occupied a broad central continuum. Microglia and TAMs formed a distinct myeloid compartment, and neurons-like cells and NK cells were distributed in separate peripheral regions. B InferCNV analysis using neurons-like cells and oligodendrocytes as references highlighted broad and recurrent large-scale copy-number variation patterns in the glioma-like compartment, supporting its assignment to the malignant glioma population. Non-malignant microenvironmental populations lacked comparable CNV patterns. C Subdivision of the malignant glioma compartment into three transcriptional states—Glioma_MES, Glioma_PN, and Glioma_CL—based on inferCNV-supported clustering and curated marker programs. Glioma_CL cells were distributed along the central elongated glioma trajectory, Glioma_PN cells were enriched in the upper region, and Glioma_MES cells were located in the lower-central area adjacent to astrocyte-rich regions InferCNV-guided refinement of the malignant glioma compartment To further distinguish malignant glioma cells from non-malignant populations, inferCNV analysis was applied to the preliminarily annotated atlas using neurons-like cells and oligodendrocytes as reference populations. The reference populations showed relatively stable copy-number profiles, whereas the glioma-like compartment displayed broader and recurrent large-scale copy-number variation patterns across multiple chromosomal regions, supporting its assignment to the malignant glioma compartment (Fig. 3B). In contrast, the other annotated microenvironmental populations lacked comparable large-scale CNV abnormalities and were therefore retained as non-malignant cell populations. Based on this inferCNV-supported refinement, the malignant glioma compartment was subsequently subdivided into three transcriptional states, namely Glioma_MES, Glioma_PN, and Glioma_CL (Fig. 3C). These three glioma states showed both spatial continuity and regional enrichment within the glioma manifold, consistent with substantial intratumoral transcriptional heterogeneity. Specifically, Glioma_CL was distributed along the elongated central glioma trajectory, Glioma_PN was enriched in the upper part of the glioma manifold, and Glioma_MES was mainly located in the lower-central region adjacent to the astrocyte-rich area (Fig. 3C). The dot plot of representative marker genes further supported the reliability of this annotation framework by showing state-specific expression patterns across the final cell populations (Supplementary Fig. 7). In addition, the relative proportions of annotated cell populations and malignant glioma states varied substantially across samples, indicating marked inter-sample compositional heterogeneity (Supplementary Fig. 8). Together, these analyses established a refined single-cell atlas containing both major non-malignant neural and immune populations and state-resolved malignant glioma cells for downstream marker-centered analyses. Cell-type-specific distribution of prognostic markers Further analysis of the single-cell atlas revealed distinct cell-type preferences for the four prognostic markers. The top three cell populations with the highest expression of BCL2 were oligodendrocytes-like cells, oligodendrocytes, and microglia. CHEK2 was preferentially expressed in microglia, TAMs, and Glioma_PN. The highest-expressing populations for CTNNB1 were TAMs, microglia, and Glioma_CL, whereas NDRG1 was most strongly enriched in Glioma_CL, oligodendrocytes, and oligodendrocytes-like cells. Overall, BCL2 and NDRG1 were mainly associated with oligodendroglial-related populations and selected malignant glioma states, while CTNNB1 and CHEK2 were more enriched in myeloid-associated populations, indicating distinct cellular contexts for these prognostic markers within the glioma ecosystem (Fig. 4, Supplementary Fig. 9). Fig. 4Cell-type-specific distribution of the four prognostic markers in the refined snRNA-seq atlas. A BCL2 expression was highest in oligodendrocytes-like cells, followed by oligodendrocytes and microglia. B CHEK2 expression was preferentially enriched in microglia, tumor-associated macrophages (TAMs), and Glioma_PN cells. C CTNNB1 showed the highest expression in TAMs, followed by microglia and Glioma_CL cells. D NDRG1 was most strongly enriched in Glioma_CL, oligodendrocytes, and oligodendrocytes-like cells Cell-type-specific distribution of the four prognostic markers in the refined snRNA-seq atlas. A BCL2 expression was highest in oligodendrocytes-like cells, followed by oligodendrocytes and microglia. B CHEK2 expression was preferentially enriched in microglia, tumor-associated macrophages (TAMs), and Glioma_PN cells. C CTNNB1 showed the highest expression in TAMs, followed by microglia and Glioma_CL cells. D NDRG1 was most strongly enriched in Glioma_CL, oligodendrocytes, and oligodendrocytes-like cells CHEK2- and NDRG1-associated transcriptional alterations in malignant glioma cells To further explore the single-cell functional contexts of the four prognostic markers, we first identified the top three cell populations with the highest expression for each marker and then performed marker-positive versus marker-negative differential expression analysis within these major expressing populations. Among the four key proteins, only CHEK2 and NDRG1 yielded more than 10 significant differentially expressed genes when malignant glioma cells were stratified into positive and negative groups. Therefore, subsequent downstream analyses were focused on these two markers in the glioma compartment.(Supplementary Fig. 10). Comparison between CHEK2-positive and CHEK2-negative malignant glioma cells identified 876 differentially expressed genes (DEGs), including 711 upregulated and 165 downregulated genes in the CHEK2-positive group (Fig. 5A). GO enrichment analysis showed that the upregulated genes were mainly associated with chromosome segregation, DNA replication, DNA-templated replication, nuclear division, mitotic progression, and double-strand break repair, whereas the downregulated genes were primarily related to responses to hypoxia and oxygen fluctuation, cytoplasmic translation, peptide transport/secretion, and adhesion-related processes (Fig. 5B, C). Consistently, KEGG analysis further showed that the upregulated genes were enriched in the cell cycle, DNA replication, homologous recombination, Fanconi anemia pathway, mismatch repair, base excision repair, and nucleotide excision repair, while the downregulated genes were mainly enriched in ribosome-associated translation and reactive oxygen species-related stress pathways (Fig. 5D, E). In agreement with these findings, GSVA demonstrated that CHEK2-positive glioma cells were significantly associated with DNA damage response, checkpoint regulation, G2/M transition, replicative senescence, radiation response, and p53-related signaling (Fig. 5F). Fig. 5CHEK2-associated transcriptional alterations in malignant glioma cells. A Volcano plot showing differentially expressed genes (DEGs) between CHEK2-positive and CHEK2-negative malignant glioma cells. A total of 876 DEGs were identified, including 711 upregulated and 165 downregulated genes in the CHEK2-positive group. B, C Gene Ontology (GO) enrichment analysis of upregulated and downregulated DEGs in CHEK2-positive malignant glioma cells. Upregulated genes were predominantly enriched in biological processes related to chromosome segregation, DNA replication, DNA-templated replication, nuclear division, mitotic progression, and double-strand break repair, whereas downregulated genes were mainly associated with responses to hypoxia and oxygen fluctuation, cytoplasmic translation, peptide transport/secretion, and adhesion-related processes. D, E Kyoto Encyclopedia of Genes and Genomes (KEGG) pathway enrichment analysis of upregulated and downregulated DEGs. Upregulated genes were mainly enriched in cell cycle, DNA replication, homologous recombination, Fanconi anemia pathway, mismatch repair, base excision repair, and nucleotide excision repair, whereas downregulated genes were primarily enriched in ribosome-associated translation and reactive oxygen species-related stress pathways. F Gene set variation analysis (GSVA) comparing CHEK2-positive and CHEK2-negative malignant glioma cells. CHEK2-positive glioma cells showed higher enrichment scores for pathways related to DNA damage response, checkpoint regulation, G2/M transition, replicative senescence, radiation response, and p53-associated signaling CHEK2-associated transcriptional alterations in malignant glioma cells. A Volcano plot showing differentially expressed genes (DEGs) between CHEK2-positive and CHEK2-negative malignant glioma cells. A total of 876 DEGs were identified, including 711 upregulated and 165 downregulated genes in the CHEK2-positive group. B, C Gene Ontology (GO) enrichment analysis of upregulated and downregulated DEGs in CHEK2-positive malignant glioma cells. Upregulated genes were predominantly enriched in biological processes related to chromosome segregation, DNA replication, DNA-templated replication, nuclear division, mitotic progression, and double-strand break repair, whereas downregulated genes were mainly associated with responses to hypoxia and oxygen fluctuation, cytoplasmic translation, peptide transport/secretion, and adhesion-related processes. D, E Kyoto Encyclopedia of Genes and Genomes (KEGG) pathway enrichment analysis of upregulated and downregulated DEGs. Upregulated genes were mainly enriched in cell cycle, DNA replication, homologous recombination, Fanconi anemia pathway, mismatch repair, base excision repair, and nucleotide excision repair, whereas downregulated genes were primarily enriched in ribosome-associated translation and reactive oxygen species-related stress pathways. F Gene set variation analysis (GSVA) comparing CHEK2-positive and CHEK2-negative malignant glioma cells. CHEK2-positive glioma cells showed higher enrichment scores for pathways related to DNA damage response, checkpoint regulation, G2/M transition, replicative senescence, radiation response, and p53-associated signaling Comparison between NDRG1-positive and NDRG1-negative malignant glioma cells identified 1,585 DEGs, including 832 upregulated and 753 downregulated genes in the NDRG1-positive group (Fig. 6A). GO enrichment analysis showed that the upregulated genes were mainly involved in responses to hypoxia and decreased oxygen levels, carbohydrate catabolism, pyruvate metabolism, steroid hormone response, and extracellular matrix- and vesicle-associated processes, whereas the downregulated genes were predominantly enriched in DNA replication, chromosome segregation, cell-cycle phase transition, and homologous recombination-mediated DNA repair (Fig. 6B, C). KEGG analysis further revealed that the upregulated genes were enriched in the HIF-1 signaling pathway, glycolysis/gluconeogenesis, central carbon metabolism in cancer, AMPK signaling, FoxO signaling, MAPK signaling, autophagy, mitophagy, and protein processing in the endoplasmic reticulum, while the downregulated genes were mainly concentrated in cell cycle, cellular senescence, homologous recombination, DNA replication, and the Fanconi anemia pathway (Fig. 6D, E). Consistently, GSVA showed that NDRG1-positive glioma cells were characterized by enhanced glucose uptake, glycolysis, hypoxia response, oxygen-level adaptation, and mTORC1-related metabolic programs (Fig. 6F). Fig. 6NDRG1-associated transcriptional alterations in malignant glioma cells. A Volcano plot showing differentially expressed genes (DEGs) between NDRG1-positive and NDRG1-negative malignant glioma cells. A total of 1,585 DEGs were identified, including 832 upregulated and 753 downregulated genes in the NDRG1-positive group. B, C Gene Ontology (GO) enrichment analysis of upregulated and downregulated DEGs in NDRG1-positive malignant glioma cells. Upregulated genes were mainly involved in responses to hypoxia and decreased oxygen levels, carbohydrate catabolism, pyruvate metabolism, steroid hormone response, and extracellular matrix- and vesicle-associated processes, whereas downregulated genes were predominantly enriched in DNA replication, chromosome segregation, cell-cycle phase transition, and homologous recombination-mediated DNA repair. D, E Kyoto Encyclopedia of Genes and Genomes (KEGG) pathway enrichment analysis of upregulated and downregulated DEGs. Upregulated genes were enriched in HIF-1 signaling, glycolysis/gluconeogenesis, central carbon metabolism in cancer, AMPK signaling, FoxO signaling, MAPK signaling, autophagy, mitophagy, and protein processing in the endoplasmic reticulum, whereas downregulated genes were mainly enriched in cell cycle, cellular senescence, homologous recombination, DNA replication, and Fanconi anemia pathway. F Gene set variation analysis (GSVA) comparing NDRG1-positive and NDRG1-negative malignant glioma cells. NDRG1-positive glioma cells showed increased pathway activity related to glucose uptake, glycolysis, hypoxia response, oxygen-level adaptation, and mTORC1-associated metabolic programs NDRG1-associated transcriptional alterations in malignant glioma cells. A Volcano plot showing differentially expressed genes (DEGs) between NDRG1-positive and NDRG1-negative malignant glioma cells. A total of 1,585 DEGs were identified, including 832 upregulated and 753 downregulated genes in the NDRG1-positive group. B, C Gene Ontology (GO) enrichment analysis of upregulated and downregulated DEGs in NDRG1-positive malignant glioma cells. Upregulated genes were mainly involved in responses to hypoxia and decreased oxygen levels, carbohydrate catabolism, pyruvate metabolism, steroid hormone response, and extracellular matrix- and vesicle-associated processes, whereas downregulated genes were predominantly enriched in DNA replication, chromosome segregation, cell-cycle phase transition, and homologous recombination-mediated DNA repair. D, E Kyoto Encyclopedia of Genes and Genomes (KEGG) pathway enrichment analysis of upregulated and downregulated DEGs. Upregulated genes were enriched in HIF-1 signaling, glycolysis/gluconeogenesis, central carbon metabolism in cancer, AMPK signaling, FoxO signaling, MAPK signaling, autophagy, mitophagy, and protein processing in the endoplasmic reticulum, whereas downregulated genes were mainly enriched in cell cycle, cellular senescence, homologous recombination, DNA replication, and Fanconi anemia pathway. F Gene set variation analysis (GSVA) comparing NDRG1-positive and NDRG1-negative malignant glioma cells. NDRG1-positive glioma cells showed increased pathway activity related to glucose uptake, glycolysis, hypoxia response, oxygen-level adaptation, and mTORC1-associated metabolic programs CellChat analysis of CHEK2- and NDRG1-associated malignant glioma cells CellChat analysis was performed to compare the intercellular communication patterns of CHEK2-positive (Gli_CHEK2+), CHEK2-negative (Gli_CHEK2−), NDRG1-positive (Gli_NDRG1+), and NDRG1-negative (Gli_NDRG1−) malignant glioma cells within the tumor microenvironment. Overall, all four glioma subpopulations participated in broad communication networks involving astrocytes, neurons-like cells, oligodendrocytes, oligodendrocytes-like cells, microglia, NK cells, and TAMs, although their incoming and outgoing signaling profiles were not identical. For CHEK2-associated glioma cells, both Gli_CHEK2 + and Gli_CHEK2 − cells showed extensive incoming signaling across pathways related to cell adhesion, neural interaction, and extracellular matrix-associated communication, including NCAM, NRXN, CADM, PTN, APP, NRG, CLDN, CNTN, CDH, COLLAGEN, SPP1, and JAM, together with additional contributions from FGF, IGF, WNT, TGFβ, EPHA, FN1, THBS, ANGPTL, and EGF (Fig. 7). In the outgoing direction, both CHEK2-defined glioma populations also displayed broad ligand-associated signaling involving adhesion molecules, neural interaction pathways, extracellular matrix-related signaling, and growth factor-associated communication. Compared with Gli_CHEK2 − cells, Gli_CHEK2 + cells showed broader participation across selected incoming and outgoing pathways. Fig. 7CellChat analysis of intercellular communication patterns associated with CHEK2-positive and CHEK2-negative malignant glioma cells. A Heatmap showing the relative strength of incoming signaling patterns across major cell populations and marker-defined glioma subgroups, including astrocytes (Ast), CHEK2-negative malignant glioma cells (Gli_CHEK2−), CHEK2-positive malignant glioma cells (Gli_CHEK2+), microglia (Micro), neurons-like cells (Neu_like), NK cells, oligodendrocytes (Oligo), oligodendrocytes-like cells (Oligo-like), and tumor-associated macrophages (TAMs). Both Gli_CHEK2 + and Gli_CHEK2 − cells exhibited broad incoming signaling participation, particularly in pathways related to cell adhesion, neural interaction, and extracellular matrix-associated communication, including NCAM, NRXN, CADM, PTN, APP, NRG, CLDN, CNTN, CDH, COLLAGEN, SPP1, and JAM, together with additional contributions from FGF, IGF, WNT, TGFβ, EPHA, FN1, THBS, ANGPTL, and EGF. B Heatmap showing the relative strength of outgoing signaling patterns across the same cell populations. Both CHEK2-defined glioma subgroups displayed broad outgoing signaling involving adhesion-related, neural-interaction, extracellular matrix-associated, and growth factor-related pathways, whereas Gli_CHEK2 + cells showed broader participation across selected incoming and outgoing pathways than Gli_CHEK2 − cells CellChat analysis of intercellular communication patterns associated with CHEK2-positive and CHEK2-negative malignant glioma cells. A Heatmap showing the relative strength of incoming signaling patterns across major cell populations and marker-defined glioma subgroups, including astrocytes (Ast), CHEK2-negative malignant glioma cells (Gli_CHEK2−), CHEK2-positive malignant glioma cells (Gli_CHEK2+), microglia (Micro), neurons-like cells (Neu_like), NK cells, oligodendrocytes (Oligo), oligodendrocytes-like cells (Oligo-like), and tumor-associated macrophages (TAMs). Both Gli_CHEK2 + and Gli_CHEK2 − cells exhibited broad incoming signaling participation, particularly in pathways related to cell adhesion, neural interaction, and extracellular matrix-associated communication, including NCAM, NRXN, CADM, PTN, APP, NRG, CLDN, CNTN, CDH, COLLAGEN, SPP1, and JAM, together with additional contributions from FGF, IGF, WNT, TGFβ, EPHA, FN1, THBS, ANGPTL, and EGF. B Heatmap showing the relative strength of outgoing signaling patterns across the same cell populations. Both CHEK2-defined glioma subgroups displayed broad outgoing signaling involving adhesion-related, neural-interaction, extracellular matrix-associated, and growth factor-related pathways, whereas Gli_CHEK2 + cells showed broader participation across selected incoming and outgoing pathways than Gli_CHEK2 − cells For NDRG1-associated glioma cells, both Gli_NDRG1 + and Gli_NDRG1 − cells retained broad incoming signaling patterns, with major contributions from pathways associated with cell adhesion, neural interaction, and matrix-related communication, including NCAM, NRXN, CADM, PTN, APP, NRG, CLDN, CNTN, CDH, COLLAGEN, SPP1, JAM, IGF, TGFβ, WNT, ANGPTL, FN1, and THBS (Fig. 8). In contrast, the difference between the two groups was more evident in outgoing signaling. Gli_NDRG1 -cells showed broad outgoing participation across multiple pathways, whereas Gli_NDRG1 + cells displayed a more restricted outgoing signaling repertoire, with only a limited subset of pathways remaining prominent. Fig. 8CellChat analysis of intercellular communication patterns associated with NDRG1-positive and NDRG1-negative malignant glioma cells. A Heatmap showing the relative strength of incoming signaling patterns across major cell populations and marker-defined glioma subgroups, including astrocytes (Ast), NDRG1-negative malignant glioma cells (Gli_NDRG1−), NDRG1-positive malignant glioma cells (Gli_NDRG1+), microglia (Micro), neuron-like cells (Neu_like), NK cells, oligodendrocytes (Oligo), oligodendrocyte-like cells (Oligo-like), and tumor-associated macrophages (TAMs). Both Gli_NDRG1+ and Gli_NDRG1− cells exhibited broad incoming signaling participation, with major contributions from pathways related to cell adhesion, neural interaction, and extracellular matrix-associated communication, including NCAM, NRXN, CADM, PTN, APP, NRG, CLDN, CNTN, CDH, COLLAGEN, SPP1, and JAM, together with additional involvement of IGF, TGFβ, WNT, ANGPTL, FN1, and THBS. B Heatmap showing the relative strength of outgoing signaling patterns across the same cell populations. Compared with the incoming signaling profile, the difference between the two NDRG1-defined glioma subgroups was more evident in outgoing communication. Gli_NDRG1− cells showed broad outgoing participation across multiple signaling pathways, whereas Gli_NDRG1+ cells displayed a more restricted outgoing signaling repertoire, with only a limited subset of pathways remaining prominent CellChat analysis of intercellular communication patterns associated with NDRG1-positive and NDRG1-negative malignant glioma cells. A Heatmap showing the relative strength of incoming signaling patterns across major cell populations and marker-defined glioma subgroups, including astrocytes (Ast), NDRG1-negative malignant glioma cells (Gli_NDRG1−), NDRG1-positive malignant glioma cells (Gli_NDRG1+), microglia (Micro), neuron-like cells (Neu_like), NK cells, oligodendrocytes (Oligo), oligodendrocyte-like cells (Oligo-like), and tumor-associated macrophages (TAMs). Both Gli_NDRG1+ and Gli_NDRG1− cells exhibited broad incoming signaling participation, with major contributions from pathways related to cell adhesion, neural interaction, and extracellular matrix-associated communication, including NCAM, NRXN, CADM, PTN, APP, NRG, CLDN, CNTN, CDH, COLLAGEN, SPP1, and JAM, together with additional involvement of IGF, TGFβ, WNT, ANGPTL, FN1, and THBS. B Heatmap showing the relative strength of outgoing signaling patterns across the same cell populations. Compared with the incoming signaling profile, the difference between the two NDRG1-defined glioma subgroups was more evident in outgoing communication. Gli_NDRG1− cells showed broad outgoing participation across multiple signaling pathways, whereas Gli_NDRG1+ cells displayed a more restricted outgoing signaling repertoire, with only a limited subset of pathways remaining prominent At the network level, neural- and glial-lineage populations were prominently represented in adhesion- and neural interaction-related communication, whereas myeloid populations, particularly TAMs, were more strongly represented in immune-, inflammatory-, and matrix-associated signaling modules. Overall, CHEK2- and NDRG1-associated malignant glioma cells were both connected to complex neural–glial–immune communication networks, but differed in their relative incoming and outgoing signaling patterns. Identification of CTDbase-derived compounds associated with prognostic markers in GBM Using glioblastoma as the disease filter, CTDbase analysis identified curated chemical–gene interactions for all four prognostic markers. Specifically, 25 compounds were associated with increased BCL2 expression, 8 compounds with increased CHEK2 expression, 16 compounds with increased CTNNB1 expression, and 9 compounds with decreased NDRG1 expression in GBM-related contexts (Supplementary Tables 3–6). Several compounds recurred across multiple marker-associated lists, including arsenic trioxide, cisplatin, trichostatin A, valproic acid, sodium arsenite, testosterone, and tretinoin, whereas other compounds were more gene-specific. These results summarize the currently curated evidence for compound-associated modulation of the four prognostic markers in GBM. Discussion GBM is among the most aggressive and heterogeneous malignant tumors of the central nervous system, characterized by poor prognosis and limited treatment options. The advent of snRNA-seq has substantially advanced our understanding of the GBM tumor microenvironment, intratumoral heterogeneity, and intercellular communication. However, effective molecular markers for individualized prognostic assessment and targeted therapy in GBM remain lacking. In this study, we integrated proteomic and snRNA-seq data to construct a prognostic scoring model based on four proteins: BCL2, CHEK2, CTNNB1, and NDRG1. Within the multivariable model, BCL2, CHEK2, and CTNNB1 contributed in a direction consistent with more favorable prognosis, whereas NDRG1 was associated with poorer outcome. At the same time, the composite risk score—calculated as a weighted combination of all four proteins—remained significantly associated with worse survival in the high-risk group. Accordingly, the biological interpretation of individual proteins should be distinguished from that of the integrated multivariable risk model. These findings were supported by multivariate Cox regression analysis and remained statistically significant after adjusting for clinical covariates, indicating their potential as independent prognostic markers. Previous studies have reported diverse roles for these proteins in cancer biology. BCL2, a classical anti-apoptotic molecule, is often regarded as an oncogene in various tumors [19, 20]. In our study, however, BCL2 expression was mainly enriched in oligodendroglial-related populations rather than broadly distributed across all malignant glioma states, suggesting that its prognostic contribution in this cohort may reflect a more context-dependent role than a simple pan-oncogenic effect. CHEK2, a key DNA damage checkpoint protein, has well-established tumor-suppressive roles [21–23]. Consistent with this biology, CHEK2-positive malignant glioma cells in our analysis were enriched for DNA damage response, chromosome segregation, DNA replication, and checkpoint-related programs, together with p53-associated signaling, indicating that CHEK2 expression marked a transcriptional state linked to genome surveillance and proliferative control rather than unchecked tumor-promoting activity. CTNNB1 (β-catenin), a central component of Wnt signaling, plays critical roles in both neural development and gliomagenesis [24–26]. Our single-cell analysis showed that CTNNB1 was preferentially expressed in myeloid-associated populations, particularly TAMs and microglia, and in a subset of glioma cells, indicating that its biological significance in GBM may depend strongly on cell type and microenvironmental context. In contrast, NDRG1 has been frequently associated with tumor progression, invasiveness, and radioresistance in GBM [27, 28]. Consistent with previous reports, NDRG1-positive malignant glioma cells in our analysis displayed enrichment in hypoxia response, glycolysis, HIF-1 signaling, autophagy/mitophagy, and central carbon metabolism, together with relative suppression of DNA replication and cell-cycle-related programs, supporting an association with a hypoxia-adaptive and metabolically reprogrammed glioma state. To place these findings in a broader GBM signaling context, it is important to note that EGFR is among the most frequently altered receptor tyrosine kinases in GBM and remains a central node in glioma biology [29]. EGFR-centered signaling intersects with pathways relevant to both apoptosis regulation and DNA damage responses, thereby providing a plausible biological framework linking EGFR-associated signaling architecture with BCL2- and CHEK2-related programs in glioma cells. In addition, developmental and injury-associated studies have shown that EGFR expression is tightly connected to tissue homeostasis, wound response, and proliferative signaling [29], which is relevant to the adaptive and regenerative-like transcriptional features often observed in GBM. The extracellular matrix (ECM) and cell–matrix adhesion context is also highly relevant to the interpretation of our data. Recent work has shown that injury-associated neural niches undergo marked upregulation of ECM- and adhesion-related genes, with coordinated remodeling of the regenerative microenvironment [30]. This provides a useful neurobiological framework for understanding why ECM remodeling, adhesion signaling, vesicle-associated processes, and matrix-related communication pathways repeatedly emerged in our single-cell analyses, especially in NDRG1-associated glioma states and in the inferred communication networks of the tumor microenvironment. Rather than being viewed only as generic stromal signatures, these pathways may reflect a broader injury-like or reactive tissue program that is co-opted during GBM invasion and adaptation. Furthermore, we performed a drug-gene interaction analysis using CTDbase to identify pharmacologic agents reported to modulate the expression of BCL2, CHEK2, CTNNB1, and NDRG1 in GBM-related contexts. This analysis should be interpreted as a curated, hypothesis-generating overview of chemical–gene interactions rather than direct evidence of therapeutic efficacy in GBM. Several compounds were identified across multiple marker-associated lists, including Arsenic Trioxide, Cisplatin, Trichostatin A, Valproic Acid, Sodium Arsenite, Testosterone, and Tretinoin [21–35]. Notably, some compounds appeared across more than one marker-associated set, suggesting potential overlap in their reported regulatory scope, whereas others were more gene-specific. These results provide a structured starting point for future pharmacogenomic prioritization, but experimental validation will be required before any therapeutic implications can be established. Despite the multi-omics integration and the comprehensive downstream analyses performed in this study, several limitations should be acknowledged. First, the prognostic model was derived from public proteomic data and its validation relied on a limited public snRNA-seq cohort, and the absence of independent, multi-center clinical cohorts with harmonized treatment and follow-up information may constrain generalizability and clinical translation. In addition, the TCPA/TCGA GBM cohort represents a historically defined GBM population and has not been fully reannotated according to the 2021 WHO classification, which should be considered when interpreting the broader applicability of the findings. Second, integrating bulk proteomics with single-nucleus transcriptomics can introduce cross-platform and batch-related biases, and mRNA-level signals do not necessarily reflect protein abundance, post-translational modifications, or pathway activity in a one-to-one manner. Third, the CellChat-inferred communication networks are computational predictions that depend on curated ligand–receptor databases and modeling assumptions, so the identified signaling differences should be interpreted as hypothesis-generating and require experimental confirmation. Finally, although the CTDbase-based drug–gene interaction results provide a useful starting point for candidate prioritization, they do not establish therapeutic efficacy in GBM, and the mechanistic roles of the prognostic proteins within specific microenvironmental contexts still need to be validated through targeted functional experiments and prospective studies. In conclusion, our study highlights the distinct functional spectra and intercellular communication profiles of BCL2, CHEK2, CTNNB1, and NDRG1 in GBM. By integrating proteomic prognostic modeling with single-nucleus transcriptomic analysis, we further localized these markers to specific cellular contexts and refined their association with malignant glioma states and microenvironmental compartments. These proteins serve as potential biomarkers to elucidate tumor heterogeneity and inform targeted therapeutic strategies. Our findings provide a biological foundation for future precision interventions in GBM. Supplementary Information Below is the link to the electronic supplementary material. Supplementary Material 1: Fig. 1. Study flow chart of this study. Supplementary Material 1: Fig. 1. Study flow chart of this study. Supplementary Material 2: Fig. 2. Expression distribution of the four model proteins and risk-score landscape in the TCPA GBM cohort. (A) Heatmap showing the expression patterns of the four prognostic proteins (NDRG1_pT346, CHK2, BCL2, and BETACATENIN) across patients stratified by the protein-based risk score. Samples are grouped into high-risk and low-risk categories according to the median risk score. Relative protein expression levels are displayed using a scaled color gradient. (B) Distribution of the protein-based risk score among all patients in the TCPA GBM cohort, arranged in ascending order of risk score. Patients classified as high risk and low risk are indicated in different colors, and the dashed lines denote the median cutoff used for risk stratification. (C) Survival status plot of the TCPA GBM cohort ranked by increasing risk score. Each point represents an individual patient, with the y-axis indicating survival time (years). Patients are colored according to survival status (dead or alive), illustrating the relationship between increasing risk score and survival outcome. Supplementary Material 2: Fig. 2. Expression distribution of the four model proteins and risk-score landscape in the TCPA GBM cohort. (A) Heatmap showing the expression patterns of the four prognostic proteins (NDRG1_pT346, CHK2, BCL2, and BETACATENIN) across patients stratified by the protein-based risk score. Samples are grouped into high-risk and low-risk categories according to the median risk score. Relative protein expression levels are displayed using a scaled color gradient. (B) Distribution of the protein-based risk score among all patients in the TCPA GBM cohort, arranged in ascending order of risk score. Patients classified as high risk and low risk are indicated in different colors, and the dashed lines denote the median cutoff used for risk stratification. (C) Survival status plot of the TCPA GBM cohort ranked by increasing risk score. Each point represents an individual patient, with the y-axis indicating survival time (years). Patients are colored according to survival status (dead or alive), illustrating the relationship between increasing risk score and survival outcome. Supplementary Material 3: Fig. 3. Assessment of the proportional hazards assumption for the protein-based multivariable Cox model. (A–D) Residuals for individual proteins in the risk model (BCL2, BETACATENIN, CHK2, NDRG1_pT346) plotted against follow-up time. Solid lines indicate smoothed estimates of residuals, and dashed lines represent 95% confidence intervals. Deviations from a horizontal line suggest potential violation of the proportional hazards assumption. (E–G) Residuals for clinical covariates Age (E), Gender (F), and Race (G) plotted similarly to assess time-dependent effects. Age showed a minor non-proportional pattern, whereas Gender and Race residuals remained approximately horizontal over time. (H) Residuals for the composite protein-based risk score plotted over follow-up time, indicating no significant violation of the proportional hazards assumption. Supplementary Material 3: Fig. 3. Assessment of the proportional hazards assumption for the protein-based multivariable Cox model. (A–D) Residuals for individual proteins in the risk model (BCL2, BETACATENIN, CHK2, NDRG1_pT346) plotted against follow-up time. Solid lines indicate smoothed estimates of residuals, and dashed lines represent 95% confidence intervals. Deviations from a horizontal line suggest potential violation of the proportional hazards assumption. (E–G) Residuals for clinical covariates Age (E), Gender (F), and Race (G) plotted similarly to assess time-dependent effects. Age showed a minor non-proportional pattern, whereas Gender and Race residuals remained approximately horizontal over time. (H) Residuals for the composite protein-based risk score plotted over follow-up time, indicating no significant violation of the proportional hazards assumption. Supplementary Material 4: Fig. 4. Quality control summary of snRNA-seq nuclei across samples. Initial quality control was performed on 16,046 nuclei from nine GBM samples. After filtering low-quality nuclei, 13,515 nuclei were retained, corresponding to the removal of 2,531 nuclei (15.8%). The violin plots show distributions of the number of detected genes per nucleus (nFeature_RNA), the total transcript counts per nucleus (nCount_RNA), and the number of nuclei per sample before and after filtering, demonstrating improved consistency across samples. Supplementary Material 4: Fig. 4. Quality control summary of snRNA-seq nuclei across samples. Initial quality control was performed on 16,046 nuclei from nine GBM samples. After filtering low-quality nuclei, 13,515 nuclei were retained, corresponding to the removal of 2,531 nuclei (15.8%). The violin plots show distributions of the number of detected genes per nucleus (nFeature_RNA), the total transcript counts per nucleus (nCount_RNA), and the number of nuclei per sample before and after filtering, demonstrating improved consistency across samples. Supplementary Material 5: Fig. 5. Detailed quality assessment of snRNA-seq data. (A) Violin plots of nFeature_RNA and nCount_RNA across samples after filtering indicate improved uniformity in feature detection and library size. (B) Ambient RNA contamination estimated by decontX shows generally low contamination levels across nuclei (mean = 0.054). (C) Doublet detection using scDblFinder identified 689 doublets (red) and 12,655 singlets (blue), highlighting the purity of retained nuclei for downstream analysis. (D) Selection of 3,000 highly variable genes for subsequent analyses, shown as the distribution of standardized variance versus average expression, ensured that the most informative genes were retained for dimensionality reduction, clustering, and downstream cell-type-specific analyses. Supplementary Material 5: Fig. 5. Detailed quality assessment of snRNA-seq data. (A) Violin plots of nFeature_RNA and nCount_RNA across samples after filtering indicate improved uniformity in feature detection and library size. (B) Ambient RNA contamination estimated by decontX shows generally low contamination levels across nuclei (mean = 0.054). (C) Doublet detection using scDblFinder identified 689 doublets (red) and 12,655 singlets (blue), highlighting the purity of retained nuclei for downstream analysis. (D) Selection of 3,000 highly variable genes for subsequent analyses, shown as the distribution of standardized variance versus average expression, ensured that the most informative genes were retained for dimensionality reduction, clustering, and downstream cell-type-specific analyses. Supplementary Material 6: Fig. 6. Batch correction, clustering resolution optimization, and UMAP visualization of the snRNA-seq dataset. (A) Principal component analysis (PCA) of nuclei before batch correction revealed partial separation by sample, indicating inter-sample technical variation. After Harmony integration, nuclei from all samples were more evenly mixed while preserving the overall structure of the dataset, demonstrating effective batch correction. (B) Clustering tree analysis across multiple resolutions informed the selection of resolution 0.5 for final clustering. (C) UMAP embedding of nuclei colored by cluster identity (0–11) shows clear separation of 12 transcriptionally distinct clusters, supporting the quality and robustness of the processed snRNA-seq dataset for downstream cellular characterization. Supplementary Material 6: Fig. 6. Batch correction, clustering resolution optimization, and UMAP visualization of the snRNA-seq dataset. (A) Principal component analysis (PCA) of nuclei before batch correction revealed partial separation by sample, indicating inter-sample technical variation. After Harmony integration, nuclei from all samples were more evenly mixed while preserving the overall structure of the dataset, demonstrating effective batch correction. (B) Clustering tree analysis across multiple resolutions informed the selection of resolution 0.5 for final clustering. (C) UMAP embedding of nuclei colored by cluster identity (0–11) shows clear separation of 12 transcriptionally distinct clusters, supporting the quality and robustness of the processed snRNA-seq dataset for downstream cellular characterization. Supplementary Material 7: Fig. 7. Dot plot showing expression of representative marker genes across final annotated cell populations. The dot plot illustrates state- and population-specific expression patterns for selected marker genes across the refined single-cell atlas, including astrocytes, oligodendrocytes, oligodendrocytes-like cells, neurons-like cells, microglia, NK cells, TAMs, and the three malignant glioma states (Glioma_MES, Glioma_PN, and Glioma_CL). Dot size represents the proportion of cells expressing each gene within a given population, while color intensity indicates the average expression level. Marker genes for the glioma transcriptional states (e.g., CD44, VIM, CHI3L1 for Glioma_MES; OLIG2, SOX2, NES for Glioma_PN; EGFR, PDGFRA, VEGFA for Glioma_CL) show clear, state-specific enrichment patterns, supporting the reliability of inferCNV-guided refinement and subsequent glioma sub-classification. Non-malignant populations exhibit minimal or absent expression of glioma state markers, confirming the separation of tumor and microenvironmental cells. Supplementary Material 7: Fig. 7. Dot plot showing expression of representative marker genes across final annotated cell populations. The dot plot illustrates state- and population-specific expression patterns for selected marker genes across the refined single-cell atlas, including astrocytes, oligodendrocytes, oligodendrocytes-like cells, neurons-like cells, microglia, NK cells, TAMs, and the three malignant glioma states (Glioma_MES, Glioma_PN, and Glioma_CL). Dot size represents the proportion of cells expressing each gene within a given population, while color intensity indicates the average expression level. Marker genes for the glioma transcriptional states (e.g., CD44, VIM, CHI3L1 for Glioma_MES; OLIG2, SOX2, NES for Glioma_PN; EGFR, PDGFRA, VEGFA for Glioma_CL) show clear, state-specific enrichment patterns, supporting the reliability of inferCNV-guided refinement and subsequent glioma sub-classification. Non-malignant populations exhibit minimal or absent expression of glioma state markers, confirming the separation of tumor and microenvironmental cells. Supplementary Material 8: Fig. 8. Sample-wise composition of annotated cell populations and malignant glioma states. Stacked bar plots depict the relative proportions of major non-malignant populations (astrocytes, oligodendrocytes, oligodendrocytes-like cells, neurons-like cells, microglia, NK cells, TAMs) and state-resolved malignant glioma cells (Glioma_MES, Glioma_PN, Glioma_CL) across individual snRNA-seq samples. Each bar represents a single sample, with colors corresponding to specific cell populations. Considerable inter-sample heterogeneity is observed, reflecting variability in the abundance of both tumor and microenvironmental populations across the dataset. These results highlight the compositional diversity of the single-cell atlas and underscore the necessity of sample-resolved analyses for downstream functional and marker-centered investigations. Supplementary Material 8: Fig. 8. Sample-wise composition of annotated cell populations and malignant glioma states. Stacked bar plots depict the relative proportions of major non-malignant populations (astrocytes, oligodendrocytes, oligodendrocytes-like cells, neurons-like cells, microglia, NK cells, TAMs) and state-resolved malignant glioma cells (Glioma_MES, Glioma_PN, Glioma_CL) across individual snRNA-seq samples. Each bar represents a single sample, with colors corresponding to specific cell populations. Considerable inter-sample heterogeneity is observed, reflecting variability in the abundance of both tumor and microenvironmental populations across the dataset. These results highlight the compositional diversity of the single-cell atlas and underscore the necessity of sample-resolved analyses for downstream functional and marker-centered investigations. Supplementary Material 9: Fig. 9. Violin plots showing the cell-type-specific expression distributions of the four prognostic markers in the refined snRNA-seq atlas. Violin plots illustrate the expression distributions of BCL2, CHEK2, CTNNB1, and NDRG1 across the final annotated cell populations and malignant glioma states, including astrocytes, Glioma_CL, Glioma_MES, Glioma_PN, microglia, neurons-like cells, NK cells, oligodendrocytes, oligodendrocytes-like cells, and tumor-associated macrophages (TAMs). Supplementary Material 9: Fig. 9. Violin plots showing the cell-type-specific expression distributions of the four prognostic markers in the refined snRNA-seq atlas. Violin plots illustrate the expression distributions of BCL2, CHEK2, CTNNB1, and NDRG1 across the final annotated cell populations and malignant glioma states, including astrocytes, Glioma_CL, Glioma_MES, Glioma_PN, microglia, neurons-like cells, NK cells, oligodendrocytes, oligodendrocytes-like cells, and tumor-associated macrophages (TAMs). Supplementary Material 10: Fig. 10. Marker-positive versus marker-negative differential expression analyses in the top three major expressing cell populations for each prognostic marker. (A–C) Differential expression analyses for BCL2-high versus BCL2-low cells in oligodendrocytes, oligodendrocytes-like cells, and TAMs, respectively. (D, E) Differential expression analyses for CHEK2-high versus CHEK2-low cells in microglia and TAMs, respectively. (F–H) Differential expression analyses for CTNNB1-high versus CTNNB1-low cells in glioma cells, oligodendrocytes-like cells, and TAMs, respectively. (I, J) Differential expression analyses for NDRG1-high versus NDRG1-low cells in oligodendrocytes and oligodendrocytes-like cells, respectively. Supplementary Material 10: Fig. 10. Marker-positive versus marker-negative differential expression analyses in the top three major expressing cell populations for each prognostic marker. (A–C) Differential expression analyses for BCL2-high versus BCL2-low cells in oligodendrocytes, oligodendrocytes-like cells, and TAMs, respectively. (D, E) Differential expression analyses for CHEK2-high versus CHEK2-low cells in microglia and TAMs, respectively. (F–H) Differential expression analyses for CTNNB1-high versus CTNNB1-low cells in glioma cells, oligodendrocytes-like cells, and TAMs, respectively. (I, J) Differential expression analyses for NDRG1-high versus NDRG1-low cells in oligodendrocytes and oligodendrocytes-like cells, respectively. Supplementary Material 11 Supplementary Material 11 Supplementary Material 12 Supplementary Material 12 Supplementary Material 13 Supplementary Material 13 Supplementary Material 14 Supplementary Material 14 Supplementary Material 15 Supplementary Material 15 Supplementary Material 16 Supplementary Material 16