Development of A Vitamin-Related Gene Signature to Predict the Immune Characteristics and Prognosis of Glioma

Background: Vitamins not only play a pivotal role in maintaining homeostasis of the body, but also have complex impacts on the occurrence and progression of tumors. However, the effects of vitamins on glioma and the underlying mechanism have not been fully elucidated. Methods: Vitamin -related genes were extracted from the Molecular Signature Database v7.1 (MSigDB). The overlapping overall survival (OS)-related genes in The Cancer Genome Atlas (TCGA), the Chinese Glioma Genome Atlas (CGGA), and GSE16011 cohorts screened out by univariate COX regression analysis were utilized to construct a risk model based on the TCGA cohort via random survival forest analysis and multivariate COX regression analysis. The powerful prognostic predictive potential of the vitamin-related risk signature was veried by Kaplan–Meier survival analysis and receiver operating characteristic (ROC) analysis in the three datasets. The ssGSEA method of the GSVA package was used for functional enrichment and immune cell component analyses. ESTIMATE score analysis was used for auxiliary analysis of glioma immune characteristics. A nomogram was constructed and assessed based on the TCGA dataset. Results: The vitamin-related six-gene (POSTN, IRX5, EEF2, RAB27A, MDM2, and ENO1) risk signature constructed based on the TCGA dataset accurately predicted the outcomes of glioma patients and credibly distinguished between different levels and molecular subtypes of glioma in the TCGA, CGGA, and GSE16011 cohorts. Gliomas with high risk scores exhibited high immune scores, low tumor purity, and immunosuppressive features. The nomogram constructed by combining the vitamin-related risk signature and clinicopathological factors precisely predicted the 1-, 3-, and 5-year OS of glioma patients. Conclusions: Our study revealed that the vitamin-related six-gene risk signature, as an independent prognostic factor, could accurately distinguish the grade, molecular subtype, and immune characteristics of glioma.


Background
Glioma, including low-grade glioma (LGG) and glioblastoma (GBM), is the most prevalent primary malignancy of the brain, representing 81% of intracranial malignancies (1). Glioma is classi ed to grade I − IV by the World Health Organization (WHO) in 2016 in the light of their aggressiveness (2). Despite surgical resection combined with postoperative radiotherapy and chemotherapy, OS improvements in glioma patients remain unfavorable (3). The unfavorable outcome of glioma is due to its strong invasion ability and the lack of effective therapeutic targets (4). Therefore, the discovery of e cient therapeutic targets and the development of new therapeutic drugs to improve the OS of glioma patients are critical.
Vitamins are organic compounds required in small amounts in cellular metabolism and are important for overall health and normal growth of the organism. Since vitamins cannot be generated in vivo and must be ingested through diet, de ciencies in these substances are known to cause clinical disease states. In addition to the well-known lack of vitamin D that can cause rickets, it has been shown that a lack of folic acid, vitamin E, and vitamin B 3 (niacin) can increase the risk of coronary heart disease (5); lack of vitamin K can cause clotting dysfunction or hemorrhage (6). Vitamins also have an important impact on the occurrence and development of tumors. Vitamin C acts as a cofactor for many enzymes to regulate antioxidant defense, transcription and epigenetics in tumors (7). Some B vitamins, including folate, ribo avin, vitamin B 6 , vitamin B 12 , and choline, affect the transformation and progression of tumors by modulating their one-carbon metabolism (8,9). Vitamin D can inhibit tumorigenesis and prolong the survival of tumor patients (10). Vitamins also have a regulatory effect on tumor immunity (11,12). Highdose vitamin C regulates the in ltration of immune cells into the tumor microenvironment and suppresses cancer development in a T cell-dependent manner (13). Vitamin A can increase cytotoxic activity of macrophages in melanoma (14).
In glioma, vitamin A, vitamin C, and vitamin E have been proven to reduce the risk of glioma (15)(16)(17); vitamin D can not only attenuate the expression of stem cell markers in glioma stem cells, but also inhibit the self-renewal potential and mitochondrial respiration of glioma stem cells (18).
However, the role of vitamins in tumors and the underlying mechanisms are still unclear. Some studies have con rmed that vitamin-related receptors play an important role in signal transduction. For instance, the nuclear vitamin D receptor binds to vitamin D through genomic pathways and further forms a heterodimeric complex by combining with a retinoid-X-receptor to regulate the transcription of vitamin D downstream genes such as CDKN1A, C-MYC, CDH1, and CYP24A1 (19). Other studies have found that vitamins play a role in tumors through their dependent factors. A typical example is the protein S, a vitamin K-dependent factor, which can not only promote the migration of brain neural stem cells to glioma cells, but also enhance the phagocytic ability of brain neural stem cells on apoptotic glioma cells (20). Although these studies do not fully explain the mechanism of vitamins in tumors, they at least show that the expression of vitamin-related genes is crucial in tumors and these genes may serve as potential therapeutic targets.
In our study, we comprehensively investigated the vitamin-related genes in glioma. We collected vitaminrelated genes and identi ed six genes related to patients' prognosis. The vitamin-related six-gene risk signature was constructed and acted as an independent prognosis factor for glioma patients that can distinguish the immune characteristics in glioma microenvironment. Additionally, we established a nomogram that integrates the vitamin-related risk scores with clinicopathological features and assessed its capability in estimating 1-, 3-, and 5-year OS rates of glioma patients.

Methods
Human vitamin-related gene set Vitamin-related genes were extracted from MSigDB (21). All the vitamin-related gene sets are shown in Table S1. After removing the overlapped genes, 622 vitamin-related genes were obtained (Table S2).

Patients and datasets
The RNA-sequencing transcriptome data and relevant clinicopathological data of 558 glioma patients were obtained from TCGA (http://cancergenome.nih.gov/) as a training set. Similarly, data of 416 glioma patients from CGGA (www.cgga.org.cn) were downloaded as a validation set. GSE16011 gene expression and clinicopathological data were obtained through the Gene Expression Omnibus (GEO) database (http://www.ncbi.nlm.nih.gov/geo/) as another validation set. Raw data of GES16011 were processed utilizing the affy R package, and robust multi-array analysis (RMA) was used for background correction and normalization. Patient characteristics are shown in Table S3.

Construction of a vitamin-related risk model
Univariate and multivariate Cox regression analyses, and the random survival forests-variable hunting (RSFVH) algorithm were adopted to examine the prognostic signi cances of vitamins-related genes in estimating glioma patients' OS. Genes were considered as potential prognostic factors at p < 0.05. RSFVH algorithm and multivariate Cox regression analyses with the Akaike information criterion (AIC) method were performed to further screen and narrow of the prognostic related genes. Hazard ratios (HRs) and regression coe cients were counted for each gene, and six reasonable genes were eventually ltered out. Afterwards, the vitamin-related risk model was constructed via a linear integration of the regression coe cient extracted from the multivariate Cox regression analysis and expression level of the six genes. The formula for calculating the risk scores was as follows: Risk score (patient) = Coe cient mRNA1 × Expression mRNA1 + Coe cient mRNA2 × Expression mRNA2 + Coe cient mRNA3 × Expression mRNA3 + ⋯ + Coe cient mRNAn × Expression mRNAn .
Assessing of the risk signature All samples in the training and validation cohorts were divided into low-and high-risk groups according to the median value of risk scores. Kaplan-Meier (KM) survival curves of the two groups and the timedependent receiver operating characteristic (ROC) curves for OS evaluation were implemented to evaluating the accuracy of the vitamin-related risk model.

Expression and prognostic characteristics of the six genes
The correlation analysis between the six genes was performed using the corrplot package in R. The expression level of the six genes in normal brain and glioma tissues was detected via Gene Expression Pro ling Interactive Analysis (GEPIA) (22). The Human Protein Atlas (HPA) (http://www.proteinatlas.org) was used to analyze the expression of proteins encoded by the six genes in glioma. The genetic alteration features of the six genes were extracted from the cBioPortal (23).

Functional enrichment analysis
The ssGSEA method of GSVA package in R was employed to calculating the GSVA scores of each gene set from the Gene Ontology (GO) biological processes and Kyoto Encyclopedia of Genes and Genomes (KEGG) pathways for each sample. The limma package in R was applied to determine differentially expressed genes and differential gene sets scores in the low-and high-risk groups. GO and KEGG analyses were conducted with the differentially expressed genes between low-and high-risk groups via the clusterPro ler package in R. GSEA analysis was performed via the GSEA v4.0.3 desktop program (http://software.broadinstitute.org/gsea/datasets.jsp) to further verify the GO analysis results.

Estimation of the glioma immune microenvironment
The ESTIMATE R package was employed to calculate the ESTIMATE score of each sample (24). The ssGSEA method of the GSVA package was used to calculate the in ltration levels of different immune cells in glioma microenvironment (25)(26)(27). The immuneSubtypeClassier package was adopted to identify different immune subtypes of glioma samples (28).
Independence of the risk signature in predicting prognosis Univariate and multivariate Cox regression analyses were adopted to assess the independent prognostic prediction potential of the vitamin-related risk model for glioma patients. To better evaluation the prognostic signi cance of the risk signature and clinicopathological features, the ggplot2 package in R was adopted to draw the forest plots.

Establishment and evaluation of the nomogram
To improve the quality with a quantitative tool, we developed a nomogram using the rms package to predict the 1-, 3-, and 5-year OS of glioma patients. Calibration plots were employed to assess the accuracy of the nomogram. Decision curve analysis (DCA) has been recommended to evaluate the latent clinical effectiveness of predictive nomograms. In this study, DCA was employed to assess the clinical utility of the nomogram by calculating the net bene ts for a range of threshold possibilities. DCA was performed as a statistical method that integrated the results of a decision to determine the clinical utility of a model. Therefore, the DCA can examine the net bene t to evaluate the clinical utility of a nomogram (29,30). A net reduction curve was used as a supplementary veri cation for DCA (31).

Statistical analysis
Statistical analysis of this exploratory research was implemented utilizing R software (v3.6.3) or GraphPad Prism v7.00 (GraphPad Software Inc., USA). Quantitative data are presented as the mean ± standard deviation (SD). The Wilcoxon test was employed to weigh the statistical differences between two groups; the Kruskal − Wallis H test was adopted to compare multiple groups. p < 0.05 was believed statistically signi cance. The Venn diagram was plotted by jvenn (32).

Results
Six vitamin-related genes were screened out to develop a risk signature To identify vitamin-related genes that were signi cantly associated with glioma OS, we rst performed univariate COX regression analysis based on the TCGA, CGGA, and GSE16011 datasets. There were 482, 354, and 290 vitamin-related genes, respectively, in the TCGA, CGGA, and GSE16011 datasets that were obviously linked with OS of glioma patients (Fig. 1A, Tables S4 − 6). Eleven vitamins-related genes closest linked with the prognosis of glioma patients were selected out from the 193 overlapped genes (Fig. 1A, Table S7) in the three databases according to the permutation importance score by the RSFVH algorithm based on TCGA data (Fig. 1B). To identify an optimal prognostic model and avoid over tting of the risk signature, the above 11 genes were subjected to multivariate Cox regression analysis with a stepwise model selection to establish an optimal gene combination for constructing the risk signature based on TCGA data. Finally, six genes (POSTN, IRX5, EEF2, RAB27A, MDM2 and ENO1) were ltered out (Fig. 1C).
Among the six genes, POSTN, IRX5, RAB27A, MDM2 and ENO1 were unfavorable factors for glioma survival with HR > 1, and EEF2 was a conservatory factor with HR < 1.
Construction and evaluation of the vitamin-related risk signature datasets also indicated that the vitamin-related risk signature had high sensitivity and speci city in prognostic prediction (Fig. 2B, C). To further examine the prognostic prediction ability of the vitaminrelated risk model in glioma, KM analysis was employed to determine the difference of OS between the low-and high-risk groups, which revealed that the low-risk group had a considerably increased OS compared with the high-risk group in the TCGA, CGGA, and GSE16011 datasets (Fig. 2D − F). The distribution of risk scores and survival status displayed in Fig. 2G and H and Figure S1 further con rmed the excellent prognostic prediction potential of the risk model. These results con rmed that the vitaminrelated risk signature could act as a reliable prognostic predictor for glioma patients.

Genetic and expression characteristics of the six genes
To better understand the six vitamin-related genes, we analyzed the correlations of the six genes in the TCGA, CGGA, and GSE16011 datasets. We found that POSTN and RAB27A had a signi cantly positive correlation while EEF2 was negatively correlated with RAB27A in the three datasets ( Fig. 3A − C). We next explored the expression of the six genes in glioma and normal brain tissue. GEPIA, the online tool with data sourced from TCGA and GTEx, was adopted to detect the mRNA expression of these genes (22). As show in Fig. 3D, except for EEF2, the expression of the remaining ve genes in GBM was higher than that in LGG, and the expression of the six genes in glioma was higher than that in normal brain tissue. We further explored the protein expression of POSTN, EEF2, MDM2, and ENO1 in the HPA dataset; characteristic images are shown in Fig. 3E. However, we did not nd IRX5 and RAB27A proteins expression in the HPA database. Consistent with the multivariate Cox regression analysis results, KM survival curves presented that samples with higher expression levels of POSTN, IRX5, RAB27A, MDM2, or ENO1 had worse outcomes and samples with higher expression levels of EEF2 had a more favorable outcome in the TCGA cohort (Fig. 3F − K). Then, we examined the genetic alteration of these riskassociated genes in glioma; among the 794 patients included in the cBioportal for Cancer Genomics database, 66 (8.3%) patients showed genetic alterations. Ampli cation was the most prevalent genetic alteration ( Figure S2).

The correlation of risk scores and clinicopathologic features
We further detected the risk scores in glioma patients classi ed by WHO grades, IDH status, 1p19q status, MGMT promoter, and subtype. The risk score was signi cantly increased along with WHO grades in the TCGA, CGGA, and GSE16011 cohorts ( Fig. 4A − C), and the remarkably increased risk scores were also observed in IDH wild type (Fig. 4D − F), 1p19q non-codeleted (Fig. 4G, H), MGMT promoter unmethylated (Fig. 4I, J), and mesenchymal subtype (Fig. 4K, L) glioma patients. These data suggested a notable association between the vitamin-related risk signature and clinical molecular features of glioma patients.

Prognostic signi cance of the risk signature in strati ed patients
To comprehensively examine the risk model in glioma, we detected the correlation between risk scores and OS of glioma patients classi ed by WHO grades, IDH status, 1p19q status, and MGMT promoter. KM analysis showed samples in the low-risk group had an obviously longer OS than those in the high-risk group for LGG patients in the TCGA, CGGA, and GSE16011 datasets ( Fig. 5A − C). Consistent results were also found in most strati ed samples expect in 1p19q codeleted patients (Fig. 5D − T). These ndings demonstrated the powerful potential of the vitamin-related risk model in predicting the outcome of glioma patients.

Functional enrichment analysis
To unearth the vitamin-related risk signature associated biological processes, the GSVA score of KEGG pathways and GO biological processes (c5.bp.v7.1.symbols.gmt) was investigated between the low-and high-risk groups in the TCGA, CGGA, and GSE16011 datasets. The top 30 KEGG pathways and top 30 GO biological processes signi cantly enriched in the high-risk group (p < 0.05) were screened out. We found that most of the processes were linked with immune and in ammatory responses in the TCGA (Fig. 6A −  B), CGGA ( Figure S3A − B), and GSE16011 ( Figure S4A − B) datasets. To con rm this result, we performed GO and KEGG analyses of the differentially expressed genes between the two groups. The biological processes and pathways enriched in the high-risk group was similar to the results of GSVA in the three datasets (Fig. 6C, D, Figure S3C, D, Figure S4C, D). Additionally, the GSEA analysis of GO biological processes in the three databases was also consistent with the GO analysis ( Fig. 7A − C). These data denoted that the vitamin-related risk model may be able to distinguish the immune characteristics of gliomas.
Immune features of the low-and high-risk groups Considering the role of vitamins in the immune system (11,12) and the results of the above-mentioned functional enrichment analysis, we speculated that the vitamin-related risk signature may be linked with the tumor immune characteristics of glioma. To verify this conjecture, we rst compared the ESTIMATE scores between the low-and high-risk groups. ESTIMATE scores in the high-risk group were signi cantly higher than that in the low-risk group in the TCGA (Fig. 8A), CGGA (Fig. 8E), and GSE16011 ( Figure S5A) datasets. Similarly, the immune and stromal scores were obviously higher in the high-risk group than in the low-risk group in the three datasets; while tumor purity showed the opposite result ( Fig. 8B − D, F − H, Figure S5B − D). Correlation analysis presented that the ESTIMATE, immune and stromal scores were all signi cantly positively correlated with the risk score, but tumor purity was negatively correlated with the risk score in the TCGA, CGGA, and GSE16011 datasets ( Fig. 8I − P, Figure S5E − H). These results indicated that the higher the risk score, the lower the purity of the glioma, indicating that the higher the risk score, the more complex the microenvironment of the glioma. We previously con rmed that, in glioma, lower tumor purity indicated a worse prognosis (33). This was accordant with the prognosis prediction of the risk signature of this study, and further con rmed that the vitamin-related risk model was accurate in predicting the tumor purity of glioma and the prognosis of glioma patients.
Immune cells, as an important component in the glioma microenvironment, affect tumor purity of glioma (34,35). As glioma purity is reduced, the local immune status can be enhanced (33) and the vitaminrelated risk score is signi cantly negatively correlated with tumor purity. We speculated that the risk score may be linked with the distribution of immune cells in the glioma microenvironment. To investigate their relationship, we employed GSVA to estimate signatures of multiple immune cells and compared the distribution of immune cells in low-and high-risk groups. We found that most immune cells, whether innate or adaptive immune cells, were notably more enriched in the high-risk group in the TCGA (Fig. 9A,  B), CGGA (Fig. 9C, D), and GSE16011 ( Figure S6A, B) datasets. This indicated that the higher the risk score, the more immune cells were enriched in the glioma microenvironment. At the same time, it also suggested that the vitamin-related risk signature can directly predict the local immune status of glioma, highlighting its strong predictive ability. It has been shown that higher immune cells in ltration indicates a worse outcome in both LGG and GBM (36). This further con rmed the accuracy and reliability of the vitamin-related risk signature in predicting the local immune response of the glioma microenvironment and the prognosis of glioma patients.
To further characterize intratumoral immune states, we analyzed the components of immune subtypes in the TCGA, CGGA, and GSE16011 datasets using the immuneSubtypeClassier package. In the three datasets, the C4 subtype was the main component of the high-risk group, while the C5 subtype was mainly concentrated in the low-risk group (Fig. 9E). The OS of tumor patients classi ed as C4 subtype is worse than that of tumor patients classi ed as C5 subtype (28), which is consistent with the prognosis of glioma in the low-and high-risk groups differentiated by the vitamin-related risk model. These results indicated that the vitamin-related risk model could not only accurately predict the immune status, but also could exactly reveal the prognosis of glioma patients.

Vitamin-related signature indicates an immunosuppressive microenvironment
In the process of the immune system's elimination of tumor cells, a series of anti-tumor immune response processes are initiated and expanded iteratively; these processes are de ned as the cancer-immunity cycle (37). Immune cells play a vital role in tumor clearance in this cycle. However, in glioma, the high-risk group had signi cantly higher immune cell enrichment than the low-risk group, but the prognosis was worse than that of the low-risk group, indicating that there may be other immunosuppressive mechanisms in the high-risk group. To explore this, we rst detected the expression of cancer-immunity cycle inhibitors in the two groups (38). The expression levels of most cancer-immunity cycle inhibitors in the high-risk group were obviously higher than those in the low-risk group in the TCGA, CGGA, and GSE16011 (Fig. 10A, B, Figure S7A) datasets. TGFB1, VEGFA, IL10, ARG1, CSF1, and FGL2 are immunosuppressive factors secreted by glioma cells, while CD70 and FASLG are glioma cell surface immunosuppressive factors (39). We further compared the expression levels of these genes in the two groups in the TCGA, CGGA, and GSE16011 datasets. As presented in Fig. 10C − E, the expression levels of these immunosuppressive factors in the high-risk group were evidently higher than those in the low-risk group in the three datasets.
In addition to cancer-immunity cycle inhibitors, immune checkpoints also play a vital role in tumor immune tolerance. Some immune checkpoint blockers have emerged as powerful weapons in the oncological armamentarium (40). Similar to cancer-immunity cycle inhibitors, the expression of most immune checkpoints in the high-risk group was notably higher than that in the low-risk group in the TCGA (Fig. 11A), CGGA (Fig. 11B), and GSE16011 ( Figure S7B) datasets. Since CTLA4, PD1 and PD-L1 blockers have shown clinical bene t in tumor treatment (41,42), we separately explored their expression in the two groups and their relationship with risk scores in the TCGA and CGGA datasets. As shown in Fig. 11C − N, the expression levels of CTLA4, PD1 and PD-L1 in the high-risk group were signi cantly higher than those in the low-risk group and signi cantly positively correlated with the risk scores.
These ndings indicated that glioma with high vitamin-related risk scores tended to form an immunosuppressive microenvironment via overexpression of cancer-immunity cycle inhibitors and immune checkpoints. This also explained why the high-risk group had more immune cells in ltration than the low-risk group but their prognosis was worse.
To examine whether the prognostic signi cance of vitamin-related risk model did not depend on other clinicopathological factors, univariate and multivariate Cox regression analyses were conducted in the TCGA dataset We detected that the risk signature was obviously linked with OS even when adjusted by other clinical factors (Fig. 12A, B). The same results were also obtained in the CGGA and GSE16011 datasets ( Figure S8A − D).
To develop a clinically utilizable method for predicting the 1-, 3-, and 5-year OS of glioma patients, we constructed a nomogram based on the TCGA cohort. The factors of the nomogram included clinicopathological variables (age, gender, WHO grade, IDH status, 1p19q status, and MGMT promoter status) and the vitamin-related risk score (Fig. 12C). Calibration curves were employed to evaluate the accuracy of the nomograms in predicting the OS of glioma patients. The 45-degree line denoted the most accurate prediction. Calibration curves displayed that the nomogram performed excellently in both the TCGA and CGGA cohorts (Fig. 12D, E). The DCA based on the TCGA dataset suggested that the risk scorebased nomogram could amplify net bene ts and display a broad range of threshold possibility in the prediction of 1-, 3-, and 5-year OS (Fig. 12F − H). Consistent results were also found in the CGGA dataset ( Figure

Discussion
Vitamins play a multi-faceted role in tumors. In addition to affecting their occurrence and progression, vitamins can also in uence the effects of radiotherapy and chemotherapy in tumor patients, and have an effect on the treatment of cachexia in tumor patients (43,44). However, different vitamins can have completely different effects. The speci c mechanisms of the different effects of vitamins in tumors are complex and diverse and involve a variety of signaling and biological pathways (19,20).
In this study, we rst selected all vitamin-related genes from MSigDB. From these, genes related to the OS of glioma patients in the TCGA, CGGA, and GSE16011datasets were screened out. Based on the TCGA database, vitamin-related six-gene risk model was constructed using survival-related genes. Among these six genes (POSTN, IRX5, EEF2, RAB27A, MDM2, and ENO1), vitamin D can suppress the metastasis potential of breast cancer cells by inhibiting the expression of POSTN through its receptor (45). IRX5 is also regulated by vitamin D in prostate cancer (46). RAB27 is related to the level of 25(OH)D in serum (47) and can regulate the invasion ability of melanoma cells (48). The expression of MDM2 is affected by folic acid and can inhibit the effects of ADR and p53 (49)(50)(51). ENO1 can regulate the metabolism of vitamin D3 and indole (52). EEF2 plays a vital role in the tumorigenesis and progression of some tumors (53)(54)(55).
We checked and veri ed the prognostic prediction potential of the risk signature in the above three databases via ROC curve, KM curve and risk score distribution plots. We also revealed correlations, expression characteristics, prognostic effects, and mutation characteristics of the six genes in glioma. These six genes were signi cantly linked with the outcomes of glioma patients, and the vitamin-related six-gene risk signature could accurately predict the outcome of glioma patients. Additionally, the risk signature can accurately distinguish different grades and molecular subtypes of glioma, and can predict the outcomes of glioma patients with different grades and molecular subtypes.
Functional enrichment analysis found that there was an accumulation of immune response-related pathways in the high-risk group. The ESTIMATE score displayed that the high-risk group had higher immune scores and lower tumor purity than the low-risk group, indicating that the risk model might be linked with the characteristics of the immune microenvironment of glioma. Immune cell component analysis con rmed this conclusion. The expression characteristics of cancer-immunity cycle suppression molecules and immune checkpoints veri ed that high vitamin-related risk scores were related to glioma immunosuppression.
To make full use of the vitamin-related risk signature, we combined it with clinicopathological factors to construct a nomogram. This nomogram has strong potential in predicting the outcome of glioma patients.
Although the vitamin-related risk signature has great potential in predicting the outcome of glioma patients, it still has some limitations. First, the construction of this model is based on the mRNA expression of the six genes, which needs to be veri ed at the protein expression level. In addition, due to the various types and functions of vitamins, we cannot enumerate the speci c role of every vitaminrelated gene set in glioma.

Conclusion
Overall, our study screened out six vitamin-related genes, which were used to construct an independent prognostic risk signature, that can accurately distinguish different WHO grades and molecular subtypes of glioma, and accurately predict the immune characteristics of glioma. Availability of data and materials: