Integrative analysis of single-cell and bulk transcriptome data reveals age-related immune cell alterations in primary glioblastoma associated with prognosis

Abstract

Background

Glioblastoma (GBM) is the most malignant and highly recurrent brain tumor. Although over half of the GBM patients are elderly patients, the understanding of how aging affects GBM progression remains limited.

Methods

Clinical and genomic variation data of GBM patients from TCGA and CGGA databases were used for prognostic analysis. We collected single-cell transcriptome data of 88,908 cells from 13 primary GBM (pGBM) and 12 recurrent GBM (rGBM) patients. Age-related immune cells were identified through cell–cell communication and trajectory analysis. The results were validated by projecting the single-cell transcriptome profiles onto bulk data. Finally, we experimentally validated the results on syngeneic orthotopic models of younger and older mice.

Results

Prognostic analysis indicated the effect of age in pGBM patients is stronger than that in rGBM patients. Moreover, the mutational signatures in pGBM does not affect prognosis. Single-cell RNA sequencing analysis revealed age-related differences in immune cells between pGBM and rGBM, and identified microglia underwent significant cell state changes with aging only in pGBM patients. Next, we validated that high expression of HSPB1 in microglia from older pGBM patients is associated with poor prognosis. Finally, the syngeneic orthotopic model for aged mice exhibited more tumor invasion, with a shorter median survival time. Furthermore, the microglia within the tumor microenvironment (TME) of aged mice showed markedly high expression level of HSPB1.

Conclusions

Our study highlights the crucial role of microglial aging in pGBM, reveals distinct age-related changes of immune cells in the TME between pGBM and rGBM, and offers valuable insights into clinical treatment strategies targeting elderly GBM patients.

Supplementary Information

The online version contains supplementary material available at 10.1007/s00262-025-04206-w.

Introduction

Gliomas, accounting for 80% of all malignant brain tumors [1], are the leading cause of death from primary brain tumors. These highly heterogeneous tumors are thought to originate from oligodendroglial lineage precursor and neural progenitor stem cells [2]. Among all types of gliomas, glioblastoma (GBM) is the most common, the most malignant, has a high recurrence rate [3], and carries the worst prognosis. The relative survival rate of GBM patients for 5 years is only 6.9% [2]. Given the long-standing lack of effective therapeutic options for recurrent glioblastoma (rGBM), the treatment strategies for rGBM have remained largely consistent with those for primary glioblastoma (pGBM), particularly surgery, radiotherapy, and temozolomide [4].

The rapid development of single-cell technology has provided an opportunity for deeper insights into tumor microenvironment (TME) in several cancers, including GBM [5–7], revealing the diversity of malignant cell subtypes and the complexity of interactions between different cell populations. However, single-cell transcriptome studies investigating the cellular composition and states heterogeneity of rGBM remain limited, particularly in comparison to pGBM. In recent years, several studies have begun to reveal the differences of cellular composition and states between pGBM and rGBM at single-cell level [8–10]. These in-depth transcriptome studies contribute to our understanding of the similarities and differences between pGBM and rGBM, facilitating the development of more effective therapeutic strategies.

It is a consensus that age is a risk factor for GBM [11], and more than half of GBM patients are elderly patients [1]. Nevertheless, there is a gap in our understanding of how aging influences the molecular mechanisms underlying GBM progression, particularly in the context of immune cells within the TME. As individuals age, immune cells undergo varying degrees of senescence, leading to cellular dysfunction [12, 13]. Previous studies have shown that age-induced immunosenescence in the TME is closely associated with tumor progression, particularly involving lymphoid cells [14–17]. While in GBM, due to the unique immune environment of the brain [18], the immune cells within the TME are mainly composed of tumor-associated macrophages and microglia (TAMs), with lymphoid cells relatively scarce [19–21]. A recent study demonstrated that aging leads to blood–brain barrier dysfunction in the healthy brain, accompanied by increased infiltration of immune cells other than microglia, while chronic inflammation profile of the aging brain promotes myeloid cell dysfunction [22]. In short, it is essential to investigate the impact of aging on immune cells within the TME of GBM, as this will contribute to the development of age-specific precision therapy.

In this study, we identified the prognostic effect of age in pGBM patients is stronger than that in rGBM patients. Firstly, we performed an integrative analysis of single-cell RNA sequencing (scRNA-Seq) data, comparing the immune components between younger and older patients in both pGBM and rGBM. Furthermore, we identified that microglia experienced significant state changes with increasing age, but only in pGBM. Next, by correlating with bulk transcriptome clinical cohort data, we identified HSPB1 was associated with both microglial aging and poor prognosis only in pGBM patients. Finally, we experimentally validated these findings by using syngeneic orthotopic models of younger and older mice. Our study reveals age-related differences of immune cells between pGBM and rGBM, and contributes to the development of personalized treatments for pGBM patients across different ages.

Materials and methods

Animals

According to the guideline for mouse models in aging research [23], we established syngeneic orthotopic models using male C57BL/6J mice at 3 months and 18 months old. Before the experiment, the mice were acclimatized for one week at the Peking Union Medical College Animal Center. All procedures were authorized by the Institutional Animal Care and Use Committee of the Chinese Academy of Medical Sciences and Peking Union Medical College (ACUC-A01-2023–057).

Cell culture and transfection

The mouse high grade glioma cell line GL261 was kindly provided by the Beijing Tiantan Hospital. The cells were cultured in DMEM (Gibco) supplemented with 10% FBS (ScienCell) and 1% P/S solution (ScienCell), and maintained at 37 ℃ with 5% CO₂. According to the manufacturer’s protocol, the GL261 cells were transfected with lentivirus carrying an empty vector expressing mCherry (Genechem: GV298). Puromycin (2 µg/mL) was used to select for stable cell lines after transfection.

Orthotopic implantation

After anesthetizing each mouse with 1.25% tribromoethyl alcohol (Aibei Biotechnology; 0.02 mL/g body weight), 3 × 10⁵ GL261-mCherry cells in 5 μL saline were injected into the striatum (CPu; 1 mm rostral to the bregma, 2 mm lateral to the midline, 3 mm deep) using a microsyringe (Hamilton). Magnetic resonance imaging (MRI) was used to confirm the existence of the tumor. Tumor-bearing mice were used for survival analysis, and for brain section after 21 days of tumor development.

Tissue section

The mice for brain section were subjected to transcardiac perfusion with 4% PFA-PBS after being anesthetized with the same agent. Subsequently, the tumor-containing brains were fixed in 4% PFA-PBS for ~ 48 h, followed by dehydration in 30% sucrose-PBS solution for another ~ 48 h. After that, the brains were embedded in O.C.T. (SAKURA) and stored at −80 ℃. They were finally cut into 8 μm thick sections at tumor-containing regions using a Leica CM1950 cryostat.

Immunofluorescence staining

The brain sections were desiccated at 50 ℃ for 30 min and then washed twice in 1 × PBS for 5 min each. Subsequently, the sections were incubated at room temperature for 2 h with 5% donkey serum (YEASEN) in buffer (1 × PBS with 0.3% Triton X-100). The tissue sections were then incubated at 4 ℃ overnight with rabbit anti-MMP2 (1:250; Abcam; #ab92536), rabbit anit-TMEM119 (1:200; Proteintech; #27585-1-AP), mouse anti-HSPB1 (1:100; Proteintech; #66767-1-Ig), prepared in the buffer containing 5% donkey serum. After that, the primary antibody mixture was removed and the sections were washed three times in 1 × PBS for 5 min each. The tissue sections were then incubated at room temperature for 1 h with secondary antibody (1:800; Invitrogen; #A21245, #A11001) diluted in the buffer. After three washes with 1 × PBS, the sections were stained with DAPI, mounted, and photographed with a Leica Stellar system (STELLARIS 5). The signal intensity of the immunofluorescence was analyzed using the software LAS X.

Survival analysis

Clinical data for pGBM and rGBM patients were collected and merged from The Cancer Genome Atlas (TCGA) and Chinese Glioma Genome Atlas (CGGA) [24] databases. Samples missing survival and age information were filtered out. A subgroup Cox proportional hazards model (Cox model) was used with age as the key feature to determine its association on survival across different subgroup. Survival differences between different age groups were computed using the Kaplan–Meier method and the log-rank test. Survival analysis was performed using the R package survival (v3.7.0). The interaction P-values for the subgroup Cox model were calculated using the R package jstable (v1.3.5). Hazard Ratios (HR) with 95% Confidence Intervals (CI) for each feature were visualized using the R package forestplot (v3.1.3). Survival curves were visualized using the R package survminer (v0.4.9).

Genomic variation analysis

The GBM simple nucleotide variation data were downloaded from TCGA database, and the total numbers of mutation events were calculated and further correlated with patient survival data. We used the total number of mutation events as the mutational signature for each patient, and the median, 42, was applied to stratify patients into groups with fewer or more numbers of variations for survival analysis. The mutation data of GBM were analyzed and visualized using the R package maftools (v2.21.2) [25] with default parameter settings.

Single-cell RNA sequencing data collection and quality control

We collected four high-quality GBM scRNA-Seq datasets (Supplementary Table 1) containing the information of patient age from published studies: Mei Y et al. (GSE235676) [26], Xiao Y et al. (GSE135045) [27], Yuan J et al. (GSE103224) [28], and Cheng YL et al. (GSE148842) [29]. The selection of these data was also based on additional standards (see in Results). A total of 25 patient samples were obtained after selection, including 13 pGBM samples and 12 rGBM samples.

Data were processed and analyzed using the R package Seurat (v5.1.0) [30] based on the raw gene expression (UMI counts; UMIs, unique molecular identifiers) matrices. Unless otherwise stated, functions are executed using default parameters. We only included cells with more than 200 features and genes expressed in at least 10 cells for subsequent analysis. After removing doublets from each sample individually using the R package scDblFinder (v1.19.1) [31] with parameters ‘nfeatures = 2000’ and ‘dims = 30’, cells with more than 500 and fewer than 40,000 UMI counts, fewer than 6,000 detected genes, and lower than 30% mitochondrial, 40% ribosomal, 0.5% hemoglobinous features were retained.

Single-cell RNA sequencing data integration and analysis

After quality control of each sample, we integrated the scRNA-Seq data of all 25 GBM patients into an aggregate Seurat object using the merge function followed by the JoinLayers function from Seurat. Subsequently, we used the NormalizeData function to convert the UMI counts of each gene into TPM-like values by using the ‘LogNormalize’ method with a scale factor of 10,000. Following that, we used the FindVariableFeatures function with a selection method ‘vst’, selecting 2,000 highly variable genes for the downstream analysis. The expression matrix of highly variable genes was then scaled using the ScaleData function after regressing out the total count of UMIs, the number of detected genes, the percent of mitochondrial genes, and the percent of ribosomal genes. After that, dimension reduction for the scaled data matrix of highly variable genes was performed using the RunPCA function, followed by batch effect correction across different samples using the R package harmony (v1.2.0) [32]. Finally, we identified different cell clusters using the FindNeighbors function for the first 30 components from ‘harmony’ reduction and the FindClusters function (resolution = 0.7). For the resolution, we selected it based on the R package clustree (v0.5.1) [33] and the expression levels of representative features for the cell types frequently observed in GBM scRNA-Seq data. Uniform Manifold Approximation and Projection (UMAP) was used for result visualization.

We used the AverageExpression function to calculate the average expression levels of highly variable genes for each cell cluster and assessed the correlation. The results were visualized using the R package ComplexHeatmap (v2.18.0) [34]. The expression level of gene was visualized using the R package Nebulosa (v1.0.1) [35]. Differentially expressed genes (DEGs) in each cell cluster or age group were identified using the FindAllMarkers or FindMarkers function with ‘min.pct = 0.1’ and the Wilcoxon rank-sum test, together with the R package presto (v1.0.0) for rapid DEGs calculation in large-scale scRNA-Seq data. Gene set enrichment at the single-cell level was performed using the AddModuleScore function from Seurat. Gene set functional analysis of DEGs was performed using Gene Set Enrichment Analysis (GSEA) through the R package clusterProfiler (v4.10.1) [36].

Copy number variations estimation for single-cell RNA sequencing data

The R package infercnv (v1.16.0; https://github.com/broadinstitute/infercnv) was used to estimate copy number variations (CNVs) of each cell in the scRNA-Seq object. To maximize the efficiency of computing CNVs in large-scale scRNA-Seq data and minimize potential bias caused by the proportion differences of reference cells among samples, we followed the developer’s guideline and regrouped the 25 GBM patient samples into 7 groups (cnv_group; Supplementary Table 2), based on the sequencing protocol of each sample and the number of cells after quality control. For each cnv_group, we used the same normal cell types (macrophages, microglia, T cells/NK cells) as a baseline to infer the CNVs of other cells, following the same computational approach. A cut-off of 0.1 was applied for the minimum average read counts per gene among reference cells. Genes expressed in more than 20 cells were sorted based on their loci on each chromosome. A slide window size of 101 genes was used to smoothen the relative expression on each chromosome.

For quantifying CNV scores for each cell, according to Peng J et al.’s method [37], we re-normalized the gene expression values to a range of -1 to 1 for infercnv results of each cnv_group, and then computed the quadratic sum of the adjusted values for each cell as the CNV score. For each cnv_group, the threshold for malignancy was calculated as the sum of median CNV score and three times standard deviation from reference cells. Cells with CNV scores exceeding this threshold were defined as high-CNV cells.

Analysis and annotation of cell sub-populations

We performed a second round of analysis and clustering for the major cell populations that require further recognition. Each cell population was extracted and processed to a new round of UMI counts normalization, highly variable genes selecting, expression matrix scaling, dimensionality reduction, batch effect correction (harmony), using the same parameters as in the initial analysis. For the clustering and annotation of cells, we performed the same approach as in the initial analysis. Briefly, for the myeloid and lymphoid cell major populations, we chose resolutions of 0.8 and 0.7, respectively, for sub-type annotation (see in Results).

Cell–cell communication analysis

We applied the R package CellChat (v1.6.1) [38] on all immune and malignant cell populations for intercellular interaction analysis according to canonical ligand-receptor pairs. The cells were extracted as a new Seurat object and the UMI counts matrix was re-normalized. Then, we ran the pipeline of CellChat with default settings to analyze characteristics of cell–cell communication between younger and older patients in both pGBM and rGBM. We used the projectData function in the pipeline to project the inference results from gene expression to protein–protein interaction networks thereby enhancing reliability.

Developmental trajectory analysis

The diffusion model was used to reduce the dimensions of cellular transcriptome profiles, illustrating the continuous transitions of cell states. For each sub-population, we performed the same pipeline as for the second round of analysis, up to batch effect correction (harmony). Based on the first 30 components from ‘harmony’ reduction of each sub-population, we generated its diffusion model using the DiffusionMap function from the R package destiny (v3.17.0) [39]. A 3D diffusion map of the model was generated using the plot3d function from the R package rgl (v1.3.1).

GBM bulk transcriptome data collection and processing

Bulk RNA sequencing (bulk RNA-Seq) data of GBM patients with age and survival information were downloaded from TCGA (whole transcriptome sequencing) and CGGA (mRNA sequencing) databases. For consistency, both datasets are referred to as bulk RNA-seq in this study. These data were organized and reclassified into pGBM and rGBM clinical cohort datasets. The TPM values were used for integration, while FPKM values were converted to TPM values. We used the R packages FactoMineR (v2.11) [40] and sva (v3.50.0) [41] to evaluate and remove batch effects between the data from TCGA and CGGA datasets.

Linking scRNA-Seq data to bulk RNA-Seq data to reveal prognosis-related genes in specific cell populations

According to the method described by Li et al. [42], we corrected the bulk RNA-Seq data to approximate the gene expression profiles of specific cell populations in the TME. Briefly, the feature gene sets of cell populations were identified using the FindMarkers function (log2FC > 1, adjusted P-value < 0.05 and minimum percentage = 0.3). Then, we used the single sample Gene Set Enrichment Analysis (ssGSEA) algorithm [43] from the R package GSVA (v1.53.3) [44] to quantify the relative cell abundance of the cell populations based on their feature gene sets in the batch effect removed bulk datasets. After that, we adjusted the bulk datasets for the cell abundance using linear regression. For each gene, a linear model was fitted with gene expression as the dependent variable and cell abundance as the independent variable. The residuals from the model were used as the adjusted values. The subsequent analyses in clinical cohorts were all based on the abundance-adjusted bulk cohorts. Gene set expression level scoring was also performed using the ssGSEA algorithm.

Statistical analysis

All statistical analyses and graph generation were performed in R (v4.3.3), with the R package ggplot2 (v3.5.1) used for visualization. Survival analysis was assessed using log-rank test. Pearson’s correlation coefficient was used for correlation analysis. The selection of DEGs was based on adjusted P-values using the false discovery rate (FDR) method. Data were assessed using the Wilcoxon rank-sum test.

Result

The prognostic effect of age in pGBM patients is stronger than that in rGBM patients

We collected and integrated available clinical survival data from 997 GBM patients in TCGA and CGGA databases, with ages ranging from 20 to 80 years old. The other clinical features, which were used for subgroup analysis in Cox model, include tumor progression (primary / recurrent), sex (female / male), race (Asian / White / Black or African American), preoperative treatment status (untreated / pharmaceutical only / radiation only / both), IDH status (mutant / wild-type), 1p19q status (codeletion / non-codeletion), and MGMTp status (methylated / unmethylated). Based on the integrated data, we divided the patients into two groups: younger (20–50 years old) and older (50–80 years old). We then used Cox model to compute the HR for each subgroup with age as the core grouping feature (Fig. 1a). The results indicated that patients older than 50 years old had a higher risk of death in all features, except for those with IDH mutations. According to the latest 2021 WHO classification for glioma subtypes [2, 45], the original IDH mutant GBM is reclassified as WHO-IV astrocytoma, while only IDH wild-type GBM is classified as GBM. Moreover, pediatric and adult GBM patients are now categorized as independent subtypes. We further calculated the P-value for interaction between the age and each subgroup feature. Notably, the interaction P-values between age and race, as well as between age and preoperative treatment status, were statistically significant (P < 0.05), indicating that the impact of age on prognosis varies significantly among patients from different regions and treatment regimens.

Fig. 1.

The prognostic effect of age in pGBM patients is stronger than that in rGBM patient. a Forest plots showing Hazard Ratios with 95% Confidence Intervals and P-values calculated by age-related subgroup Cox model for GBM patients from TCGA and CGGA databases. (b and c) Kaplan–Meier survival curves for pGBM IDH wild-type patients (b) and rGBM IDH wild-type patients (c) from TCGA and CGGA databases within age group at diagnosis (ages 20–40, 40–60, 60–80 years). d Heatmap showing the top mutation events for pGBM patients from TCGA database sorted by increasing age; statistical graph of mutation events for each gene is shown in the right panel, colored by variant classifications. e Pearson’s correlation between increasing age and the number of variations for pGBM patients from TCGA database. f Kaplan–Meier survival curves for pGBM patients from TCGA database grouped by the number of variations using median cutoff 42. In (a and b), * P < 0.05, ** P < 0.01, *** P < 0.001

Interestingly, in the groups of primary and recurrent, we found that although the HRs were greater than 1 for both groups, age had a significant impact on the poor prognosis only for pGBM patients. We further regroup the patients into three age groups: 20–40 years old, 40–60 years old, and 60–80 years old. Kaplan–Meier models of the data showed that the median survival time (MST) significantly decreased with aging in pGBM IDH wild-type patients (Fig. 1b). However, in rGBM IDH wild-type patients, there was no significant decrease in MST (Fig. 1c). A classic hallmark of aging is genomic instability leading to the accumulation of somatic mutations [46]. We used the simple nucleotide variation data from TCGA database and selected 344 pGBM patients aged 20–80 years to analyze the relationship between age, prognosis, and mutational signatures. The results indicated that the top mutation events in pGBM patients did not show an age-related tendency (Fig. 1d). Moreover, the number of genomic variations in pGBM patients significantly increased with aging (Fig. 1e), but a higher number of variations did not correlate with worse prognosis (Fig. 1f).

Integration of scRNA-Seq data and annotation of cell populations

As mentioned above, immunosenescence in the TME is closely associated with tumor progression. To further investigate how aging affects the poor prognosis of pGBM patients and to compare them with rGBM patients, we collected four high-quality scRNA-Seq datasets from previous studies (see in Materials and methods), and analyzed the immune components within the TME at single-cell level. After applying the same quality control and filtering processes to each sample (Supplementary Fig. 1a), we selected 25 scRNA-Seq samples from the four GBM datasets (Fig. 2a; Supplementary Table 1), including 13 pGBM samples and 12 rGBM samples. A total of 88,908 cells were included for subsequent analysis. The ages of the patients ranged from 20 to 80 years, and all patients had IDH wild-type statuses. To minimize the impact of pharmaceutical or radiation therapy on the phenotypes of tumor cells, we only selected pGBM samples that had not received preoperative chemotherapy or radiotherapy. However, due to the nature of rGBM, very few rGBM patients had not received any preoperative treatment. Therefore, among the 12 rGBM patients, only 3 did not report prior treatment, while the remaining 9 had received chemotherapy (temozolomide), radiotherapy, or neoadjuvant therapy.

Fig. 2.

Integration and analyses of pGBM and rGBM scRNA-Seq data identified different cell populations. a Main information of each scRNA-Seq data. b Strategy of batch effect correction for 25 scRNA-Seq data, colored by patients. c UMAP visualization of 88,908 cells from 25 GBM patients, colored by cell types. d Dot plot of representative feature genes among each cell type; dot size represents abundance, and color represents gene expression level. e UMAP visualization of infercnv prediction for all 88,908 cells. f Malignant cell proportions of each cell type. (g and h) UMAP visualizations of myeloid (g) and lymphoid (h) cells selecting from the aggregate scRNA-Seq data, colored by cell subtypes. i Heatmap showing the proportions of immune cell subtypes in scRNA-Seq data sorted by increasing age

We integrated the 88,908 cells in to an aggregate Seurat object and removed the batch effects among patients (Fig. 2b). Subsequently, the Seurat object was divided into 27 clusters through the first round of unsupervised clustering (Supplementary Fig. 1b). Some clusters exhibited similar expression patterns of highly variable genes (Supplementary Fig. 1c–d). After detailed identification and annotation, we observed six populations of malignant cells (highly express SOX2 [28]), myeloid cells, lymphoid cells, oligodendrocytes, fibroblasts, and endothelial cells (Fig. 2c–d). Among these, the malignant cells were further classified into two proliferation groups based on high expression of TAOK1 and MKI67, as well as four resting groups with low expression of proliferation markers. These four groups of malignant cells were classified as mesenchymal-like (MES-like; CHI3L1, ADM), oligodendrocyte precursor cell-like (OPC-like; PDGFRA, OLIG1), astrocyte-like (AC-like; GFAP, AQP4), and neural progenitor cell-like (NPC-like; CD24, DCX). These four subtypes were also recognized as the major malignant cell subtypes in recent GBM transcriptome studies [7, 47].

CNV level contributes to identifying true malignant cells in tumor scRNA-Seq data. Using macrophages, microglia, and T/NK cells as normal cell baseline, we computed the CNVs profiles of all other cells in each cnv_group (Supplementary Fig. 2a). Subsequently, we quantified the CNV level of each cell and successfully divided malignant cells with high-CNV scores from the Seurat object (Fig. 2e). It is important to note that in follow analyses, we considered the populations expressing GBM malignant cell gene features with high-CNV as malignant cells, and those expressing immune cell gene features with low-CNV as normal immune cells in the TME (Fig. 2f).

The immune landscapes between pGBM and rGBM have significant differences

After filtering out the high-CNV score cells, we performed a second round of analysis and clustering for normal myeloid and lymphoid cells, respectively. According to recent studies on the immune landscape of TME in pGBM and rGBM [21, 48], we annotated the immune cell subclusters based on feature genes from the scRNA-Seq data (Supplementary Fig. 2b–c). We identified 17 distinct and meaningful clusters in myeloid cells and 16 clusters in lymphoid cells (including one mixed myeloid-derived cluster and one pDCs cluster) (Fig. 2g–h).

In myeloid cells, we identified five major populations, including proliferating myeloid cells (Mye-cycling: MKI67), microglia (MG: CX3CR1, TMEM119, P2RY12), macrophages (MΦ: only general markers CD14, CD68, FCGR3A, CD163), monocytes (Mono: FCN1, S100A8, S100A9), and dendritic cells (DCs). In addition to these five major populations, we also identified a special group of bone marrow-derived immunosuppressive cells (MDSC: CD274) that highly express interferon-stimulated genes (ISGs; MDSC-ISG). Upon further detailed annotation, we distinguished five sub-populations within MG, including two resting clusters (MG.1, MG.2) and three inflammatory activated clusters. Among these activated MG clusters, one cluster highly express both TNF and IL1A, while another expressed ISGs. Next, we also separated five sub-populations from MΦ, including (1) an inflammatory activated cluster, (2) a resting cluster with relatively high MHC-I expression, (3) an angiogenic cluster, (4) a perivascular macrophage-like cluster, and (5) a phagocytic cluster expressing lysosomal genes. For monocytes, we identified three sub-populations: one cluster with relatively low inflammation and higher expression of MDSC-associated genes (LILRB2, JAML), another cluster with higher inflammation, and a third cluster with high expression of interferon-γ induced monocyte factors CXCL9. Finally, DCs were classified into conventional type 1 and type 2 DCs (cDC1, cDC2).

In lymphoid cells, we identified five major populations, including proliferating lymphocytes (Lym-cycling: MKI67), CD8 + T cells (CD3D, CD8A/B^high, CD4^low), CD4 + T cells (CD3D, CD8A/B^low, CD4^high), B cells (MS4A1, CD79A), and natural killer cells (NK: NCAM1, KLRB1). For B cells, there was no further annotation performed. In NK cells, we distinguished three sub-populations: one cluster with high expression of the NK cell receptor KLRB1; an activated group with high expression of GZMH/B; and an activated group with high expression of chemokines XCL1/3. For T cells, we performed annotation based on the gene features related to exhaustion, activation, memory and effector functions. According to the T cell dynamics classification model [49], we further classified CD8 + T cells into six sub-populations, including (1) an activated cluster with significantly higher expression of interferon-γ (IFNG), (2 & 3) two early activated cell clusters with GZMK/H^high and GZMB^low (CD8 + Tearly.act.1/2), (4) a cluster specifically expressing interferon-stimulated genes (IFIT1/2/3, ISG15), (5) a terminally exhausted T cell cluster (CD8 + Tte: PDCD1^high, HAVCR2^high, TCF7^low), and (6) a tissue-resident memory T cell cluster (CD8 + Trm: ZNF683^high). In CD4 + T cells, we identified two sub-populations: a regulatory T cells cluster (Treg: IL2RA, FOXP3) and a central memory T cells cluster (CD4 + Tcm: IL7R^high, CCR7^high). Additionally, we identified a particular sub-population of CD4-CD8- effector memory T cells (CD4-CD8-Tem: IL7R^high, CCR7^low).

For the immune cell types clustered in our aggregate Seurat object, we did not observe a notable trend of cell proportion associated with the increasing age (Fig. 2i; Supplementary Table 3). In the TME of pGBM patients, the immune cell populations were largely myeloid cells, with a small proportion of lymphoid cells. In contrast, the proportion of lymphoid cells in the TME of rGBM was notably higher may due to the preoperative treatment (Figs. 2i, 3a). Further analysis aimed to determine which immune cells experience significant changes with the increasing age.

Fig. 3.

Interactome analyses among malignant and immune cells revealed age-related differences between pGBM and rGBM. a Cell proportions of myeloid and lymphoid cell subtypes in each patient, colored by patients. b Number of inferred interactions of pathways for malignant and immune cells in each patient group. c Interaction strength of pathways for malignant and immune cells in each patient group. d Venn diagram showing the significant core signaling pathways from the TME of elderly patients in pGBM and rGBM scRNA-Seq data. e Heatmaps of different GBM groups showing the overall interaction strength of the overrepresented pathways in Fig. 3b; the bottom histograms represent the sum of normalized interaction strength per cell subtype

Malignant cells are more active and immune cells are more suppressed in older pGBM patients, in contrast to older rGBM patients

The presence of Mono-CXCL9, MDSC-ISG, and CD4-CD8-Tem was only observed in a few patient samples and may be due to patient-specific differences. All other cell types were present at varying proportions in the TME of different GBM patients (Fig. 3a). Next, we used CellChat to compare the potential ligand-receptor interactions in detail between malignant cells, myeloid and lymphoid cells in younger and older patients. We re-integrated all malignant cells and immune cells (excluding Mono-CXCL9, MDSC-ISG, and CD4-CD8-Tem cells) into a new Seurat object. After re-normalizing the UMI counts matrix, we divided the samples into four groups based on the cutoff of 50 years old: the pGBM younger group (5 samples) and older group (8 samples), and the rGBM younger group (7 samples) and older group (5 samples). In both pGBM and rGBM, aging induced more total ligand-receptor interactions (Fig. 3b). Unexpectedly, the interaction strength in older pGBM patients was only about half that in the younger patients, while in rGBM, the strength increased with the rise of total interactions (Fig. 3c).

We further ranked the CellChat’s curated signaling pathways based on the ligand-receptor interaction scores between older and younger patients (Supplementary Fig. 3a–b), and compared the dominant cell signaling pathways in the TME between older pGBM and older rGBM patients (Fig. 3d). In the TME of older pGBM patients, signaling pathways associated with tumor cell proliferation and invasion were significantly upregulated. While in the TME of older rGBM patients, the dominant pathways were mainly related to immune cell function and activity. Moreover, we performed an analysis of the interaction strength between different cell populations in older compared to younger patients. We observed that in pGBM patients, the interaction strength related to malignant cell populations was significantly increased in older patients, while the interaction strength between immune cell populations was notably reduced, with microglia being the major contributors (Supplementary Fig. 4). Among them, the changes in the MG-TNF/IL1A cells were the most prominent. Furthermore, we analyzed the strength changes of the signaling pathways which significantly enhanced in older patients. As shown in the bottom histogram (Fig. 3e), older pGBM patients exhibited a marked increase in the strength of signaling pathways related to cell proliferation and migration for malignant cell populations. However, immune cells showed a significant decrease in the strength of pathways related to their functions, particularly in microglia, where the changes in MG-TNF/IL1A cells were the most evident. On the contrary, the situations were different even reversed in older rGBM patients (Fig. 3e, Supplementary Fig. 4).

Microglia in pGBM scRNA-Seq data indicates age-related changes

To identify which immune populations in the TME of GBM are associated with aging in the changes of cell states, we applied a diffusion model to re-dimension the transcriptome and mapped the age information of patients onto each cell. The color scheme for each cell sub-population is consistent with that in Fig. 2g–h to ensure model accuracy. Based on the age information, we observed a most pronounced age-related trend in the microglia of pGBM patients (Fig. 4a). Other cell populations in both pGBM and rGBM patients did not show detectable age-related cell state changes in the diffusion map. (Supplementary Fig. 5a–b), suggesting that the transcriptomic states of these cells did not exhibit significant differences as individuals age. Furthermore, we performed a diffusion model analysis for each microglial sub-population. We observed age-related changes in the cellular states of all microglial subtypes in pGBM (Fig. 4b), with the most prominent changes occurring in the MG-TNF/IL1A, consistent with the results presented in Fig. 3.

Fig. 4.

The cellular states of microglia underwent distinct age-related changes between pGBM and rGBM. a 3D diffusion map of microglia in pGBM and rGBM scRNA-Seq, colored by patient age; the smaller projection in the lower right panel is colored by cell subtype, as shown in Fig. 2g. b 2D diffusion map of microglia subtype, colored by patient age. c Volcano plot showing age-related DEGs from microglia in pGBM, with log2 Fold-Change in expression versus the difference in the percentage of cells expressing the gene (ΔPercentage Difference). d Bubble chart displaying the significantly altered signal pathways for GSEA-GO enrichment analysis of age-related DEGs from microglia in pGBM. e Heatmaps showing single cell level enrichment of M1/M2 phenotypes and microglial identity gene signatures across different age groups in microglia between pGBM and rGBM. f Heatmap showing the Pearson’s correlation of average gene expression level between microglia and monocytes in pGBM and rGBM. g Venn diagrams showing the number of up- and down- regulated genes that overlapped in the replication analysis. h Scatter plot displaying the differences expression of overlapped gene set in Fig. 4f; point color represents P-value calculated by the Wilcoxon rank-sum test, and size represents ΔPercentage Difference

We further identified the DEGs of microglia in older patients compared to younger patients in both pGBM and rGBM, respectively. Using 50 years old as the cutoff, the number of DEGs for microglia in pGBM was higher than that in rGBM (|log2FC|> 1, adjusted P-value < 0.05) (Fig. 4c, Supplementary Fig. 5c; Supplementary Table 4–5). GSEA of Gene Ontology (GO) terms revealed that microglia in older pGBM patients exhibited a broader range of significantly enriched signaling pathways (Fig. 4d), whereas few such changes were observed in older rGBM (Supplementary Fig. 5d). For Biological Processes, microglia in older pGBM patients mainly upregulated the pathways related to oxidative stress metabolism, while downregulating many pathways associated with myeloid cell migration and chemotaxis (Fig. 4d; Supplementary Table 6). These findings are consistent with previous reports on microglia aging [50, 51]. Nevertheless, in older rGBM patients, the significantly enriched pathways were not related to myeloid cell functions. (Supplementary Fig. 5d; Supplementary Table 7).

M1/M2 phenotypes of TAMs in the TME have been used as a classical identification basis for their anti-tumor and pro-tumor functions, while many in vivo studies have shown that TAMs actually exist in a more complex intermediate state, with both M1 and M2 phenotypes coexisting [20]. We collected the core gene sets corresponding to M1 and M2 phenotypes from previously published TAMs transcriptional atlases [52]. For microglia, we observed that the M2 phenotype was present in patients across all age groups in both pGBM and rGBM, while the M1 phenotype is more pronounced in rGBM than in pGBM, particularly in older patients over 50 years old (Fig. 4e). This suggests that the anti-tumor activity of microglia in the TME of older patients is stronger in rGBM compared to pGBM. Moreover, we observed a marked reduction in the expression of microglial identity signatures in older pGBM patients, but not in older rGBM (Fig. 4e). We then re-integrated microglia and monocytes. After re-normalizing the UMI counts matrix and re-selecting highly variable genes, we compared the cell states similarity of these two cell populations by age group. The results showed that the cell state of microglia from older pGBM patients are more similar to monocytes, whereas microglia from older rGBM patients did not display such a phenotype (Fig. 4f). Furthermore, we compared the DEGs identified from the microglia in pGBM with the age-related gene set of microglia from Lopes KP et al.’s research [51]. Although the number of genes that overlapped between the gene sets was small (Fig. 4g), the significantly altered genes were broadly expressed (Fig. 4h) and associated with microglial functions. In microglia, the expression of S100A4 increases the pro-inflammatory phenotypes of microglia [53], which may drive progression of GBM [54]. Previous research has demonstrated that high level of LGALS1 in microglia deactivates M1 phenotype while promoting polarization toward an M2 phenotype [55]. There are few reports on the expression of TMIGD2 in myeloid cells or microglia.

High expression of HSPB1 in microglia from older pGBM patients leads to poor prognosis

Previous studies have shown that cell type profiles derived from scRNA-Seq data can be projected to bulk transcriptome data and reflect clinical outcomes in cancer [42, 56]. Therefore, we aimed to combine clinical cohort data from pGBM and rGBM patients to identify core genes in microglia associated with aging and poor prognosis. After merging and removing batch effects, we obtained bulk RNA-Seq data for 505 cases from TCGA and CGGA databases, including patient age and clinical survival information. These included 365 pGBM patients (147 from TCGA, 218 from CGGA) and 140 rGBM patients (13 from TCGA, 127 from CGGA). Firstly, we identified the feature gene set for microglia (containing 407 genes; Supplementary Table 8) in a Seurat object containing malignant cells and normal immune cells. This gene set was then used to correct the bulk datasets to obtain the microglia abundance-adjusted bulk cohorts for pGBM and rGBM, reflecting the gene expression of this cell type in the TME.

Next, we calculated age-related DEGs scores (Age Score) for microglia in the pGBM and rGBM adjusted bulk cohorts. The Age Score was defined as the age-related upregulated gene expression level minus the downregulated gene expression level, and the age-related DEGs were identified from the corresponding single-cell data. To refine the list of age-related DEGs, we progressively filtered genes based on the percentage of microglia expressing each gene (pct.1/2). Through this stepwise refinement, we ultimately identified a core age-related gene set from microglia in pGBM (Fig. 5a; Supplementary Table 9) that was positively associated with both increasing age and poor prognosis (Fig. 5b). As expected, the core age-related gene set was not significantly associated with either increasing age or prognosis in the adjusted rGBM bulk cohort (Fig. 5c). Notably, no such gene set was derived from the rGBM scRNA-Seq data and adjusted bulk cohorts.

Fig. 5.

HSPB1 from microglia were associated with aging and poor prognosis in pGBM. a 3D diffusion map of microglia in pGBM scRNA-Seq, colored by Age Score. (b and c) Pearson’s correlation and Kaplan–Meier survival curves showing the correlation of Age Score with aging and prognosis in pGBM (b) and rGBM (c) adjusted bulk cohort. d Scatter plot showing the results of HSPB1 and ANXA1 in age-related differential gene analysis for microglia in scRNA-Seq data. e 2D diffusion map visualizations of HSPB1 expression levels for microglia subtypes in pGBM. f Scatter plot showing the results of HSPB1 in age-related differential gene analysis for microglia subtypes in pGBM scRNA-Seq data. (g and h) Pearson’s correlation and Kaplan–Meier survival curves showing the correlation of HSPB1 expression levels with aging and prognosis in pGBM (g) and rGBM (h) adjusted bulk cohort

We then calculated, separately for pGBM and rGBM, the correlation between each gene in the core gene set and patient age, as well as between each gene and prognosis, based on adjusted expression levels. The results indicated that no genes in rGBM were significantly associated with both increasing age and prognosis, whereas in pGBM, we identified HSPB1 and ANXA1 as two key genes correlated with both patient age and poor prognosis. As expected, HSPB1 and ANXA1 were specifically upregulated in microglia from older pGBM patients, but not from older rGBM patients (Fig. 5d). Although previous studies have shown that the expression of ANXA1 promotes M2 polarization and facilitates microglial migration [57], it is not as broadly expressed as HSPB1 in aggregate scRNA-Seq object (Fig. 5d). Therefore, we focused on HSPB1 in the subsequent analyses.

We observed that, except for the MG.1 subtype, HSPB1 expression level exhibited a clear age-related upregulation across the remaining four microglial subtypes (Fig. 5e–f). In MG.1 from pGBM, HSPB1 was broadly expressed in both younger and older patients. In the adjusted pGBM bulk cohort, we validated that HSPB1 expression levels were positively correlated with aging, and the high expression levels were associated with poor prognosis (Fig. 5g). But in the adjusted rGBM bulk cohort, the HSPB1 expression levels did not show an association with patient age, and high expression levels did not affect prognosis (Fig. 5h). Our results suggest that HSPB1 may play a distinct role in the TME between pGBM and rGBM patients, and could serve as potential microglia-related therapeutic target to improve the prognosis of older pGBM patients.

Aged mice bearing high-grade glioma exhibit poorer prognosis and increased microglial HSPB1 expression in the TME

Given that the TME of older pGBM patients who received no preoperative chemotherapy or radiotherapy closely resembles that of natural aged individuals, we employed naturally aged C57BL/6J mice as hosts for high-grade glioma syngeneic orthotopic models. We established the models in younger (3 months) and older (18 months) C57BL/6J mice using the GL261 cell labeled with mCherry (Fig. 6a). The MST of tumor-bearing older mice was significantly shorter than that of younger mice (Fig. 6b). Matrix Metallopeptidase 2 (MMP2) is a classic marker of invasion in GBM [58]. After immunofluorescence staining, we confirmed that the expression level of MMP2 was notably higher in the tumor tissues of older mice (Fig. 6c–d).

Fig. 6.

Tumor-bearing elderly mice exhibit more invasive orthotopic tumors and increased microglial HSPB1 expression in the TME. a Experimental strategy of high-grade glioma syngeneic orthotopic models followed by survival analysis and immunofluorescence staining; original figure created with BioRender.com. b Kaplan–Meier survival curves for tumor-bearing mice at 3 months and 18 months old. c Immunofluorescence staining of MMP2 on tumor in situ sections. d Statistical graph of MMP2-level for each section from different mice. (e and f) UMAP visualizations of microglia population (e) and expression levels of TMEM119 (f) in myeloid cells from the aggregate scRNA-Seq data. g Immunofluorescence staining of TMEM119 and HSPB1 on tumor in situ sections. h Statistical graph of the proportion of microglia that highly expressed HSPBI for each section from different mice. Statistical graphs are presented as the mean ± SEMs; statistical differences were calculated by the Wilcoxon rank-sum test

TMEM119 is a transmembrane protein specifically expressed on the surface of microglia. Consistent with our scRNA-seq data, we observed that TMEM119 was broadly and exclusively expressed in microglia (Fig. 6e–f). Therefore, we used TMEM119 to identify microglia in the TME via immunofluorescence staining. Co-staining with anti-HSPB1 allowed us to quantify the proportion of microglia that highly expressed HSPB1 within the TME. The result demonstrated that microglia within the TME of older mice expressed significantly a higher level of HSPB1 (Fig. 6g–h), which was consistent with our bioinformatics findings described above.

Discussion

GBM is the most common and highly recurrent malignant brain tumor, and the cellular composition and states within the TME differ between pGBM and rGBM [8–10]. Since over half of GBM patients are elderly [1], precision treatment and care for elderly patients are crucial [11]. However, the understanding of how aging affects GBM progression remains limited.

The accumulation of somatic mutations is one of the most notable hallmarks of aging [46]. Although we found that the number of genomic variations increased in older pGBM patients, this did not lead to worse prognosis. Further in-depth analysis in scRNA-Seq data revealed that aging-induced transcriptome changes of cells in the TME were the key drivers of poor prognosis in older pGBM patients. Aged immune cells and extracellular matrix in the TME plays a critical role in tumor progression [59]. As for GBM, microglia are the special and predominant immune cells in the TME [19, 20], and we found that microglial dysfunction was most pronounced in older pGBM patients, whereas no significant changes were observed in older rGBM patients (Fig. 3e, Supplementary Fig. 4). Recent studies have established age-related gene expression models for microglia from various perspectives [50, 51, 60], and revealed that aged microglia exhibit upregulated lipid metabolism signaling, along with downregulated cell motility and polarity signaling. A previous experimental study demonstrated lipid droplet accumulation and defective phagocytic function in aged microglia, as well as an increase in reactive oxygen species and proinflammatory cytokines [61]. Nevertheless, most of the samples used in these studies are derived from naturally aging humans and mice, as well as patients with psychiatric disorders, with a notable lack of samples from GBM and other brain tumors.

In our preliminary clinical data analysis, we found the prognostic effect of age in pGBM patients is stronger than that in rGBM patient. We also indicated that, although the mutational signatures in pGBM tumor tissue significantly increased with aging, it did not affect patient prognosis (Fig. 1). In addition, contrary to previous reports [62], we observed that the MST of 18 months old mice was significantly shorter than that of 3 months old mice in the high-grade glioma syngeneic orthotopic models (Fig. 6a–b). These results prompted us to further investigate changes in cellular composition and states within the TME.

We found the TME in older pGBM patients is more immunosuppressive. In contrast, immune cells were more active in older rGBM patients, particularly those microglia that were suppressed in pGBM (Supplementary Fig. 4). In the TME of older pGBM patients, signaling pathways associated with tumor cell proliferation and invasion were significantly upregulated, including the colony stimulating factor 3 (CSF3), contactin (CNTN), tenascin-C (TNC), CD155, heparin-binding growth factor pleiotrophin (PTN), the NRG1/ErbB4 axis, the HGF/c-Met axis, cadherin 5 (CDH5), and c-KIT-related pathways [63–73]. While in the TME of older rGBM patients, the dominant pathways were mainly related to immune cell function and activity, including antigen presentation pathways (MHC-II), complement system-related pathways, various immune cytokines that promote TME inflammation and immune cell recruitment [74–76], TNF pathway regulating immune and inflammatory responses [77], type II interferon-related pathways (IFN-II) coordinating immune responses in the TME [78], and immune regulatory factors (CD86, CD39, CD45). These distinct differences suggest that, as individuals age, immune cells in the TME of pGBM and rGBM undergo different changes. We hypothesize that this may be due to the significantly higher proportion of lymphoid cells in rGBM compared to pGBM. Although GBM is generally considered an immune-excluded “cold tumor” [20], the proportion of lymphoid cells in rGBM significantly increases following treatment [26, 79]. Since most rGBM patients are inevitably subjected to preoperative treatment, this could affect the transcriptome profiles of malignant and immune cells from scRNA-Seq data, potentially masking the differences caused by natural aging in rGBM patients. Thus, therapy effects may dominate over age-related changes.

In the subsequent analysis, we identified microglia with the most significant changes in intercellular interaction strength observed in older pGBM patients. The diffusion map of trajectory analysis revealed that each microglia subtype showed a clear age-related change in cellular states in pGBM patients. In contrast, the cell populations in rGBM patients did not show a pronounced and detectable age-related cell state change (Fig. 4a–b, Supplementary Fig. 5a–b). Although the age-related transcriptome changes in microglia from our scRNA-Seq data were generally similar to the signaling pathways primarily enriched in previous studies [50, 51], the overlap of DEGs intersection was modest. This may suggest that the effects of aging on microglia are not entirely the same under different physiological conditions, such as under healthy individuals, neuropsychiatric disorders, and brain tumor.

Moreover, we used the feature gene set of microglia and the same pipeline to adjust pGBM and rGBM bulk transcriptome data. We first identified an age-related gene set in microglia from pGBM (Supplementary Table 9). This gene set remained positively associated with aging and was predictive of poor prognosis in the adjusted pGBM bulk cohort but not in the rGBM cohort. Notably, similar associations were not observed in the rGBM data. Further analysis highlighted HSPB1 as a key gene within this age-related gene set. We identified HSPB1 in the microglia from pGBM scRNA-Seq data was associated with both aging and poor prognosis for pGBM patients. No such results were found in the rGBM patients, and HSPB1 was also not associated with aging or poor prognosis in the adjusted rGBM bulk cohort. High expression of HSPB1 in microglia or macrophages represents a high level of neuroinflammation and a pro-tumor environment [80–82]. These results further support that microglia in the TME of pGBM and rGBM may experience distinct age-related changes.

Finally, we performed validation of our bioinformatics findings using a high-grade glioma syngeneic orthotopic mouse model. We first observed that orthotopic tumors in older mice were more invasive, which may contribute to their shorter MST. In addition, immunofluorescence revealed that microglia from the TME of older mice showed significantly increased HSPB1 expression, demonstrating that the aged mouse model could partially recapitulate features observed in older pGBM patients.

Investigating the TME in older GBM patients are critical for achieving precision therapy for elderly patients. Our study highlights the crucial role of microglial aging in pGBM and reveals distinct age-related changes of immune cells in the TME between pGBM and rGBM. We then identified HSPB1 from microglia as highly expressed only in older pGBM patients, and its higher expression was associated with poorer prognosis in the patients. By performing a workflow that projects scRNA-Seq data to bulk RNA-Seq data, we explained why age has a stronger impact on prognosis in pGBM than in rGBM patients. We further validated our bioinformatics findings using high-grade glioma syngeneic orthotopic models in older and younger mice. However, our study does not investigate the mechanisms of HSPB1 in microglia. The current findings are based on correlations of gene expression and survival analysis. The role of HSPB1 in the microglia within the TME of GBM requires further experimental validation. In addition, our study does not confirm whether microglia in older patients exhibit high expression levels of HSPB1 in additional clinical samples. Although the results were validated in mouse models, the lack of validation in clinical samples may limit the translational impact of the study. Validation using human specimens will need to be performed in future studies. Moreover, we primarily focus on immune cells, but the aging of extracellular matrix in the TME also plays a significant role in tumor progression [59]. Additionally, some studies have suggested that therapy-induced senescence in GBM can lead to drug resistance, ultimately promoting recurrence [83–87]. A recent study identified a 13-gene set associated with both aging and poor prognosis, based on GBM bulk RNA-Seq data [88]. And we do not observe significant changes in the proportion of immune cells with aging (Fig. 2i), while in Wu S et al.’s work [89], they found that more extracranial monocyte-derived macrophages infiltrate in older GBM patients and leads to a poor prognosis. In conclusion, our study provides insights into the differential impact of aging on pGBM and rGBM patients and offers valuable perspectives for future individualized treatment strategies in pGBM patients of different ages.

Supplementary Information

Below is the link to the electronic supplementary material.

Author contributions

Zefan Jing: writing—original draft, conceptualization, data curation, formal analysis, investigation, methodology, software, validation, visualization. Bojun Qiu: conceptualization, investigation, validation. Chenyang Ai: investigation, validation. Chunhui Wang: investigation, validation. Xinrun Wang: investigation, validation. Boyang Li: investigation, validation. Lin Hou: project administration, supervision. Bin Yin: project administration, supervision. Wei Han: Writing—review and editing, conceptualization, funding acquisition, project administration, resources, supervision. Xiaozhong Peng: Writing—review and editing, conceptualization, funding acquisition, project administration, resources, supervision.

Funding

This study was supported by the National Key R&D Program of China (2022YFC3401000, 2022YFA1103803), the CAMS Innovation Fund for Medical Sciences (CIFMS) grant (2021-I2M-1–034, 2021-I2M-1–024), the National Natural Science Foundation of China (82173373) and State Key Laboratory Special Fund (2060204).

Data and codes availability

All public scRNA-Seq data for GBM patients are available from GEO database under accession numbers GSE235676, GSE135045, GSE103224 and GSE148842. Bulk RNA-Seq data and clinical data for GBM patients were obtained from TCGA and CGGA databases. The codes are available on request from the authors. No custom code or software was generated in this study. All analyses were conducted using default pipelines or functions, with parameter settings detailed in Materials and methods.

Declarations

Conflict of interest

The authors declare no conflict of interest.

Consent to participate

All public data were obtained from GEO, TCGA and CGGA databases. Ethical approval for the use of human samples was obtained by the respective data providers. All animal experimental procedures were authorized by the Institutional Animal Care and Use Committee of the Chinese Academy of Medical Sciences and Peking Union Medical College (ACUC-A01-2023–057).

Consent for publication

Not applicable.