2.1 Data sets and data collection

GeneCards (https://www.genecards.org/) is a searchable, integrative database that provides comprehensive, user-friendly information on all annotated and predicted human genes. ER stress-related genes were extracted from GeneCards, and genes with a relevance score ≥7 were selected. TCGA RNA-seq transcriptome data and clinical information were obtained from TCGA database (http://cancergemome.nih.gov/). The Chinese Glioma Genome Atlas (CGGA) mRNA expression data (mRNAseq_325 and mRNA-array) and corresponding clinicopathological features were collected from the CGGA database (http://www.cgga.org.cn). The GSE16011 data set was procured from the Gene Expression Omnibus (GEO) database (http://www.ncbi.nlm.nih.gov/geo/). Raw data were processed via the R package, affy, and robust multi-array analysis (RMA) was used for background correction and normalization. Biospecimens and clinical data from related research were used as supplements.²³ Patient characteristics, including detailed data for each research object, are shown in Table S1.

2.2 Samples and quantitative real-time PCR analysis

From June 2018 to September 2019, we gathered 12 glioma tissues (4 cases each of grade II, III and IV glioma) from patients at the First Hospital of China Medical University. The clinical characteristics of 12 patients were shown in Table S2. Total RNA extraction and quantitative real-time PCR (qRT-PCR) was performed as previously described.²⁴ After extraction using TRIzol reagent (Invitrogen/Thermo Fisher Scientific), total RNA was transcribed to first-strand cDNA and then underwent qRT-PCR (SYBR Green Master Mix). Each sample was assayed in triplicate. After being normalized to GAPDH expression levels, the expression values were log₂ transformed. Primer sequences of target genes are shown in Table S3. This research was approved by the Medical Ethics Committee of the First Affiliated Hospital of China Medical University. All participants provided written informed consent.

2.3 Western blot assay

Western blot assay was performed as previously described.²⁴ Briefly, the total protein of glioma tissues was extracted using a protein extraction kit (Beyotime). Equivalent amounts of protein (25-50 μg) were then electrophoresed and transferred to polyvinylidene fluoride membranes (0.45 μm; Millipore). After being blocked, membranes were incubated with antibodies against ATF6 (24169-1-AP, Proteintech), EIF2α (sc-133132, Santa), p-EIF2α (# 3398S, CST), p-IRE1α (sc-390960, Santa) and GAPDH (60004-1-Ig, Proteintech) at 4°C for 16 hours. The membranes were then incubated with appropriate secondary antibodies (ProteinTech). Protein bands of interest were detected and quantified using a chemiluminescence kit (Beyotime), the ChemiDoc™ Touch detection system (Bio-Rad Laboratories) and Image J software (National Institutes of Health).

2.4 ER stress-related risk signature construction and validation

We first conducted univariate Cox regression and Kaplan-Meier (KM) analysis using the survival R package to identify ER stress-related genes associated with patients’ overall survival (OS) time from TCGA, CGGA (CGGA refers to mRNAseq_325 data unless otherwise specified), CGGA (mRNA-array) and GSE16011 data sets. Only when the p values of both analysis methods were ≤0.05 were the genes included in the next step. The intersection genes related to OS from the above four data sets were analysed by least absolute shrinkage and selection operator (LASSO) regression, using the glmnet R package in TCGA database, to narrow the range of prognosis-related genes. Subsequently, the Akaike information criterion (AIC) method of multivariate Cox regression analysis was performed using the survival package to establish an optimal ER stress-related risk signature based on linear integration of the regression coefficient obtained from the multivariate Cox regression analysis and expression level of the selected ER related genes. The risk score was computed as follows:

where Exp_i is the expression value of the ER stress-related genes and Coef_i is the corresponding regression coefficient calculated by multivariate Cox regression analysis. TCGA data were used as the training cohort, and CGGA and GSE16011 data were used for the validation cohorts.

2.5 Survival analysis

Kaplan-Meier survival analyses were performed using the survival and survminer packages in R to compare the OS between different groups of glioma patients. The R package survivalROC was used to establish time-dependent receiver operating characteristic (ROC) curves to check the accuracy of the risk signatures in predicting the outcomes of glioma patients. The larger the area under the ROC curve (AUC), the stronger the predictive ability of the risk model. A risk plot was drawn using Pheatmap R package to display the distribution of survival status of samples in different risk groups.

2.6 Functional enrichment analysis

The GSVA package in R was applied to estimate the Gene Ontology (GO) biological processes and Kyoto Encyclopedia of Genes and Genomes (KEGG) pathways that correlated with the risk signature. The GSVA package scored the GO biological processes and KEGG pathways in each sample, and then by comparing the score differences in different risk groups, we identified the different biological processes enriched in the high- and low-risk groups. The limma package in R was adopted to identify differentially expressed genes and gene sets in different groups. To further verify the GO processes and KEGG pathways related to the signature, GO and KEGG analyses were performed for the differentially expressed genes using the clusterProfiler package in R. All heat maps were drawn using the Pheatmap package in R.

2.7 Evaluation of immune cell fractions and immune subtypes

Myeloid-derived suppressor cells (MDSCs),²⁵ regulatory T cells (Tregs),²⁶ natural killer (NK) cells²⁷ and neutrophils²⁸ in the glioma microenvironment are considered to be involved in immunosuppression. The ssGSEA method of the GSVA package was used to identify the immune cell composition mentioned above to evaluate the level of immune cell enrichment in the tumour microenvironment through the gene expression level in a single tumour sample.²⁹ Vésteinn Thorsson and his colleagues categorized tumours into six immune subtypes, C1 (Wound Healing) has high angiogenic gene expression, high proliferation rate, and Th2 cells trending to the adaptive immune infiltrate; C2 (IFN-γ Dominant) has the highest M1/M2 macrophage polarization level, a strong CD8 signal, and together with C6, the greatest T cell receptor diversity; C3 (Inflammatory) is determined by increased expression of Th17 and Th1 genes, low to moderate tumour cell proliferation, and, along with C5, lower aneuploidy levels and overall somatic copy number alterations compared with the other subtypes; C4 (Lymphocyte Depleted) displays a more prominent macrophage signature, with Th1 suppressed and a high M2 response; C5 (Immunologically Quiet) exhibits the lowest lymphocyte count, and highest macrophage responses, dominated by M2 macrophages; C6 (TGF-β Dominant) displays the highest TGF-β signature and a high lymphocytic infiltrate with an even distribution of Type I and Type II T cells. Subtype C4 is enriched in adrenocortical carcinoma (ACC), pheochromocytoma and paraganglioma (PCPG), hepatocellular carcinoma (LIHC) and gliomas, while low-grade gliomas (LGG) consist mostly of subtype C5.³⁰ The glioma samples were classified into different immune subtypes using the ImmuneSubtypeClassifier package in R.³⁰

2.8 Independent prognostic role of the risk signature

To identify whether the ER stress-related risk signature depended on other clinicopathological factors (including age, gender, tumour grade, IDH mutation status, 1p19q codeletion status and MGMT promoter methylation status) in predicting patients’ OS, univariate and multivariate Cox regression analyses were performed using the survival R package. The results of independent prognostic factor analysis were displayed in the form of forest plots using the forestplot R package.

2.9 Development and assessment of the nomogram

The nomogram shows the probability of clinical events through simple graphs of statistical prediction models to form a personalized prediction model. Age, gender, tumour grade, IDH mutation status, 1p19q codeletion status, MGMT promoter methylation status and ER stress-related risk scores were combined to develop a nomogram using the R packages survival and rms. Calibration curves and decision curve analysis (DCA) were adopted to evaluate the accuracy of the nomogram in predicting one-, three- and five-year survival rates of glioma patients.^{31, 32} For the calibration curve, the higher was the degree of agreement between the predicted curve and the actual curve, the stronger was the predictive ability of the nomogram. For DCA, the decision curve of the nomogram was compared with that of other independent prognostic factors.

2.10 Statistical analyses

R software (version 3.6.3) and GraphPad Prism v7.00 (GraphPad Software Inc.) were used as statistical analysis tools in this study. Quantitative data are presented as the mean ± standard error of the mean (SEM) or standard deviation (SD). The Wilcoxon test was applied to compare the statistical differences between the two groups, and the Kruskal-Wallis H test was employed to compare multiple groups. Statistical significance was defined as P < .05. Venn diagrams were drawn using jvenn.³³ The other plots were constructed using R software or GraphPad Prism.

3.1 Sixteen ER stress-related genes were identified to develop risk model

A total of 787 ER stress-related genes with a relevance score of ≥7 were extracted from the GeneCards database to generate prognostic gene signatures (Table S4). These genes were subjected to univariate Cox regression and KM analyses. In the TCGA, CGGA, CGGA (mRNA-array) and GSE16011 data sets, 593, 504, 422 and 330 genes, respectively, were significantly related to OS of glioma patients (Figure 1A, Tables S5-8). The overlapping genes (190 genes, Table S9) were included in the LASSO regression analysis to avoid overfitting problems in the risk signature (Figure 1B,C, Table S10). The AIC method of multivariate Cox regression analysis was applied to the genes returned from the LASSO regression analysis (28 genes) to construct the optimal model, which included sixteen genes (CYP2E1, SLN, BRCA1, CISD2, LRRK2, BMP2, MYH7, HSPB1, DNM1L, SHISA5, RNF185, RCN1, SPP1, RPN2, PDIA3 and ATP2A2) (Figure 1D). Among these genes, CYP2E1, SLN, BMP2, MYH7, RNF185 and PDIA3 were protective factors for glioma survival, with hazard ratios (HRs) <1, and BRCA1, CISD2, LRRK2, HSPB1, DNM1L, SHISA5, RCN1, SPP1, RPN2 and ATP2A2 were risk factors with HRs > 1.

Furthermore, we analysed the correlation of these 16 genes in the TCGA, CGGA and GSE16011 data sets and found that few genes were highly correlated with each other in the three data sets (Figure S1A-C). We retrieved BRCA1, PDIA3, DNM1L, RCN1, HSPB1, RPN2, LRRK2 and SPP1 protein expression levels in gliomas using the Human Protein Atlas (HPA) (https://www.proteinatlas.org/). Expression levels of risk factors BRCA1, DNM1L, RCN1, HSPB1, RPN2, LRRK2 and SPP1 in high-grade gliomas were greater than those in lower-grade gliomas, while the protein expression levels of protective factor PDIA3 were exactly the opposite (Figure S2). We further examined the genetic alterations of these OS-associated genes in glioma. A data set (http://www.cbioportal.org) comprising a merged cohort of low-grade glioma (LGG) and GBM containing 794 patients/samples harbouring both mutations and CAN data was queried. Genetic alterations were found in 108 (13.60%) of the queried patients/samples (Figure S1D). CYP2E1 had the highest frequency of genetic alterations.

3.2 Establishing and evaluating the ER stress-related risk signature

The Cancer Genome Atlas glioma data were used to construct the risk signature. The risk score was calculated as follows: risk score = (−0.1318 × CYP2E1 expression) + (−0.1122 × SLN expression) + (0.4543 × BRCA1 expression) + (0.3481 × CISD2 expression) + (0.1811 × LRRK2 expression) + (−0.1236 × BMP2 expression) + (−0.1183 × MYH7 expression) + (0.4296 × HSPB1 expression) + (0.3839 × DNM1L expression) + (0.5514 × SHISA5 expression) + (−0.5315 × RNF185 expression) + (0.2889 × RCN1 expression) + (0.1260 × SPP1 expression) + (0.5187 × RPN2 expression) + (−0.7779 × PDIA3 expression) + (0.2649 × ATP2A2 expression). Time-dependent ROC curves were used to assess the efficiencies of the prognostic prediction of the ER stress-related risk signature. As presented in Figure 2A, the AUCs for 1-, 3- and 5-year OS in the TCGA data set were 0.890, 0.929 and 0.900, respectively. The AUCs for predicting 1-, 3- and 5-year OS in the CGGA data set were 0.719, 0.795 and 0.804, respectively, and those in the GSE16011 data set were 0.758, 0.814 and 0.808, respectively (Figure 2B,C).

Samples in the TCGA, CGGA and GSE16011 cohorts were then divided into low- and high-risk groups based on the median risk score in each cohort. KM analysis showed that patients in the low-risk group had a more favourable outcome than patients in the high-risk group (Figure 2D-F). The three-year OS rates were 26.56% (95% confidence interval (CI): 21.20-33.28) vs 92.43% (95% CI: 88.72-96.32) for the high- and low-risk groups in the TCGA data set, 22.38% (95% CI: 16.56-30.25) vs 69.4% (95% CI: 62.40-77.21) for the high- and low-risk groups in the CGGA cohort, and 10.81% (95% CI: 6.60-17.72) vs 47.88% (95% CI: 39.97-57.37) for the high- and low- risk groups in the GSE16011 data set, respectively. Figure 2G-H and Figure S3A show the OS-related prediction model distribution of patients in the TCGA, CGGA and GSE16011 data sets. These results indicate the accuracy of the ER stress-related risk signature in predicting the outcomes of glioma patients. The expression patterns of the 16 genes in the TCGA, CGGA and GSE16011 data sets are shown in Figure S3B-D.

3.3 The ER stress-related signature is associated with clinicopathological features

We investigated whether the ER stress-related risk signature was correlated with the clinicopathological features of glioma patients. First, we examined the prognostic effect of the signature in LGG and GBM patients using KM analysis. As displayed in Figure 3A-F, the high-risk group showed reduced OS in terms of both LGG and GBM in the TCGA, CGGA and GSE16011 data sets. Additionally, the risk scores were notably different among stratified patients, with high-risk scores in high-grade glioma, wild-type IDH, 1p19q non-codeletion, mesenchymal subtype and MGMT promoter unmethylated patients (Figure 3G-N, Figure S4A-D). These results suggest that the ER stress-related risk signature can accurately distinguish different clinicopathologic features of glioma patients.

3.4 Functional annotation of the risk signature

Many studies on ER stress have used the intracellular expression levels of related proteins, such as ATF6, HSPA5, XBP1 and ATF4, as indicators to examine the intensity of ER stress in cells or tissues.^34-36 To explore ER stress status in different glioma groups, we measured the expression levels of these markers in the TCGA, CGGA and GSE16011 data sets. Most of them had higher expression levels in the high-risk group, demonstrating that the ER stress in this group was markedly more intense than that in the low-risk group (Figure 4A-C). To further verify the relationship between the risk score and the ER stress intensity of gliomas, we used qRT-PCR to measure the expression levels of the 16 risk genes in the 12 glioma samples. The samples were scored based on the expression levels of these genes and grouped according to the score. We then used Western blotting to measure the expression levels of ATF6, EIF2α, p-EIF2α and p-IRE1α in the samples from the high- and low-score groups. The Western blotting results were consistent with those of the above databases analysis (Table S11, Figure S5). These results further confirmed the correlation between the signature and ER stress activation. Next, we used the GSVA package to explore the KEGG pathways and GO biological processes associated with risk signature in the TCGA and CGGA data sets. At P < .05, the top 20 KEGG pathways and the top 30 GO biological processes were identified based on the logFC value in the TCGA and CGGA data sets. Most of the KEGG pathways and GO biological processes enriched in the high-risk group were associated with the immune and inflammatory responses as well as cell biosynthesis and degradation (Figure 4D,E, Figure S6A,B). To confirm these results, differentially expressed genes between the two groups were selected and underwent GO and KEGG analyses using the R package, clusterProfiler, in the TCGA, CGGA and GSE16011 data sets (Figure S7A-F). The results were similar to the results of the GSVA analysis.

The typical characteristic of ER stress is cellular protein synthesis in excess of the ER folding ability. This may be why biosynthesis- and degradation-related pathways were enriched in the high-risk groups. In addition, studies have confirmed that ER stress inhibits anti-tumour immunity and promotes tumour cell escape from immunosurveillance.^{13, 37, 38} Here, the enrichment in immune and inflammatory response-related pathways and processes in the high-risk group further demonstrates that the immune environment of the high-risk group is complex.

3.5 High ER stress-related risk score demonstrates an immunosuppressive feature

The process of tumour eradication by the immune system involves the cancer-immunity cycle.³⁹ We explored the expression characteristics of genes that inhibited this cycle in the TCGA, CGGA and GSE16011 data sets. These genes were acquired from the Tracking Tumor Immunophenotype website (http://biocc.hrbmu.edu.cn/TIP/index.jsp).⁴⁰ In the three cohorts, most genes were highly expressed in the high-risk group (Figure 5A-C). TGFB1, VEGFA, ARG1, FGL2 and IL10 are secreted immunosuppressive factors in glioma,^41-46 while CD95L and CD70 are glioma cell-surface immunosuppressive factors.^{47, 48} As shown in Figure 5D-O and Figure S8A-I, they were all obviously overexpressed in the high-risk groups. Furthermore, some immune cells, such as MDSCs, Tregs, NK cells, and neutrophils, infiltrate into the tumour microenvironment and exhibit immunosuppressive effects to promote tumour occurrence, progression, and therapy resistance.^25-28 These immunosuppressive cells were also found to be enriched in the ER stress-related high-risk groups (Figure 5P-R). To further explore the characteristics of the immune microenvironment in glioma of the high- and low-risk groups, the ImmuneSubtypeClassifier package was used to divide the samples into different immune subtypes in the three cohorts. We found that in both the high- and the low-risk groups, the main subtypes were C4 and C5, but the high-risk group had considerably more C4 subtypes than the low-risk group (Figure 5S). The prognosis of the C4 immune subtype in tumours is worse than that of the C5 immune subtype,³⁰ which was consistent with the prognosis of high- and low-risk glioma patients. This further confirmed the accuracy of the ER stress-related risk signature in predicting immune subtypes and prognosis in gliomas.

In addition to the cancer-immunity cycle inhibitors mentioned above, immune checkpoints can also suppress the immune system's ability to clear tumours.⁴⁹ In recent years, immune checkpoints have become potential therapeutic targets for many malignant tumours and play an important role in tumour immunotherapy.⁵⁰ Comparison of the expression of immune checkpoints in the high- and low-risk groups showed that most immune checkpoints were up-regulated in the high-risk groups in the TCGA, CGGA and GSE16011 cohorts (Figure 6A-C). Considering the critical roles played by PD1 and PD-L1 in tumour immunosuppression and immunotherapy, we separately investigated the relationship between their expression levels and the ER stress-related risk score. We found that the expression levels of PD1 and PD-L1 were obviously positively correlated with the risk score (Figure 6D-G), and their expression levels in the high-risk group were markedly higher than those in the low-risk group (Figure 6H-K). These data suggest that the ER stress-related risk signature can accurately predict the immune characteristics of glioma.

3.6 Construction and validation of the nomogram

In addition to the ER stress risk score, there are numerous known prognostic factors for glioma, such as age, gender, WHO grade, IDH mutation status, 1p19q codeletion status and MGMT promoter methylation status. Therefore, it was necessary to examine whether the ER stress risk signature could independently predict prognosis. In the TCGA training cohort, univariate Cox analysis revealed that ER stress-related risk score was negatively correlated with the OS of glioma patients. Moreover, age, grade, IDH mutation status, 1p19q codeletion status and MGMT promoter status were also significantly related to OS (Figure 7A). Subsequent multivariate Cox regression analysis indicated that the ER stress-related risk signature, grade and IDH mutation status were significantly correlated with OS (Figure 7B). These results were further confirmed in the CGGA and GSE16011 data sets (Figure S9A-D). These findings indicate that the ER stress-related risk signature constructed using the TCGA data set was an independent prognostic factor for glioma patients.

We established a nomogram to predict 1-, 3- and 5-year OS in the TCGA data set, by integrating ER stress risk signature, age, gender, WHO grade, IDH mutation status, 1p19q codeletion status and MGMT promoter methylation status. In the nomogram, each signature was assigned points according to its risk contribution to OS (Figure 7C). The calibration curves illustrated a remarkable consensus between the predicted and actual survival time in terms of the 1-, 3- and 5-year OS rates in the TCGA and CGGA cohorts (Figure 7D,E). The DCAs constructed using the TCGA (Figure 7F-H) and CGGA (Figure 7I-K) cohorts for this nomogram demonstrated that the nomogram performed well in predicting the 1-, 3- and 5-year OS rates in glioma patients.