2.1 Patients and datasets

We collected 550 and 309 diffuse gliomas with RNA-seq data and clinical information from TCGA and CGGA database, respectively.11, 12 TCGA cohort was used as training set and CGGA cohort as validation set. All tissues and clinicopathologic information were obtained with written informed consents. This study was approved by ethics committee of Tiantan Hospital. The patient characteristics of these two cohorts were summarized in Table S1.

2.2 Gene set selection

Four lipid metabolism-related gene sets (Reactome metabolism of lipids and lipoproteins, Reactome phospholipid metabolism, Hallmark fatty acid metabolism and KEGG glycerophospholipid metabolism) were collected from the Molecular Signature Database v5.1 (MSigDB).13 After removing the overlapped genes, 614 lipid metabolism-related genes were obtained. The differential lipid metabolism genes between LGG and GBM were selected. By using the R package ‘survival’, univariate Cox analysis performed to prefilter the genes based on the P values. Then, the Cox proportional hazards model with elastic net penalty was applied for selecting signature gene, which was performed with the R package ‘glmnet’.14, 15 A linear combination of signature genes expression level weighted by regression coefficients (Coeffs) was developed to calculate the risk score of each patient in training set. Then, the regression Coeffs from training set was used to compute the risk scores for cases of validation set.

2.3 Bioinformatic analysis

Gene ontology (GO) analysis was performed for function annotation of differential genes.16 Gene set enrichment analysis (GSEA) was applied for identifying statistically different gene sets between two groups with GSEA v3 software.13 Principal components analysis (PCA) was carried out using the R package ‘princomp’ to analyse the expression pattern of grouped patients.17, 18 Utilizing the gene expression data, stromal and immune score of each sample was calculated with R package ‘ESTIMATE’ which reflected the gene signature enrichment of stromal and immune cells.19

2.4 Statistical analysis

Patients in both training and validation cohorts were assigned into high- or low-risk group based on the median value of risk score. Kaplan-Meier curves and 2-sided log-rank test were applied to assess the survival difference between high- and low-risk groups. Chi-square test was conducted to detect the pathologic differences between high- and low-risk patients. Univariate and multivariate Cox regression analyses were performed to assess the independent prognostic factors by using SPSS software. ROC curve analysis was used to predict overall survival (OS) with R package ‘pROC’. P value <.05 was considered significant statistically.

3.1 LGG and GBM show distinct lipid metabolism phenotypes

To detect the lipid metabolism differences during the progression of diffuse gliomas, we collected 550 patients with RNA sequencing data and clinical information from TCGA database and four lipid metabolism-related gene sets, which were integrated into one set containing 614 genes. Gene clustering using the R package ‘pheatmap’ found that the profile of lipid metabolism-related genes between LGG and GBM showed obvious differences (Figure 1A). Principal components analysis based on these selected genes showed that GBM and LGG were distributed in different regions, suggesting distinct lipid metabolism phenotypes between them (Figure 1B). To further explore the lipid metabolism phenotypes, we performed GO analysis and found that GBM mainly exhibited an enrichment of glycosphingolipid metabolic progress, whereas LGG displayed enrichment of phosphatidylinositol metabolic progress (Figure 1C). Gene set enrichment analysis analysis also confirmed this finding (Figure 1D,E). In addition, we also analysed the CGGA cohort of 309 glioma samples using the above methods, and the same results were observed between LGG and GBM (Figure S1). Heat maps showed the differential genes between LGG and GBM, involving in glycosphingolipid and phosphatidylinositol metabolic progress (Figure S2). These results indicated LGG and GBM displayed distinct lipid metabolic phenotypes.

3.2 Identification of a lipid metabolism-related gene set for prognostic prediction

Considering the distinct profile of lipid metabolism between LGG and GBM, we proposed to build a lipid metabolism-related gene set for predicting prognosis. By performing univariate Cox regression analysis, 297 prognosis-related genes remained (P < .05). Thirty-one out of prognosis-related genes were involved in glycosphingolipid and phosphatidylinositol metabolic progress (Figure 2A). Then, we performed the Cox proportional hazards model with elastic net regression for gene selection (Figure 2B). Consequently, a nine-gene signature was obtained as a classifier (Figure 2C,D), and risk score of each patient was computed with expression value and the coeffs of multivariable Cox regression.

Then, based on the median risk score, patients were assigned into high- or low-risk group. Kaplan-Meier analysis found the high-risk cases had a significantly shorter OS than low-risk ones (P < .001, Figure 2E). To validate this gene set, we also calculated patients' risk scores of CGGA cohort with same regression Coeffs. Heat map showed the expression of signature genes in CGGA cohort (Figure S3A). As expected, we acquired consensus result (Figure S3B).

3.3 The nine-gene set shows strong prognostic power

We next performed univariate and multivariate Cox regression analyses to determine the prognostic value of the acquired gene set. The results showed that the lipid metabolism-related gene set was independently correlative with OS (P = .017) (Table 1). Consistently, this gene set could also be served as an independent prognostic factor in CGGA validation set (P = .003) (Table 1). By computing the AUC of risk score, age and grade, we next assessed the predictive accuracy with ROC curve and found that AUC of risk score (0.86) was much higher than that of age (0.801) or grade (0.83) (Figure S4A). Similar results were also observed in CGGA validation set (Figure S4B). These results indicated that the acquired lipid metabolic gene set had strong power for prognosis prediction.

3.4 The acquired nine-gene set is correlated with pathologic features in diffuse gliomas

We further detected whether the gene set was associated with pathologic features. As shown in Figure 3, higher level of risk scores preferred to distribute in higher grade, classical, mesenchymal, IDH-wt, MGMT promoter unmethylated or 1p/19q non-codeleted patients. We also assessed the distributive differences of these pathologic features between high- and low-risk groups by performing chi-square test. In both cohorts, most of pathologic features had significantly different distribution between risk groups except gender (Table S1). These results suggested a significant association between the lipid metabolism gene set and clinical molecular features.

3.5 Application of the nine-gene panel in stratified patients

We further explored the prognostic significance of the gene panel in patients stratified by grade, IDH, MGMT promoter and 1p/19q status. In both cohorts, Kaplan-Meier analysis showed that cases with high-risk score had shorter overall survival than the low-risk ones in most stratified patients (Figure 4, Figure S5). The similar trend occurred in GBM or 1p/19q codeleted cases despite of no statistical difference (Figure 4B,G). After that, patients were also stratified by WHO 2016 molecular subtype. Consensus results were obtained in cases of IDH-mutant LGG, whereas in IDH-wt LGG, IDH-wt GBM and IDH-mutant GBM found no significant differences (Figure S6). These data revealed that acquired signature could accurately predict the unfavourable outcome in most stratified patients.

3.6 High-risk cases show enhanced cell division and immune response phenotypes

To detect the biological function differences, we further compared gene expression of patients between low- and high-risk groups. PCA found that low- and high-risk cases distributed in two regions clearly (Figure S7). Based on the differentially expressed genes (P < .05) which were identified by SAM, GO analysis found that cell division and immune response were significantly enriched in high-risk patients, whereas low-risk cases showed enrichments of chemical synaptic transmission and neurotransmitter secretion (Figure 5A,B). Moreover, GSEA also confirmed these findings (Figure 5C,D).