2.1 Patients and data sets

A total of 1978 glioma samples were enrolled in this study. The Cancer Genome Atlas (TCGA) mRNA sequencing data set, which contains a total of 960 glioma samples was downloaded from public databases (https://portal.gdc.cancer.gov/) and used as discovery cohort.¹⁸ The Chinese Glioma Genome Atlas (CGGA) data set1 and data set2 (including 693 and 325 samples, respectively) were served as independent validation cohorts. In CGGA data sets, the transcriptome sequencing data and clinical information were downloaded from CGGA databases (http://www.cgga.org.cn). The single-cell RNA sequencing data were also obtained from CGGA databases. The details of establishing and managing CGGA databases have been mentioned in our previous studies.¹⁹ This study was approved by Institutional Review Board (IRB) of Beijing Tiantan Hospital Affiliated to Capital Medical University, and wrote informed consent was obtained from each patient. GSE16011 data set was enrolled in this study to further validated the prognostic value of the pathway-based subtypes.²⁰ The cohort deposited in SRA PRJNA482620 was enrolled in this study to investigate the association between risk signature and response to PD1 inhibitor treatment.²¹ Besides, the GSE121810 data set was employed to find an optimal immunotherapy strategy for patients with gliomas.²²

2.2 Identification of TME-related subtypes of gliomas

The enrichment score of TME-related signalling pathways were calculated by R package ‘GSVA’. The median absolute deviation (MAD) was used to assessed the underlying variation among patients for enrichment score of all signalling pathways. Unsupervised clustering was performed using the R package ‘ConsensusClusterPlus’ for cluster discovery based on the activation of hypoxia response, metabolism and immune-related signalling pathways.²³ The optimal K was selected by cumulative distribution function (CDF) curves and consensus heatmap.

2.3 Construction of the risk signature

The R package ‘limma’ was applied to identify differentially expressed genes among four clusters.²⁴ Then, the genes with the highest expression level in Cluster3 were selected to construct a Cluster3-related risk signature. Subsequently, the least absolute shrinkage and selection operator (LASSO) analysis, which was suitable for high-dimensional regression analysis, was employed to identify the most valuable predictive genes to construct an optimal signature with R package ‘glmnet’.²⁵ The risk score for each patient was calculated as following formula: $$$ Risk\ score={\sum}_{i=1}^n\beta ixi $$$. In this formula, $$$ \beta i $$$ was the regression coefficient obtained from LASSO Cox analysis and $$$ xi $$$ was the relative expression level of each selected genes.

2.4 Immunohistochemical staining

A total of 20 paraffin-embedded glioma tissues from CGGA database were collected for immunohistochemical (IHC) analysis. Anti-CA9 [1:500, 11071-1-AP, Proteintech (Wuhan, Hubei, PR China)], C20orf20 (1:500, 26040-1-AP, Proteintech) and anti-FOSL1 [1:500, A5372, ABclonal (Wuhan, Hubei, PR China)] were used to evaluate the hypoxia response of glioma samples. Anti-CD163 [1:500, ab182422, Abcam (Burlingame, CA, USA)] was used to detect M2 macrophages. Anti-CD4 (1:500, ab133616, Abcam) was used to detect CD4+ T cells. Anti-PKM2 [1:500, 15822-1-AP, Proteintech (Wuhan, Hubei, PR China)] and anti-DPYD (1:500, 27662-1-AP, Proteintech) were used to estimate glucose metabolism and pyrimidine metabolism of glioma samples, respectively. Briefly, after deparaffinizing, the four-micrometre-thick sections were boiled with EDTA antigen retrieval buffer. Then, the sections were incubated with primary antibodies overnight at 4°C. After that, the sections were incubated with propriate secondary antibody at room temperature for 1 h. Finally, the sections were stained with DAB (ZSGB-BIO, Beijing, P.R. China) and haematoxylin (Solarbio, Beijing, PR China). Finally, place the tissue sections under a microscope with a magnification of 20× for observation and capturing images. The stained slides were individually evaluated by two experienced pathologists.

2.5 Bioinformatic analysis

Graph learning-based dimensionality reduction analyses were performed to further explore the intrinsic distribution of individual patients. The discriminative dimensionality reduction with trees (DDRTree) was used to visualize the TME landscape, and the plot cell trajectory function was calculated by monocle with R package ‘monocle’.^{26, 27} Cases with somatic mutations and somatic copy number alternations (CNAs) were downloaded from TCGA database to investigate the differences of genomic alterations among four clusters. The CNAs were analysed by GISTIC 2.0.²⁸ The stromal score and immune score, which represents the degree of stromal and immune cells infiltration in glioma samples were calculated by ESTIMATE algorithm with R package ‘estimate’.²⁹ The tumour purity of each sample was obtained based on the methods proposed by Yoshihara K et al.²⁹ Principal components analysis (PCA) was applied to detect the differences of transcriptome profiles among groups.³⁰ Gene ontology (GO) analysis and gene set enrichment analysis (GSEA) were carried out for functional annotation.^{31, 32} Receiver operating characteristic (ROC) curve analysis was performed to predict patients from specific cluster, and timeROC curve analysis was performed for 1-, 3- and 5-year overall survival (OS) prediction.^{33, 34} The R package ‘pROC’ and ‘timeROC’ was used in ROC and timeROC curve analysis, respectively.

The scRNA-seq data were analysed by R package ‘Seurat’. The CellPhoneDB software was used to identified the significant ligand-receptor interactions between tumour cell and immune cells.

2.6 Statistical analysis

In this study, Student's t-test and one-way ANOVA were conducted to estimate statistical differences among groups. Pearson's correlation analysis was performed to detect the correlation between two continuous variables. The Kaplan–Meier (K–M) survival curves analysis with long-rank test was performed to investigate survival distributions of stratified patients. Cox regression analysis was used to evaluate the independent prognostic value of the risk signature. The statistical analyses were mainly conducted using R language (version 3.6.2) and SPSS (version 16.0). p < 0.05 was regarded as statistically significant.

3.1 Consensus clustering identifies four subtypes of gliomas

Hypoxia, immune responses, and cell metabolism were key components of TME of gliomas. In order to stratify the glioma patients by the differences of TME characteristics, 277 related signalling pathways were obtained from MSigDB (Supplementary Figure S1A,B, Supplementary Table S1) and our approach and workflow for this study were summarized in Supplementary Figure S1C. The enrichment score of each pathway was calculated by GSVA. After that, consensus clustering analysis of 166 pathways with highly variable enrichment score (MAD >1) in TCGA RNA sequencing data set identified four main subtypes, the optimal number of clusters (k = 4) was assessed by cumulative distribution function (CDF) curves and consensus matrices (Figure 1A–C). Then, we preformed SigClust analysis to evaluate the cluster significance and found that all cluster boundaries were statistically significant (Figure 1D). The pathways most representative of the subtypes were identified based on their positive silhouette width, indicating higher similarity to their own class than to any other class member (Figure 1E). Consistent subtypes were also obtained in two independent CGGA RNA sequencing and GSE16011 data sets (Supplementary Figure S2A–C). According to the core pathways of each subtype (Supplementary Table S2) and the functional analysis we mentioned blow, C1 subtype was defined as amino/fatty acid metabolism (AFM, marked in red throughout the manuscript, respectively), C2 subtype was defined as cytokine secretion suppression (CSS, green). C3 subtype was defined as hypoxia/immunosuppression (HIS, brown), and C4 subtype was defined as glucose metabolism (GLU, blue).

In TCGA data set, the Sankey diagram showed significant differences in clinicopathological characteristics among four subtypes (Figure 1F). PCA was performed based on the enrichment score of signalling pathways in TCGA data set. The results showed that cases in four clusters tended to distribute in different directions. (Figure 1G). Besides, the K-M survival analysis suggested that the group membership of each cluster was associated with distinct survival characteristics (Figure 1H). Among them, AFM and CSS subtypes tended to have a better clinical outcome, while the prognosis of HIS and GLU subtypes was poor (Figure 1H). Furthermore, between the HIS and GLU subtypes, patients with the HIS subtype exhibit a worse prognosis (Figure 1H). The reproducibility of these findings was also assessed in CGGA and GSE16011 data sets (Supplementary Figure S2F–K).

To explore the TME landscape of gliomas, we further applied the graph learning-based dimensionality reduction analysis to the enrichment score of TME-related signalling pathways.^{26, 27} We found that the integral distribution of AFM subtype was opposite to that of HIS and GLU subtypes (Figure 1I). Meanwhile, AFM and CSS subtypes were distributed in the same direction, suggesting these two subtypes may have similar TME characteristics. Besides, the results also revealed that compared with TCGA subtypes and DDRTree subtypes, our pathway-based classification could more efficiently stratify gliomas patients with TME characteristics (Figure 1I–K). By performing ROC curve analyses, we also found that the TME pathway-based subtypes had the highest prognostic value compared with TCGA subtypes and DDRTree subtypes (Supplementary Figure S2I).

3.2 The subtypes associated with distinct patterns of genomic alterations

Somatic mutations and copy number alterations (CNAs) have significant impact on the biological characteristics and clinical prognosis of gliomas.^{35, 36} Thus, samples with somatic mutations and CNAs information in TCGA data set were enrolled in this study to further explore the molecular mechanisms influencing the hypoxia response, immune response, and metabolic regulation of gliomas. After comparing the frequency of mutations among samples in four subtypes, we found that more somatic mutations were revealed in samples from HIS and GLU subtypes (Figure 2A). However, the IDH1 mutation and ATRX mutation tended to be enriched in AFM and CSS subtypes (Figure 2A). Besides, The CIC mutation and FUBP1 mutation were significantly enriched in AFM subtype (Supplementary Figure S3A). A higher incidence of TTN mutation was observed in both CSS, HIS and GLU subtypes (Supplementary Figure S3B–D). Meanwhile, the results also showed that the cases from HIS subtype had the highest PTEN mutation rate and the lowest TP53 mutation rate (Supplementary Figure S3C). Furthermore, the CNAs among cases in four clusters were also investigated in TCGA data set. We found that the 1p/19q codeletion, as a representative genomic alteration in oligodendroglioma was mainly observed in cases from AFM (Figure 2B). However, chromosome 7 amplification accompanied chromosome 10 loss, which were usually considered as a genomic hallmark of GBM tended to be enriched in HIS and GLU subtypes (Figure 2B). For cases from CSS, the CNAs were insignificant, suggesting CNAs may not play an essential role in this subtype of gliomas. These findings revealed that cases in four subtypes were inconsistent on genomic alterations, which may provide new insights into the hypoxia response, immune response and metabolic regulation of gliomas.

3.3 The four subtypes differ in tumour immunologic microenvironment

To explore the immune heterogeneity among the four subtypes, immune-related signatures published recently were included in this study. First, the stromal and immune score calculated by Estimate algorithm was used to examine the distribution of stromal and immune content among four subtypes in TCGA data set (Supplementary Figures S3B and S4A).²⁹ We found that HIS had the highest stromal and immune scores among the four subtypes (Supplementary Figure S4A,B). On the contrary, we observed a significant reduction in tumour purity of samples in HIS subtype (Supplementary Figure S4C), suggesting that tumours in HIS subtype contained larger number of immune cells (Supplementary Figure S4C). We also validated these findings in CGGA data sets and similar results were obtained (Supplementary Figure S4D–I). Besides, a tumour microenvironment (TME) cell network was constructed to depict an overview of immune cell interactions and their effects on the clinical outcome of patients with gliomas (Figure 3A). The NK cells activation and the enrichment of plasma cells, naïve CD4+ T cell and B cell may be the main reasons for the better prognosis of patients in AFM subtype (Figure 3A,B). For patients in HIS subtype, the M2 macrophage, neutrophils, activated DCs and activated memory CD4+ T cells may contribute to the poor prognosis (Figure 3A,B). CSS showed significant increases in the infiltration of monocytes, M1 macrophages, and activated mast cells, while GLU exhibited high infiltration of M0 macrophages, CD8+ T cells and follicular helper T cells (Figure 3A,B). To verify this, paraffin-embedded glioma tissues from CGGA database were collected for IHC analysis, and the results showed that CD4 and CD163 were overexpressed in glioma tissues of HIS subtype (Figure 3C). These findings suggested that there were significant differences in TME cell infiltration in the four subtypes of gliomas.

To further investigate the immune background characterizing each TME-related subtype, we also analysed the expression level of selected cytokine, chemokine, and immune checkpoint in glioma samples from TCGA data set (Figure 3D,E). In this part, we considered GZMB, GZMA, CXCL9, CXCL10, CD8A, PRF1, TNF and TBX2 to be immune-activated biomarkers. As shown in Figure 3C, HIS subtype was associated with high expression of T helper 1 (Th1)/cytotoxic T lymphocyte (CTL)-related genes (Especially CXCL9, CXCL10, GZMA, and GZMB)³⁷ (Figure 3D). These results suggested that the HIS has the highest level of immunocyte infiltration. Additionally, we further investigated the expression of several immune checkpoint genes, since T cell and nature killer (NK) cell exhaustion has been identified as an important mechanism of cancer progression and immune evasion.³⁸ We also found that the HIS subtype had the highest expression of several immune checkpoints (particularly IDO1, CD274 and PDCD1LG2), it means that glioma samples from HIS subtype tend to be immunosuppressed (Figure 3E). Besides, the key members of human leukocyte antigen (HLA) complex, which are required to present endogenous cellular antigens to circulating T cells were also included in this study (Figure 3F). The results revealed that CSS and HIS subtypes were associated with high expression of HLA complex member mRNAs. Moreover, the expression level of TGF-β/EMT pathway related genes (such as VIM, ACTA2, COL4A1 and TWIST1) were relatively higher in HIS and GLU subtypes (Figure 3G). In brief, the results mentioned above identified an extensive immune heterogeneity among four subtypes of gliomas.

3.4 The clusters are associated with hypoxia and metabolic status in gliomas

According to the previous researches, the hypoxia and metabolic status play an essential role in the malignant progression of human cancers.^{7, 39} The activation of anti-tumour immune response is also affected by the oxygen content and cellular metabolism in tumour microenvironment. Thus, we further investigated the differences of hypoxia response and metabolic status among the four subtypes. In this study, CA9, FOSL1, SLC2A1, SLC16A1, C20orf20 and VEGFA were considering as biomarkers of hypoxia response according to the related literature.⁴⁰ After exploring the expression level of these hypoxia response-related biomarkers in TCGA data set, we found that the glioma samples from HIS and GLU subtypes were in a more obvious hypoxia state (Figure 4A). Besides, IHC analysis was performed with paraffin-embedded glioma tissues from CGGA database, and similar results were obtained (Figure 4B). Besides, we also found that the glioma samples from four clusters have significant differences in cellular metabolic pathways (Figure 4C, Supplementary Figure S5). Notably, in HIS and GLU subtypes with poor prognosis, the metabolism of nucleotides, glucose and vitamins was more active (Figure 4C, Supplementary Figure S5). In contrast, glioma samples with active amino acid and fatty acid metabolism tend to be enriched in AFM and CSS subtypes (Figure 4C, Supplementary Figure S5). The differences in metabolic pathways among the four clusters were also verified by IHC (Figure 4). These differences in hypoxia response and cellular metabolism may be the molecular mechanism behind the heterogeneity of tumour immunologic microenvironment, which also has a significant impact on the prognosis of patients with gliomas.

3.5 Validation of pathway-based subtypes on single cell resolution

To evaluate the strength of pathway-based classification, we assessed the intersection of the four clusters with the existing and widely used multiple cellular states in single cell resolution. The four pathway-based clusters were closely correlated with cellular states (Figure 5A). We found increased fraction of HIS subtype GBM cells in MESlike1 and MESlike2 subtypes, compared with other four subtype, suggesting the mesenchymal identity and hypoxia/ immunosuppression activities were inseparable features in GBM. Meanwhile, the AFM subtype GBM cells were mostly excluded from the MESlike1 and MESlike2 subtypes (Figure 5B).

The glioma heterogeneity was driven by tumour genetic and microenvironment. To comprehensively depict the interactions between tumour microenvironment and four pathway-based clusters, we analysed the communication between tumour cell and other cells. The significant ligand-receptor (L-R) interactions were identified by the CellPhoneDB software. Hundreds of specific molecular pairs mediating cell–cell interactions pairs were found between different cell types, and communication between tumour cell and immune cell was dominated (Figure 5C). We found 15 L-R pairs specific interacted between HIS subtype cells and immune cells, including C3_C3AR1, TIMP1_FGFR2, C3_aMb2 complex, CD44_FGFR2, SPP1_a9b1 complex, NRP1_VEGFB, COL6A1_a2b1 complex, FN1_a2b1 complex, VEGFA_FLT1, FN1_a5b1 complex, TNFSF10_RIPK1, COL8A1_a1b1 complex, FGF1_FGFR2, FGFR1_FGFR2 and LTBR_LTB (Figure 5D). Meanwhile, the gene ontology analysis indicated that the genes in 15 L_R pairs mainly involved in cell proliferation and angiogenesis progress. Collectively, these results indicate that the interactions between HIS type tumour cells and immune cells can alter neoplastic glioma cells towards a more proliferative phenotype that associates with poor prognosis (Figure 5E).

3.6 Identification of a HIS-related prognostic signature for gliomas

According to the results mentioned above, glioma samples in HIS subtype had the characteristics of immunosuppression, oxygen deprivation and active energy metabolism, resulting in a poor clinical outcome. Therefore, to provide more reasonable and effective treatment strategies, it is necessary to identify patients in HIS subtype accurately. Here, we constructed a HIS-related signature based on the differential genes among four clusters in TCGA data set for patient screening and prognosis prediction. The differentially expressed genes were identified by ‘limma’ package in R-3.6.2, wherein 177 genes were significantly overexpressed in HIS subtype (Fold change >1, p < 0.05, Figure 6A). Then, LASSO Cox regression model was performed for to select genes with best prognostic value (Figure 6B, Supplementary Figure S6A). Subsequently, we constructed a five-gene signature and the risk score of each sample was calculated using the formula mentioned in ‘Materials and Methods’ section (Figure 6C, Supplementary Figure S6B,C). After exploring the distribution of risk score among four clusters TCGA and CGGA data sets, we found that HIS subtype had the highest risk score (Figure 6D, Supplementary Figure S6D,F). Afterwards, ROC curve analysis was performed to assessed the predictive accuracy by computing area under the curve (AUC) of risk score. The AUC of risk score (91.4% in TCGA data set, 90.2% in CGGA data set1 and 91.9 in CGGA data set2) suggested that the risk signature could accurately predict glioma samples from HIS subtype, which has important reference value for individualized treatment of gliomas.

To evaluated the prognostic value of the risk signature, we performed timeROC analysis by ‘timereg’ package in TCGA and CGGA data sets. The results suggested the superior performance of this signature for prediction of 1-, 3- and 5-years survival rate (Figure 6E, Supplementary Figure S6E,G). Besides, samples from TCGA and CGGA data sets were divided into low- or high-risk group according to the median value of risk scores. Kaplan–Meier (K–M) survival analysis revealed that patients in low-risk group had a better prognosis than those in high-risk group (Figure 6F, Supplementary Figure S7A,B). Considering the significant difference of pathological characteristics between lower-grade glioma (LGG, WHO Grade 2 and 3) and GBM, we also performed K-M survival analysis in LGG and GBM, respectively. Consistently, the survival curve showed that both LGG and GBM patients in low-risk group tended to have a better clinical outcome. (Figure 6G,H, Supplementary Figure S7C–F). Furthermore, univariate and multivariate Cox regression analyses was performed to evaluated the prognostic independence of this risk signature. As expected, the signature was a prognostic factor independent of other clinicopathological features in gliomas in TCGA and CGGA data sets (Table 1, Supplementary Tables S3 and S4). These results demonstrated that the HIS-related signature had great value of clinical practice for patient screening and prognosis prediction.

3.7 Functional annotation and clinical application of the risk signature

In both TCGA and CGGA data sets, PCA showed that patients in the high-risk and low-risk group tended to have differential gene expression profiles (Figures 7A–C). Thus, we hypothesize that samples in the low-risk and high-risk group have different biological features. To verify this, Pearson's correlation analysis was performed to select genes that positively correlated with the risk score (R > 0.5) in TCGA and CGGA data sets. Then, these genes were uploaded to DAVID website for functional annotation. The GO analysis revealed that the risk score was positively correlated with immune and inflammatory response, response to hypoxia, glycolysis and protein metabolic processes and negatively correlated with apoptotic process (Figure 7D,E, Supplementary Figure S8A–C). Besides, we also found that the extracellular matrix organization, angiogenesis, cell proliferation and migration were significantly enriched in the high-risk group, suggesting the risk signature was tightly associated with the malignant phenotype of gliomas (Figure 7D,E, Supplementary Figure S8A–C). To improve accuracy, GSEA was performed in TCGA data set as a validation and similar results were uncovered (Figure 7F). These results indicated that the risk signature derived from the new classification based on the immunologic, hypoxia and metabolic status could represent the malignant progress of gliomas.

To evaluate the clinical application of the risk signature, we focussed on the effect of the risk score on the efficacy of postoperative adjuvant therapy. As mentioned before, we divided patients into high-risk and low-risk groups according to the risk score. The K-M survival analysis revealed that compared with other treatment strategies, postoperative radiotherapy combined with chemotherapy can significantly improve the prognosis of patients in high-risk group (Figure 7G, left). However, for patients in low-risk group, postoperative adjuvant therapy could not significantly improve the clinical outcome (Figure 7G, right). These results implied that patients in high-risk group can benefit significantly from postoperative adjuvant therapy, and more aggressive treatment strategies should be taken. Moreover, we also found that patients in high-risk group and HIS subtypes have a higher probability of responding to anti-PD1 immunotherapy (Figure 7H and Supplementary Figure S8D, p = 0.008 and p = 0.0443, respectively). To find an optimal immunotherapy strategy for patients with gliomas. GSE121810 data set was enrolled in this study.²² The K-M survival analysis suggested that neoadjuvant administration of PD-1 blockade could significantly prolong the OS of patients in high-risk group but not the low-risk group (Figure 7I). Thus, the risk signature might serve as an important reference in individualized treatments for glioma patients.