Cellular subpopulation remodeling and biphasic molecular dynamics in glioma recurrence: a multi-omics analysis and diagnostic markers screening using liquid biopsy samples

Abstract

Objective

Glioblastoma (GBM) is a highly aggressive brain tumor with a median survival of about 15 months and frequent recurrence. Liquid biopsy of cerebrospinal fluid (CSF) and plasma enables minimally invasive detection of tumor-associated molecular signals and may improve early diagnosis of recurrence. This study aims to elucidate the distinct molecular landscapes of cerebrospinal fluid and plasma in glioblastoma, and to identify recurrence-associated biomarkers through an integrated multi-omics framework.

Methods

Using integrated bioinformatic frameworks, gene set enrichment analysis, and receiver operating characteristic curves, we systematically mapped differential gene expression patterns and their associated regulatory mechanisms.

Results

Cerebrospinal fluid yielded substantially more differentially expressed genes than plasma (3,362 vs. 1,329), with upregulated genes predominating in both compartments. We identified four genes (THNSL2, SYT5, RUVBL1, and GSKIP) that showed significant expression differences between primary and recurrent tumors, with RUVBL1 and GSKIP demonstrating the strongest functional enrichment signals. These genes were preferentially linked to glioblastoma cell subpopulations in both fluid types. The plasma-based multigene prognostic model achieved area under the curve values of 1.0 for 1-, 2-, and 3-year survival prediction in the training cohort, which requires further validation in independent cohorts.

Conclusion

Collectively, by using patients’ CSF and plasma samples, this study identifies candidate molecular markers and regulatory networks associated with primary and recurrent GBM through integrated analysis of CSF and plasma samples, and highlights complementary molecular signatures between the two fluid compartments. Our findings provide a preliminary molecular basis and candidate targets for future research on precision diagnosis, prognostic stratification, and targeted therapy of GBM.

Graphical abstract

Introduction

Glioblastoma (GBM), classified as a WHO grade 4 tumor, represents the most malignant form of astrocytic tumors. It is the most common malignant glioma in adults, accounting for 50% to 55% of neuroectodermal tumors. GBM predominantly affects males aged 45 to 65, with a male-to-female incidence ratio of approximately 3:2 [1]. Clinically, it is characterized by high invasiveness, rapid proliferation, and significant heterogeneity. Despite the current standard treatment regimen of maximal safe surgical resection combined with postoperative radiotherapy and chemotherapy, with some patients also receiving Tumor-Treating Fields therapy, the prognosis remains dismal. The median overall survival for newly diagnosed patients is only 15.6 months, and the recurrence rate is extremely high, with a median survival of 6 to 9 months after recurrence [2,3].

The primary challenge in managing recurrent GBM lies in the lack of precise early diagnostic tools and effective targeted therapies. Conventional anatomical imaging often fails to distinguish tumor recurrence from radiation-induced necrosis, a challenge compounded by the inability of standard imaging to capture molecular evolution [4]. In recent years, molecular pathological markers have played an increasingly important role in the diagnosis, classification, and prognostic evaluation of GBM. The 2021 WHO guidelines have integrated molecular markers, such as IDH and TERT promoter mutations, into the diagnostic system [5]. Liquid biopsy samples, including cerebrospinal fluid (CSF) and plasma, serve as accessible windows into tumor molecular dynamics. CSF directly contacts central nervous system tissue and captures local tumor molecular alterations with high fidelity, reflecting the intracranial pathological state; plasma, as a systemic circulating fluid, offers the advantages of non-invasive collection and serial monitoring, reflecting the systemic impact of tumor progression [[6], [7], [8], [9]]. These two fluid compartments exhibit distinct but complementary molecular dynamics during GBM recurrence, forming a "biphasic" molecular signature system. However, current studies mostly focus on a single fluid type, and a systematic comparison of the transcriptomic landscapes of CSF and plasma in recurrent GBM is lacking. Current studies have identified individual recurrence-associated molecules in GBM, but several critical knowledge gaps remain. First, the molecular landscapes of CSF and plasma in recurrent GBM have not been systematically compared, and the complementary value of dual-compartment liquid biopsy remains unclear. Second, the cellular origin of recurrence-associated signals detected in body fluids is poorly defined, limiting mechanistic interpretation. Third, there is a lack of prognostic models built on liquid biopsy transcriptomic data for recurrent GBM. To address these gaps, this study applies an integrated multi-omics framework to systematically investigate the expression patterns, functional associations, and clinical relevance of recurrence-associated molecules in both CSF and plasma, with the aim of providing candidate biomarkers and new insights for precision diagnosis of GBM recurrence.

Materials and methods

Data acquisition and preprocessing

Single-cell RNA sequencing dataset

Sample and cell counts were evaluated for the datasets listed in Table S1. SCDS0000041, which contained the largest number of cells, was selected for single-cell analysis. GBM single-cell RNA sequencing data (SCDS0000041) were obtained from https://db.cngb.org/cdcp/dataset/SCDS0000041/, comprising 36 primary and 36 recurrent samples.

Raw data quality control was performed using the Seurat package (v5.3.1; https://github.com/satijalab/seurat). Low-quality cells were removed based on the following criteria: mitochondrial gene percentage < 10% and 400–7,500 detected genes per cell. Data normalization was performed using the NormalizeData function, followed by identification of highly variable genes using FindVariableFeatures and scaling with ScaleData. The top 3,000 highly variable genes were identified, and the first 30 principal components were calculated using principal component analysis for downstream dimensional reduction and clustering. Clustering results were visualized using the DimPlot function. Stacked bar plots of cell-type proportions were generated to illustrate compositional differences across treatment groups.

Body fluid transcriptome datasets

Body fluid transcriptome datasets derived from CSF and plasma samples were obtained from public repositories (Table S2).

Single-cell transcriptomic analysis

Cell type annotations were applied to clustered subpopulations based on canonical marker genes reported in the original publication [10]. Differentially expressed genes (DEGs) in recurrent-group subpopulations were identified using the FindAllMarkers function (Wilcoxon rank-sum test; p < 0.05). Functional enrichment analysis of DEGs was conducted using GO and KEGG analyses implemented in the clusterProfiler package. Gene Set Enrichment Analysis (GSEA) was performed using the GO-related gene set (c5.go.v7.5.1.symbols.gmt) downloaded from the MSigDB database (https://www.gsea-msigdb.org/gsea/msigdb) to identify core functional pathways within each subpopulation.

Pseudotime analysis was conducted on key cell subpopulations using the Monocle package (v2.34.0; https://github.com/cole-trapnell-lab/monocle-release). Dimensionality reduction was applied to the single-cell expression matrix to reconstruct differentiation trajectories, and distinct cell fate branches were identified through branch-point detection. The distribution of each subpopulation along the pseudotime trajectory and across cell fate branches was quantified to assess differentiation dynamics during disease progression. Subpopulations enriched within specific branches, as well as putative “initiating” subpopulations located at branch points, were identified. Branch-associated DEGs were further classified into gene clusters by hierarchical clustering. GO enrichment analysis was performed for each gene cluster to determine activated pathways, and the top five genes from each cluster were selected to illustrate their expression patterns across cell subpopulations.

Comparative analysis of CSF and plasma DEGs

Differential expression analysis between the glioma (GLI) and normal control (NOR) groups was performed independently for CSF and plasma using the limma package, with DEGs identified at a threshold of adjusted p < 0.05. DEG expression patterns were visualized by heatmaps and volcano plots, and significantly up- and downregulated genes were identified in CSF and plasma samples from the recurrent group.

Venn diagram analysis was applied to these identified recurrence-associated DEGs to classify genes into three categories: genes shared between both fluids (overlapping genes), CSF-specific genes, and plasma-specific genes. GO and KEGG enrichment analyses were conducted for each gene set using the clusterProfiler package. The expression patterns of shared and fluid-specific genes were visualized using heatmaps.

Construction and evaluation of recurrence status prediction models

The prognostic model was constructed through a two-step procedure:

Univariate Cox screening: Univariate Cox proportional hazards regression was performed on all DEGs from each fluid type independently using the survival R package (v3.7-0). Genes with a p-value < 0.01 were retained as candidate prognostic genes.

LASSO-penalized Cox regression: The candidate genes were subjected to LASSO-penalized Cox regression using the glmnet R package (v4.1-8), with 10-fold cross-validation to select the optimal lambda value (lambda.min). The risk score for each sample was calculated using the following formula:

$$RiskScore=\sum _{i=1}^n(Coe{f}_i\times Expressio{n}_i)$$

where $Coe{f}_i$ is the LASSO regression coefficient of gene i, and \(Expression_i\) is the normalized expression value of gene i.

The optimal cutoff value for the risk score was determined using the surv_cutpoint function in the survminer R package as previously described [11], and samples were stratified into high-risk (risk score ≥ cutoff) and low-risk (risk score < cutoff) groups accordingly. Kaplan–Meier survival curves were generated using the survival package, and survival differences between groups were evaluated by the log-rank test. The prognostic performance of the risk score was evaluated by calculating time-dependent receiver operating characteristic (ROC) curves and the corresponding area under the curve (AUC) values at 1, 2, and 3 years using the timeROC R package (v0.4). It should be noted that all performance evaluations were conducted in the training cohort, and independent validation cohorts are required to confirm the actual predictive power.

Integration of single-cell and body fluid data

To link cell subpopulation characteristics to body fluid molecular signatures, Spearman correlation analysis was performed between the marker genes of each annotated subpopulation and the overlapping DEGs from CSF and plasma, using the rcorr function in the Hmisc package. A GMT file was constructed from subpopulation marker genes, and single-sample GSEA (ssGSEA) was applied via the GSVA package to estimate the relative abundance of each cell subpopulation in body fluid samples. Correlation analysis was subsequently conducted between the expression levels of overlapping DEGs and subpopulation abundance scores to identify cell subpopulations significantly associated with fluid-derived molecular signals and to determine mRNAs significantly correlated with specific subpopulations (p < 0.05).

Construction of a cell–molecule regulatory network

To elucidate the cellular origins and regulatory relationships of recurrence-associated molecules detected in body fluids, mRNAs significantly correlated with specific cell subpopulations (p < 0.05) were subjected to upstream regulator prediction. Transcription factors (TFs) were predicted using the ChEA3 database (https://maayanlab.cloud/chea3/), and regulatory miRNAs were predicted using the mirDIP database (https://ophid.utoronto.ca/mirDIP/index.jsp; prediction score > 0.5). The predicted regulatory interactions were integrated and imported into Cytoscape to construct and visualize the cell–molecule regulatory network, illustrating the relationships between mRNAs and their upstream regulatory factors.

Functional characterization of candidate molecules

The expression of candidate body fluid mRNAs in key cell subpopulations was visualized using violin plots generated from the single-cell dataset. To investigate their biological functions, recurrent-group samples in the body fluid dataset were stratified into high and low expression groups based on the median expression value of each mRNA. Single-gene GSEA was performed between the two groups using the GO gene set (c5.go.v7.5.1.symbols.gmt) from MSigDB and the clusterProfiler package (significance threshold: p < 0.05). The top 50 significantly enriched pathways were visualized as bubble plots using ggplot2.

Preliminary bioinformatic validation of candidate biomarkers in the TCGA-GBM cohort

Normalized transcriptome data from the TCGA-GBM cohort were downloaded from the GDC database (https://portal.gdc.cancer.gov/), including 386 glioma samples (372 primary tumors and 14 recurrent tumors). Expression levels of the previously identified key fluid-derived molecular candidates were analyzed and compared between primary and recurrent tumors using the Wilcoxon rank-sum test. Results were visualized as box plots overlaid with individual data points. ROC curve analysis was performed to assess the ability of each gene and their combination to discriminate between primary and recurrent tumor status, with AUC values calculated to quantify discriminative performance. These results provide preliminary supporting evidence for the recurrence association of the candidate genes, but further validation in independent liquid biopsy cohorts is required.

Results

Characterization of key pathogenic cell subpopulations associated with GBM recurrence

To characterize recurrence-associated cellular heterogeneity and identify key pathogenic subpopulations, we analyzed single-cell RNA sequencing data from SCDS0000041, comprising 36 primary and 36 recurrent samples. Following stringent quality control and unsupervised clustering analysis, a high-quality single-cell transcriptomic matrix comprising 110,932 cells and 27,925 genes was generated (Supplementary Fig. S1). Based on the annotation scheme described in the corresponding original publication [10], six cell types were identified: GBM cells, neurons, oligodendrocytes, myeloid cells, T cells, and astrocytes (Fig. 1A). Among these, GBM cells exhibited the largest proportional decrease in the recurrent group (−23.1%), followed by oligodendrocytes, which increased by 10% (Fig. 1B-C). Given this substantial shift, GBM cells were selected as the primary cell population for subsequent analysis.

Fig. 1.

Key pathogenic cell subpopulations associated with glioma recurrence. (A, B) Cell type annotation analysis: (A) UMAP cell annotation plot; (B, C) stacked bar chart showing the proportional distribution of cell types across the two groups. (D–H) Functional enrichment of cell-type-specific differentially expressed genes: (D–F) GO enrichment analysis of DEGs; (G) KEGG enrichment analysis of DEGs; (H) GSEA enrichment analysis.

DEGs across the six cell subpopulations in the recurrent group were identified using the FindAllMarkers function. A total of 1,721 DEGs were detected in GBM cells (p < 0.05), which were significantly enriched across 1,803 GO terms, including 1,360 Biological Process (BP), 222 Cellular Component (CC), and 221 Molecular Function (MF) terms, and 86 KEGG pathways (p < 0.05; Fig. 1D–G). For each enrichment category, the top 20 pathways by number of enriched DEGs are shown. Representative enriched Biological Process (BP) terms included small GTPase-mediated signal transduction, regulation of neuron projection development, and axonogenesis; Cellular Component (CC) terms included neuron-to-neuron synapse, postsynaptic specialization, and cell leading edge; Molecular Function (MF) terms included GTPase regulator activity, nucleoside-triphosphatase regulator activity, and actin binding; KEGG pathways included IgSF CAM signaling, Endocytosis, and the PI3K–Akt signaling pathway. GSEA performed on the 1,721 DEGs using the c5.go.v7.5.1.symbols.gmt gene set identified four significantly enriched GO pathways (Fig. 1H).

Pseudotime analysis of key cell differentiation trajectories during disease progression

Pseudotime analysis revealed bifurcation into two distinct cell fate trajectories (cell fate 1 and cell fate 2; Fig. 2A). The distribution of cell types along the pseudotime axis was markedly heterogeneous. Neurons exhibited peak density within the intermediate pseudotime window (10–20), whereas GBM cells and T cells were distributed across both early (0–10) and late (20–30) stages (Fig. 2B), suggesting pronounced heterogeneity in differentiation timing and state transitions across cell types. BEAM heatmap analysis (Fig. 2C) resolved four core gene clusters (Clusters 1–4). GO BP enrichment analysis identified significantly enriched pathways (p < 0.05); the top 4–5 pathways by gene count are shown for each cluster. For instance, Cluster 1 was enriched in neuronal action potential propagation and chromatin remodeling at centromere. Notably, Cluster 2 genes, represented by ACSL1 and AGAP1, displayed significantly higher expression levels in GBM cells relative to other cell types. Their sustained medium-to-high expression patterns across pseudotime progression suggest a stable functional role in tumor cell regulation (Fig. 2D).

Fig. 2.

Pseudotime differentiation trajectories and molecular features of key cells. (A) Pseudotime trajectory plot, in which color depth represents pseudotime value (darker colors indicate higher pseudotime values) and cells are distributed along trajectories leading to distinct fate branches. (B) Cell density distribution along pseudotime trajectories: upper panel shows the branching structure; lower panel shows cell density curves per branch. (C) BEAM heatmap and pathway enrichment analysis: left panel shows enriched pathways for differentiation-associated genes; right panel shows the corresponding BEAM heatmap of gene expression patterns. (D) Dynamic expression of the top 5 characteristic genes per cluster across cell types along the scaled pseudotime axis (0–100).

Specific effects of recurrent GBM on the CSF and plasma transcriptomes

Comparative analysis revealed that glioma-associated transcriptional alterations were more pronounced in CSF than in plasma. Relative to the NOR group, the GLI group exhibited substantially more DEGs in CSF (3,362 genes) than in plasma (1,329 genes) (Fig. 3A–B). Both fluid types were dominated by upregulated genes: upregulated DEGs accounted for 83.3% of CSF DEGs (2,801/3,362) and 69.9% of plasma DEGs (929/1,329), indicating that GLI-associated transcriptional alterations are more pronounced in CSF and are predominantly driven by gene activation (Fig. 3C–E). In plasma, DEG log fold-change (logFC) values were concentrated between −2 and 2, with no extreme outliers. In contrast, a subset of CSF DEGs exceeded logFC ±3, with the most extreme changes predominantly among upregulated genes (shown in red). These findings indicate that GLI-related transcriptional dysregulation is more extensive in CSF, consistent with the direct anatomical proximity of this compartment to central nervous system pathology.

Fig. 3.

Differential expression analysis. (A) Volcano plot and heatmap of DEGs in plasma samples. The x-axis represents log fold-change (logFC) and the y-axis represents statistical significance (−log10(p-value)); red dots indicate upregulated genes, blue dots indicate downregulated genes, and black dots indicate non-significant genes; comparisons are between the GLI group (target) and the NOR group (control). (B) Volcano plot and heatmap of DEGs in CSF samples. (C) Principal component analysis of CSF samples. (D) Principal component analysis of plasma samples. (E) Bar chart showing the number of differentially expressed genes.

Validation of body fluid molecular signatures as indicators of GBM recurrence status

For the CSF dataset, univariate Cox analysis of the 3,362 DEGs identified two genes with p < 0.01; LASSO Cox regression then refined this to 2 characteristic genes (Fig. 4A). Survival analysis supported their prognostic relevance: patients in the high-expression group showed significantly worse survival outcomes compared with the low-expression group (p = 0.00041; Fig. 4B). ROC curve analysis in CSF samples demonstrated a favorable sensitivity–specificity balance (maximum AUC > 0.9), supporting the prognostic utility of the CSF-based model (Fig. 4C–D).

Fig. 4.

Prognostic model analysis. CSF-related panels (A–D): (A) Bar plot of prognostic gene coefficients. (B) Kaplan–Meier survival curves for high- and low-expression groups. (C) Receiver operating characteristic (ROC) curves illustrating prognostic performance (x-axis: specificity; y-axis: sensitivity). (D) Risk factor distribution plot. Plasma-related panels (E–H): (E) Bar plot of prognostic gene coefficients. (F) Kaplan–Meier survival curves for high- and low-risk groups. (G) ROC curves illustrating prognostic performance. (H) Risk factor distribution plot.

For the plasma dataset, univariate Cox analysis of 1,329 DEGs identified 29 genes with p < 0.01; LASSO Cox regression further refined this to 10 characteristic genes (Fig. 4E). Risk-stratified Kaplan–Meier curves in plasma samples (Fig. 4F) likewise revealed highly significant survival differences between high- and low-risk groups (p < 0.0001), demonstrating the clinical potential of the plasma-based risk stratification model. ROC analysis showed that the multi-molecule prognostic model achieved AUC values of 1.0 for 1-, 2-, and 3-year survival prediction in the current training cohort (Fig. 4G–H). It should be noted that this result is based on a limited sample size and may be affected by overfitting; the actual prognostic performance needs to be verified in larger independent cohorts.

Comparative analysis of CSF and plasma transcriptomes to identify recurrence-associated circulating molecules

Venn analysis identified three DEG subsets (Fig. 5A): 3,304 CSF-specific genes, 1,271 plasma-specific genes, and 58 overlapping genes. Functional enrichment analysis revealed distinct biological signatures across subsets. For the 58 overlapping genes, enrichment identified 168 GO terms (113 BP, 23 CC, 32 MF) and 2 KEGG pathways (p < 0.05; Fig. 5B). Overlapping genes were primarily associated with Wnt signaling regulation and chromatin remodeling. Key enriched terms included: BP (positive regulation of canonical Wnt signaling pathway, positive regulation of Wnt signaling pathway, and regulation of canonical Wnt signaling pathway); CC (RPAP3/R2TP/prefoldin-like complex, Swr1 complex, and Ino80 complex); MF (phosphatidylinositol-4,5-bisphosphate binding, phosphatidylinositol bisphosphate binding, and phosphatidylinositol phosphate binding); KEGG (Mitophagy [animal] and ATP-dependent chromatin remodeling)

Fig. 5.

Functional enrichment analysis of characteristic genes. (A) Venn diagram of overlapping and compartment-specific genes between CSF and plasma samples. GO and KEGG enrichment analyses for overlapping genes (B), CSF-specific genes (C), and plasma-specific genes (D). (E) Heatmaps showing the expression of overlapping genes, CSF-specific genes, and lasma-specific genes in CSF and plasma samples. Colors represent relative expression levels (red denotes high expression; blue denotes low expression). Right-side group labels indicate GLI (glioma) and NDR (normal control).

For the 3,304 CSF-specific genes, enrichment revealed 886 GO terms (601 BP, 116 CC, 169 MF) and 58 KEGG pathways (p < 0.05; Fig. 5C). Notably enriched terms included: BP (regulation of membrane potential, axonogenesis, and cell–cell adhesion via plasma-membrane adhesion molecules), CC (synaptic membrane, postsynaptic specialization, and neuron-to-neuron synapse), MF (monoatomic ion channel activity, actin binding, and GTPase regulator activity), KEGG (IgSF CAM signaling, Neuroactive ligand–receptor interaction, and Cytoskeleton in muscle cells.

For the 1,271 plasma-specific genes, enrichment yielded 178 GO terms (134 BP, 17 CC, 27 MF) and 7 KEGG pathways (p < 0.05; Fig. 5D). Prominent enriched terms included: BP (response to virus, amide metabolic process, and defense response to virus); CC (ubiquitin ligase complex, ribosomal subunit, and cullin–RING ubiquitin ligase complex); MF (ubiquitin protein ligase binding, ubiquitin-like protein ligase binding, and ubiquitin-like protein binding); KEGG (NOD-like receptor signaling pathway, Toll-like receptor signaling pathway, and Relaxin signaling pathway).

Heatmap visualization of overlapping, CSF-specific, and plasma-specific gene expression in CSF and plasma samples (Fig. 5E) showed that CSF harbored substantially more genes with aberrant expression (58 overlapping + 3,304 CSF-specific) relative to plasma (58 overlapping + 1,271 plasma-specific). Moreover, high-expression regions (red) in the CSF heatmap were more densely concentrated, confirming that GBM-associated transcriptional dysregulation is more widespread and more pronounced in CSF.

Association between single-cell features and body fluid molecular signatures

In CSF, GBM cell markers EGFR and PTPRZ1 showed strong positive correlations with multiple body fluid molecules (Spearman's r = 0.3–0.6; −log10(p) = 3–4); core associated molecules included tRF-30-QKF1R3WE8RO8 and hsa-miR-9-3p/5p. T cell markers CD247 and CD74 exhibited moderate positive correlations with tRF-18-8S68BFD2 and piR-hsa-10008 (Spearman's r = 0.2–0.3; −log10(p) = 2–3). Markers from remaining cell types showed weaker and less significant associations. ssGSEA-derived GBM cell abundance scores were strongly positively correlated with hsa-miR-9-3p/5p (Spearman's r = 0.4–0.6; −log10(p) = 3–4), while T cell abundance showed moderate positive correlations with tRF-18-8RSI66D2 and RNU1-149P (Spearman's r = 0.2–0.4; −log10(p) = 2–3). Abundance scores of the remaining cell types showed no significant associations with any body fluid molecules (Fig. 6A and 6C).

Fig. 6.

Correlation analysis. (A, B) Spearman correlations between cell-type marker genes and the 58 overlapping genes in plasma (A) and CSF (B) samples. (C, D) Spearman correlations between ssGSEA-derived cell subpopulation abundance scores and the 58 overlapping genes in plasma (C) and CSF (D) samples (*p < 0.05; **p < 0.01; ***p < 0.001; ****p < 0.0001).

In plasma, GBM cell markers EGFR and CTNNA3 were strongly positively correlated with tRF-31-QJ3KYU4RRWRBD and piR-hsa-27897 (Spearman's r = 0.3–0.6; −log10(p) = 4–5). Notably, the associated molecules differed from those in CSF-CSF associations were dominated by miRNAs, whereas plasma associations were driven by tRFs and piRNAs, reflecting compartment-specific molecular signatures. T cell markers CD247 and CD96 showed moderate positive correlations with tRF-18-8S68BFD2 and ENSG00000283828 (Spearman's r = 0.2–0.3; −log10(p) = 2–3), consistent with T cell associations observed in CSF, though the specific molecules differed. GBM cell ssGSEA abundance scores were strongly positively correlated with tRF-30-QKF1R3WE8RO8 and piR-hsa-10008 (Spearman's r = 0.3–0.5; −log10(p) = 3–4), corroborating the direct cell marker correlation findings and confirming the molecular reflection characteristics of GBM cells in plasma. Astrocyte abundance scores showed moderate positive correlations with RNU2-36P and SNORD42A (Spearman's r = 0.2–0.25; −log10(p) = 1.5–2), while T cell abundance scores showed a weak positive correlation with tRF-18-8R6Q46D2 (Spearman's r ≈ 0.15; −log10(p) ≈ 1.5). Abundance scores of the remaining cell types showed no statistically significant correlations with any plasma molecules (Fig. 6B and 6D).

Taken together, these analyses demonstrate that the characteristic genes of GBM cells and T cells are significantly correlated with body fluid molecular expression. The abundance of these two subpopulations was significantly correlated (p < 0.05) with multiple body fluid molecules, with 22 in CSF and 11 in plasma. After merging and deduplication, 27 molecules were identified: RNU2-61P, RNU2-37P, RNU2-7P, RNU2-33P, RNU2-48P, RNU2-36P, RNU2-6P, tRF-16-K6S185D, RUVBL1, tRF-20-9MV47P59, GSKIP, RNU2-27P, SYT5, THNSL2, tRF-50-QNR8VP94FQFY392J1Y0S, tRF-29-5DMKYUYRLHIX, tRF-18-8S68BFD2, tRF-18-8RSI66D2, tRF-18-8R6Q46D2, ENSG00000236341, SNORD42A, ENSG00000273885, piR-hsa-10008, ENSG00000283828, MEG8, tRF-18-NSRJ7K05, and RNU1-21P, including 4 protein-coding mRNAs: GSKIP, RUVBL1, SYT5, and THNSL2.

Construction of a "cell–molecule" regulatory network to reveal the origin and regulatory relationships of recurrence-associated body fluid molecules

Database-based prediction for the four candidate mRNAs (GSKIP, RUVBL1, SYT5, and THNSL2) identified 257 TFs and 47 miRNAs with prediction scores > 0.5. As shown in Supplementary Fig. S2, each of the six annotated cell types is connected to each molecule by two edges representing CSF and plasma associations, respectively. GBM cells demonstrated strong and significant correlations with all four recurrence-associated molecules, whereas T cells and Neurons showed significant correlations only with THNSL2. These findings suggest that GBM cells are likely the primary cellular source of the four recurrence-associated body fluid molecules, with secondary contributions from T cells and Neurons.

Direct associations and regulatory mechanisms between recurrence-associated body fluid molecules and key cell subpopulations

Using violin plots combined with box plots, the expression distribution of the four target genes (GSKIP, RUVBL1, SYT5, and THNSL2) was analyzed across six cell types in Primary and Recurrent subgroups. Cell type–specific expression preferences were evident: THNSL2 was expressed at relatively high levels in T cells, Neurons, Myeloid cells, and GBM cells, whereas SYT5 showed weak signals in most cell types. Expression differences between Primary and Recurrent groups were observed in specific cell types: in GBM cells, THNSL2 expression was generally higher in the Primary group; conversely, in T cells, RUVBL1 was slightly elevated in the Recurrent group. Notably, all four genes showed increased expression specifically in the Recurrent subgroup of GBM cells (Fig. 7A), suggesting selective enrichment in post-recurrence GBM cells and further supporting their role as cell-autonomous molecular drivers of GBM recurrence.

Fig. 7.

GSEA enrichment analysis of target genes. (A) Expression distribution of GSKIP, RUVBL1, SYT5, and THNSL2 across different cell types. (B) GO-GSEA enrichment results for the four genes. The top 50 significantly enriched GO pathways are shown.

To explore the biological functions of these genes, GSEA was performed using the c5.go.v7.5.1.symbols.gmt gene set from MSigDB (https://www.gsea-msigdb.org/gsea/msigdb) with the clusterProfiler package. Pathways with p < 0.05 were considered significant. For example, stratifying samples by GSKIP expression revealed significant enrichment in GO pathways including Protein_kinase_A_regulatory_subunit_binding, Mitochondrial_calcium_ion_homeostasis, and Luteinizing_hormone_secretion (Fig. 7B).

Preliminary validation of key body fluid molecules as liquid biopsy biomarkers

In the TCGA-GBM cohort, SYT5 and GSKIP expression levels were significantly elevated in recurrent glioma samples relative to primary tumors (p < 0.01, Fig. 8A), whereas THNSL2 and RUVBL1 showed no statistically significant intergroup differences (marked "ns"). ROC curve analysis revealed variable discriminative performance across the four genes: SYT5 (AUC = 0.55) and GSKIP (AUC = 0.63) demonstrated modest discriminative capability, while THNSL2 (AUC = 0.54) and RUVBL1 (AUC = 0.52) performed near chance level, suggesting that SYT5 and GSKIP may serve as candidate biomarkers for distinguishing primary from recurrent glioma. A combined four-gene model (allgene) achieved an AUC of 0.623 (Fig. 8B), outperforming single gene and indicating that a multi-gene approach improves identification of glioma recurrence status. Moreover, Kaplan-Meier survival analysis revealed a statistically significant difference in overall survival between patient subgroups stratified by high versus low expression levels of SYT5 and THNSL2 (Fig. 8C). A volcano plot further illustrates the differential expression of SYT5 and GSKIP, highlighting their fold-changes and statistical significance (Fig. 8D). Univariate Cox proportional hazards regression confirmed THNSL2 as a robust independent prognostic predictor (hazard ratio not specified; p = 0.004; Fig. 8E).

Fig. 8.

Validation of recurrence-associated mRNAs in the TCGA-GBM cohort. (A) Box plots showing the expression levels of GSKIP, RUVBL1, SYT5, and THNSL2. (B) ROC curves showing discriminative performance. (C) Survival curves of these 4 mRNAs. (D) Volcano plots showing the distribution of mRNA variation multiples. (E) Forest plot of univariate cox analysis for these 4 mRNAs.

Discussion

Gliomas are among the most aggressive tumors, with the lowest overall survival of all cancers [12]. While immune and tumortargeted therapies have improved outcomes in extracranial cancers, similar success in gliomas remains limited, reflecting unique mechanisms of persistence within the brain [13]. Recent studies have identified various cellular and molecular alterations at GBM recurrence, including mesenchymal transition, a reduction in the proneuronal state, a higher proportion of oligodendrocytes, and enriched ECM signatures, indicating a systematic context shift in the tumor and its surrounding environment [[14], [15], [16], [17], [18], [19]]. Therefore, thoroughly characterizing rGBM tissues relative to pGBM is crucial for comprehending relapse mechanisms and identifying targeted therapies [20].

The recurrence of GBM is typically insidious, with nonspecific or subtle early clinical manifestations—leading to frequent delays in diagnosis and suboptimal management of patients with recurrent disease. In current clinical practice, molecular and immunohistochemical markers—including isocitrate dehydrogenase 1 (IDH1), epidermal growth factor receptor (EGFR), tumor protein p53 (TP53), and the proliferation marker Ki-67—are routinely employed to support the initial diagnosis of GBM [21,22]. However, these biomarkers exhibit limited sensitivity and specificity in detecting rGBM, particularly during its early stages. Consequently, identifying novel, robust, and clinically actionable biomarkers specific to rGBM remains a critical unmet need to enable timely intervention and improve therapeutic outcomes.

To date, tissue sample acquisition through biopsies has been employed for histological and molecular tumor analysis to confirm diagnoses, classify tumor types, identify tumor-specific mutations, and inform therapeutic protocols [23]. Despite their critical role, tissue biopsies are inherently invasive and often present significant challenges. In particular, surgical brain tissue biopsies carry an increased risk of hemorrhage and potential damage to critical brain regions, which can result in neurological deficits, especially in the developing pediatric brain [24]. Moreover, tissue biopsies may fall short in detecting early-stage tumors or residual lesions and are not suitable for screening purposes [25]. Additionally, the temporal and spatial heterogeneity of tumors can limit the representativeness of tissue biopsies, reducing their effectiveness for monitoring tumor progression and assessing therapeutic response [23]. In multifocal cases, multiple biopsies may be required to obtain an accurate representation of the tumor, as invasive tumors continuously evolve both spatially and temporally, particularly in response to treatment [26]. To this end, liquid biopsy offers a promising solution for tumor diagnosis and assessing risk, prognosis, and recurrence. However, its use in regular medical practice is limited due to a lack of standardized methods and genetic markers [27]. Future clinical trials should focus on biofluid collection, biomarker selection, and detection strategies to integrate liquid biopsy into clinical settings.

Although elevated expression of ECM-related genes is associated with a mesenchymal-like cell state during recurrence [28], it remains uncertain if this is characteristic of all tumor cells or specific subsets. This is feasible due to extensive studies highlighting GBM's heterogeneity, including the identification of subpopulations based on transcriptomic similarities and states defined by major expression modules. While research on rGBM heterogeneity frequently references gene modules derived from pGBM [17,29], a comprehensive understanding of the molecular heterogeneity in rGBM, particularly in relation to pGBM and various anatomical regions, remains lacking. In this study, we established an integrative analytical framework combining single-cell transcriptomics with CSF and plasma transcriptomics to systematically characterize the molecular alterations associated with GBM recurrence. We revealed distinct and complementary molecular signatures between CSF and plasma, and traced the cellular origin of key recurrence-associated candidate molecules to GBM tumor cells. Compared with previous studies focusing on a single fluid type or single molecular layer, our work provides a more comprehensive landscape of liquid biopsy signals in recurrent GBM and offers candidate targets for subsequent mechanism research and clinical verification. Considering the ECM's involvement in promoting tumorigenesis and relapse [30], its component genes have emerged as potential targets for cancer therapy [[31], [32], [33]]. Identifying key regulators of abnormal ECM production and characterizing the molecular traits of crucial cell subpopulations involved in ECM gene expression is vital. In current research, we identified THNSL2, SYT5, RUVBL1, and GSKIP as potential regulators involved in the regulation of key signature genes for rGBM. Our comprehensive single-cell and spatial transcriptomics analyses of longitudinal glioblastoma specimens enhance the mechanistic understanding of recurrent glioblastoma state transitions associated with elevated extracellular matrix gene expression, and identify a crucial subpopulation of recurrent glioblastoma cells along with key molecular regulators. As studies on rGBM progression expand, more molecular candidates will be identified and validated for future targeted therapy research.

Among the four identified candidate genes, RUVBL1 and GSKIP showed the strongest functional enrichment signals, suggesting their core roles in GBM recurrence. RUVBL1 is a member of the AAA+ ATPase family and a key component of multiple chromatin remodeling complexes. Accumulating evidence indicates that RUVBL1 participates in the regulation of oncogenic signaling pathways including c-Myc and β-catenin, and is essential for the maintenance of cancer stem cell stemness and DNA damage repair [34]. In glioma, RUVBL1 upregulation has been associated with poor prognosis and temozolomide resistance [35], which may explain its enrichment in recurrent GBM cells after chemoradiotherapy. GSKIP (GSK3β interacting protein) acts as a negative regulator of GSK3β and is involved in cytoskeletal reorganization and cell motility. Previous studies have shown that GSKIP, a GSK3β-interacting protein, modulates β-catenin signaling and is up-regulated in multiple cancer types, which is closely related to tumor cell invasion, proliferation, and therapeutic resistance [36]. Its upregulation in recurrent GBM may promote tumor invasiveness and mesenchymal transition. As for the other two candidates, SYT5 is a synaptotagmin family member involved in calcium-dependent vesicle trafficking and secretion, which may affect intercellular communication and tumor microenvironment remodeling; THNSL2 is a threonine synthase-like protein associated with nucleotide metabolism and cell cycle regulation. These functional implications are consistent with our enrichment results showing chromatin remodeling, Wnt signaling, and cell adhesion alterations in recurrent GBM, supporting the potential mechanistic roles of these candidate genes in recurrence.

From a translational perspective, the candidate recurrence-associated genes identified in this study can be further explored through structural biology approaches guided by artificial intelligence. In recent years, the AlphaFold AI model has revolutionized structural biology by accurately predicting the three-dimensional structures of almost all known proteins [37], markedly expanding the structural coverage of the human proteome and providing unprecedented opportunities for drug target discovery and virtual screening [38]. In GBM research, AlphaFold has been used to predict the structures of key therapeutic targets such as CD44 and DNMT1, and has successfully guided the virtual screening and design of targeted peptides [39]. In future research, the proteins encoded by RUVBL1, GSKIP and other candidate genes identified in this study can be used as inputs for AlphaFold structure prediction to resolve their three-dimensional conformations and potential interaction interfaces with upstream and downstream regulatory proteins. On this basis, structure-based virtual screening and molecular optimization can be carried out to develop small molecule inhibitors or targeted peptides against these recurrence-driving molecules, thereby translating liquid biopsy-derived biomarkers into precision intervention targets for recurrent GBM. This AI-driven structural biology approach will greatly accelerate the transformation from biomarker discovery to targeted therapy development.

The study encountered multiple limitations. A key limitation is that the single-cell analysis was based on a single public dataset, which may introduce biases in cell classification, pathway analysis, and cell abundance estimation due to batch effects and sample heterogeneity. The cell annotation was performed according to the marker genes reported in the original publication, and may not fully capture the full heterogeneity of recurrent GBM cells. Since the primary purpose of the single-cell analysis was to infer the cellular origin of liquid biopsy signals rather than to comprehensively characterize GBM cellular heterogeneity, the current analysis is sufficient to support our core conclusions. The scarcity of appropriate CSF and plasma samples from recurrent glioma patients means that no independent liquid biopsy cohort was available for experimental validation of our findings. All validation analyses in this study were performed using public tissue transcriptome data, and the candidate biomarkers identified herein remain at the hypothesis-generating stage. Prospective clinical cohorts and functional experiments are needed to verify their diagnostic performance and biological roles in GBM recurrence. Our goal is to enhance research by broadening sample collection and creating a strong clinical cohort to gain deeper insights into our study area. As a retrospective study, our analyses remain correlative. The use of TMAs, though pathologist-guided to capture tumor- and immune-rich regions, limited tissue coverage per patient and may have missed spatially distinct areas. More broadly, spatial profiling inherently samples small tumor regions, raising uncertainty about how representative these sections are of the whole lesion, especially given GBM variability. Besides, the dataset also lacks pediatric brain controls from distinct regions, limiting region-specific comparisons. Finally, logistical and privacy constraints across sites prevented complete clinical histories for all patients, highlighting opportunities for future datasets with broader coverage and annotation.

In summary, promising genes for the detection of tumor recurrence in patients with GBM could be identified. Further experimental research is required to independently validate the proposed biomarkers in plasma or CSF as a robust alternative to frequent MRI scan. Our biomarker for tumor recurrence showed a significant inter-individual variance among GBM patients, thus supporting a personalized medical approach for optimized diagnosis findings.

Conclusion

In summary, this study demonstrates that differentially expressed molecules in CSF and plasma robustly reflect the pathological features of gliomas. Notably, molecular alterations in CSF exhibit greater magnitude and biological relevance compared to those in plasma; conversely, a multi-molecular predictive model derived from plasma demonstrates high prognostic accuracy. Four core candidate genes—THNSL2, SYT5, RUVBL1, and GSKIP—show stage-specific expression trends, suggesting their potential value in discriminating between primary and recurrent glioma, which requires further validation in independent clinical cohorts. Collectively, these findings identify promising candidate biomarkers for precise diagnosis and risk-stratified prognosis assessment, elucidate key molecular pathways underlying glioma pathogenesis and progression, and provide a mechanistic foundation for developing individualized therapeutic interventions and targeted therapeutics.

Funding

This work is supported by grants from National Natural Science Foundation of China (No. 82103480), Zhejiang Provincial Natural Science Foundation (No. LQ22H090018) and Medical Health Science and Technology Project of Zhejiang Provincial Health Commission (No. 2025KY073).

Ethics approval and consent to participate

Not applicable.

Clinical trial number

Not applicable.

Consent for publication

Not applicable.

CRediT authorship contribution statement

Yulong Lan: Writing – original draft, Supervision, Project administration, Funding acquisition, Formal analysis, Data curation, Conceptualization. Shuang Zou: Writing – original draft, Visualization, Software, Methodology, Investigation, Funding acquisition. Xun Wang: Writing – review & editing, Resources, Formal analysis, Data curation, Conceptualization.

Declaration of competing interest

The authors declare no competing interests.

Data availability

Data is provided within the manuscript.