A novel inositol phosphate metabolism signature predicts survival prognosis in low grade glioma

Abstract

Low grade gliomas (LGGs) are heterogeneous tumors with high recurrence rates and variable clinical outcomes. However, metabolic dysregulation within LGG and its prognosis and tumor pathology remain poorly understood. We investigated the significance of inositol phosphate metabolism (IPM) in LGG and constructed an integrated machine learning survival model to improve the prognostic accuracy. By examining the transcriptomic data of LGGs, we identified a prognostic IPM (pIPM) signature consisting of 33 genes, with PLCG1 and MTMR3 showing the highest and lowest hazard ratios, respectively. Subsequent consensus clustering analysis revealed two distinct clusters (C1 and C2) with different pIPM expression levels and overall survival outcomes. Furthermore, IPM pathway analysis revealed an overall poor prognosis except for MGAT4C, suggesting that IPM plays a key role in driving LGG progression. Using machine learning algorithms, we developed and validated a robust survival model based on pIPM, in which Lasso model showed excellent discriminative power and reliability on both training and validation datasets. Furthermore, an increased IPM risk score was associated with increased cell proliferation and inflammation. Remarkably, the score was linked to treatment response, tumor recurrence, and remodeling of the immune microenvironment, emphasizing its potential as both a prognostic biomarker and a therapeutic target for LGG. In conclusion, our findings revealed a complex involvement of phosphatidylinositol metabolism in the LGG pathogenesis and elucidated its clinical significance for prognostic prediction and therapeutic intervention.

Introduction

Gliomas represent the most prevalent primary malignant neoplasms within the cerebral nervous system, constituting 81% of all malignancies affecting the central nervous system (CNS) [1]. Based on histological classification, the World Health Organization (WHO) has divided gliomas into four grades: grades I and II are designated as low grade gliomas (LGG), whereas grades III and IV are categorized as high-grade gliomas (HGGs). LGGs account for approximately 6% of primary tumors in the adult CNS, and their prognosis remains poor despite a combination of surgical resection and radiotherapy. The median survival of patients with LGG is only 5–10 years, and more than half of patients experience recurrence, with as many as 17%–32% developing HGG [2–6]. A significant gap exists between the current clinical interventions and knowledge of the molecular pathology of these tumors. Therefore, studying the complex molecular pathology of patients with LGGs is essential to reduce the recurrence rate, prevent or delay progression to HGG, and ultimately improve clinical outcomes.

Inositol phosphates are important organic compounds that act as secondary messengers in cellular processes [7]. The dynamic interconversion of phosphatidylinositol compounds through a series of phosphorylation and dephosphorylation reactions is central to inositol phosphate metabolism (IPM). This dynamic system is critical for many cellular activities, including signal transduction, cell proliferation, survival, metabolism, cytoskeletal changes, vesicular trafficking, and membrane dynamics [8, 9]. Recent studies have shown that phosphatidylinositol metabolism plays an important role in the pathophysiology of various tumors, including LGG [10, 11]. Inositol, a molecule synthesized primarily in astrocytes, is involved in osmoregulation and brain volume maintenance and has emerged as a potential biomarker [12]. Elevated inositol levels are usually detected in LGGs with a favorable prognosis, whereas decreased inositol levels are observed in HGGs [13]. Understanding metabolic alterations in LGGs, particularly those related to IPM, holds promise for the development of novel therapeutic strategies. Targeting specific enzymes or pathways in this metabolic pathway has the potential to slow tumor progression or improve the efficacy of existing therapies. Therefore, investigating a role of IPM in LGG pathogenesis is important for the prognostic assessment and treatment of gliomas.

The ever-increasing scale and inherent complexity of biological data have led to the emergence of machine learning (ML) as an important tool in biology for constructing information and predictive models of underlying biological processes [14]. The use of ML became widespread as biomarkers are increasingly studied along with the diseases mechanisms. However, the current levels of accuracy and efficiency of bioinformatic methods may not be sufficient to screen prognostic markers for IPM. Therefore, we investigated the prognostic significance of IPM in LGG using ML algorithms in combination with other bioinformatic methods.

In this study, we aimed to identify key genes associated with prognostic IPM (pIPM) signature and patients’ survival by conducting an analysis of extensive transcriptomic datasets from The Cancer Genome Atlas (TCGA). To improve LGG prognosis and explore potential prognostic biomarkers and therapeutic targets, we constructed a pIPM-based survival model using ML. By integrating multi-omics data derived from both TCGA and the Chinese Glioma Genome Atlas (CGGA), our study explored the relationships between IPM and a range of clinical parameters, including treatment response, tumor recurrence, and characteristics of the immune microenvironment as well as the linkage between IPM dysregulation and LGG progression. This research offers valuable understanding of potential prognostic biomarkers and therapeutic targets. Put together, our study investigates the complex biological roles of IPM and LGG and provides new ideas for developing personalized treatment strategies for LGG and improving its prognosis.

Materials and methods

LGG transcriptome data acquisition and processing

The LGG data used in this study were obtained from TCGA database using the R software package (TCGAbiolinks) [15]. The Genomic Data Commons (GDC) data portal was used to obtain survival statistics, tumor staging, and other clinical information pertinent to LGG (https://gdc.cancer.gov/about-data/publications/pancanatlas). Two transcriptome datasets, mRNA_seq325 and mRNAseq_693, were included from the China Glioma Genome Atlas (CGGA) study website (http://www.cgga.org.cn). We performed log2(x + 1) transformations, missing value checking, and quartile normalization on all datasets to harmonize the three LGG datasets. Outlier detection using hierarchical clustering methods did not yield significantly abnormal results. To ensure that all genes were consistently labeled in both datasets, we compared them using a common genome annotation platform. Batch correlation analysis was performed using the limma normalizeBetweenArrays function and ENSEMBLE IDs were converted to gene symbols using the biomaRt function [16].

Single-sample gene set enrichment analysis (ssGSEA)

From the GSEA website (https://www.gsea-msigdb.org/gsea/index.jsp), we obtained the marker gene set and conducted a ssGSEA analysis utilizing the Gene Set Variation Analysis (GSVA) R package, using specified parameters: method=’ssgsea’, kcdf=’Gaussian’, and abs. ranking = TRUE [17].

Survival analysis

To account for variations in patients’ survival across different scenarios, we conducted Cox hazard analysis and Kaplan–Meier (KM) survival analysis utilizing the R survival package, with the subgroups for high and low values in the KM analysis being determined by the median values of the corresponding conditions.

Estimation of cell infiltration

Utilizing the R package xCell, we conducted ssGSEA to measure the proportional infiltration of 28 immune cells within the LGG cohort, encompassing both TCGA and CGGA datasets [18].

Consensus clustering

Utilizing the infiltration patterns of various immune cells, we applied a consensus clustering sampling technique to pinpoint 33 noteworthy IPM survival genes within the TCGA LGG dataset. The optimal cluster count was ascertained by integrating the consensus score matrix, cumulative distribution function (CDF) curves, Proportion of ambiguous clustering (PAC) scores, and Nbclust. For consensus clustering, we relied on the R package ConsensusClusterPlus [19], whereas the identification of differentially expressed genes (DEGs) was conducted using the R package limma [20].

Construction and validation of IPM-based ML survival models

Owing to the variability in ML algorithms, it is currently impossible to determine the optimal and most accurate algorithm. To improve the accuracy of this study, we integrated 10 ML algorithms and 22 combinations of algorithms for IPM prediction: Random Survival Forest (RSF), Enet, Lasso, Ridge, Stepwise Cox, CoxBoost, Cox partial least squares regression (plsRcox), Supervised Principal Component (SuperPC), Generalized Boost Regression Model (GBM), and Survival Support Vector Machine (Survival-SVM). Additionally, we constructed and validated a survival model for LGG based on IPM. The specific process includes (1) selecting 22 IPM genes identified by survival analysis in the TCGA LGG cohort as features for ML survival models; (2) building survival models based on a combination of 22 algorithms in the TCGA LGG cohort; (3) evaluating the constructed survival models using the TCGA LGG training set and two CGGA validation sets; and (4) computing the survival metrics, namely Harrell’s consistency index (C-index) and area under the curve (AUC), for each model at 1, 3, and 5 years in both the training and validation datasets. Considering the highest C-index and AUC values, along with Occam’s razor principle, we determined the Lasso model to be the most suitable choice.

Statistical analysis

R software (version 4.3.1) was used to perform all data processing and statistical analyses. The ssGSEA scores for different conditions were statistically analyzed using the Wilcoxon test. The relationship between two continuous variables was evaluated using the Pearson correlation coefficient. For the comparison of categorical variables, the chi-squared test was utilized. In contrast, the Wilcoxon rank-sum test or t-test was applied for the comparison of continuous variables. The visualizations of KM survival plots were created by employing the “ggsurvplot” function from the R package survminer. Cox regression and Kaplan-Meier analyses were executed using the Survminer package. To correct for multiple comparisons, we incorporated a false discovery rate (FDR), and set a q-value threshold of < 0.05 to determine statistical significance.

Results

Determination of prognostic IPM features in the TCGA LGG cohort

Figure 1 depicts the study design. Initially, a prognostic IPM signature was established through survival analysis of the TCGA LGG transcriptome. Among the 73 IPM genes analyzed, 33 genes were significant in the survival analysis using KM and Cox tests (pIPM; Figs. 2A B). Among these 33 pIPM genes, PLCG1 and MTMR3 had the highest and lowest hazard ratios (HR), respectively (Fig. 2B). We then performed a consensus clustering analysis based on these pIPM genes to classify all TCGA LGG tumor subjects into k clusters. The consensus score matrix’s cumulative distribution function (CDF) curve and the proportion of ambiguous clustering (PAC) statistics are the basis for this (Figs. 2C, D), two was chosen as the optimal value among the tested k values (k = 2–9).The two clusters (C1 and C2) obtained from the consensus clustering analysis differed significantly in the expression of pIPMs, with the overall infiltration frequency of C2 being substantially higher than that of C1 (Fig. 2E). In the KM curves, the overall survival rate of C2 was substantially higher than that of C1 (Fig. 2F). Notably, C1 was more inclined to synthesize inositol, whereas C2 was more inclined to synthesize 1D-inositol-1,3,4,5,6P5 (Fig. 2G).

Fig. 1.

The Workflow of the study

Fig. 2.

Inositol phosphate metabolism genes were prognostic in LGG.A The Venn plot of IPM pathway, KM-significant and univariate-COX-significant genes. B For each k, the CDF curves of the consensus matrix are presented. The color indicates the relative k value. C When k equals 2, the consensus score matrix of all samples is presented. A lower consensus score between two samples suggests that they are more prone to being clustered into the same group in different iterations. D The forest plot of prognostic IPM signatures. The HR > 1 indicates the poor prognosis while HR < 1 represents the opposite. E The heatmap of the prognostic IPM genes in two identified clusters by ConsensusClusterPlus. The relative sample information is annotated on the top. F KM curves of overall survival between to two sub-clusters. G The IPM pathway of 2 clusters. The log2FoldChange values are annotated by the color while red represents high-expressed in C1 and green represents high-expressed in C2

Construction and validation of IPM-based hierarchical ML survival models

To improve the results of LGG, we developed pIPM-based survival models. Using a ML integration procedure, we constructed 22 candidate survival models from 33 pIPMs. Utilizing the leave-one-out cross-validation (LOOCV) method, the program was trained on the TCGA LGG dataset. Subsequently, the C-index for each model was computed for both the TCGA LGG training set and two CGGA validation datasets, as depicted in Fig. 3A. Of the 22 candidate models constructed, 21 showed good discriminative power (average C-index > 0.7), mainly for Enet-related models (Fig. 3A). In a subsequent receiver operating characteristic (ROC)analysis, the Enet-related models still showed good performance, with high C-index and AUC rankings on both the training and validation datasets (Fig. 3B). Additionally, based on Occam’s razor, we selected the Enet (α = 1) survival model for further analysis, which is also known as Lasso model. Within this model, the optimal λ value was ascertained through LOOCV, and seven pIPMs were utilized. Notably, PLCG1 exhibited the highest coefficient, underscoring its significance as a key feature (as depicted in Figs. 3C, D). The AUCs for the TCGA training dataset at 1, 3, and 5 years were 0.857, 0.854, and 0.737, respectively (as shown in Figs. 3E–G). Comparable AUCs were noted in the CGGA325 and CGGA639 datasets, demonstrating high reliability of this survival model across both the training and validation datasets (Figs. 3E–G). Moreover, the survival model demonstrated remarkable consistency (p < 0.0001) for patients in the high-risk group, in both the training and validation sets, attesting to its robustness (as illustrated in Figs. 3H–J).

Fig. 3.

Construct and validate the integrated machine learning survival models based on IPM.A The heatmap of Concordance Index (C-index) of 33 prediction models via Leave-One-Out Cross-Validation (LOOCV) framework in the TCGA training dataset and 2 CGGA validation datasets. The mean C-index is on the right. B The heatmap of Area Under the Curve (AUC) of 33 prediction models via LOOCV framework in the TCGA training dataset and 2 CGGA validation datasets. C Partial likelihood deviance was plotted against log (lambda) in the Lasso model. The vertical dotted lines indicate the lambda value with minimum error and with one standard error (SE) of the minimum. D The coefficients of Lasso model. E–G The 1,3 and 5 years of ROC curve of the lasso model in TCGA (E), CGGA325 (F) and CGGA693 (G). The colors indicate the relative time. H–J The KM curve of risk score predicted by Lasso model in in TCGA (H), CGGA325 (I) and CGGA693 (J)

Lasso survival model is the best to improve LGG outcomes

To further clarify the molecular characteristics associated with IPM, we conducted DEG analyses in both low- and high-risk patients with LGG using all three datasets. By considering an intersection of analyzed results from high-risk patients in the three cohorts, 431 and 493 DEGs were obtained in the enriched and suppressed groups, respectively (Figs. 4A, B). Among these DEGs, the upregulated genes were mainly enriched in terms of infections and immune disorders such as Staphylococcus aureus infection, rheumatoid arthritis, and type I diabetes mellitus, whereas the downregulated genes were mainly associated with normal neurological activity (Figs. 4C, D). Additionally, we compared a proliferative status of tumor cells in the high- and low-risk groups and found that high-risk patients had greater proliferation scores in all three LGG cohorts (Fig. 4E). Subsequent assessment of marker signaling heterogeneity between the high- and low-risk groups revealed that hypoxia, inflammatory response, and epithelial-mesenchymal transition (EMT) levels were higher in high-risk patients than those in the low-risk group (Fig. 4F).

Fig. 4.

The elevated risk score of IPM was associated with increased levels of both cell proliferation and inflammation. A, B The Venn plot of enriched (A) and suppressed (B) DEGs in high-risk patients among the TCGA and CGGA cohorts. C, D The dot plot of shared enriched (C) and suppressed (D) DEGs KEGG enrichment analysis. E Violin plot of proliferation score between high- and low-risk groups across TCGA and CGGA cohorts. F, G The heatmap illustrates the statistical differences in hallmark (F) and KEGG metabolic pathways (G) ssGSEA scores between the high- and low-risk groups across both the TCGA and CGGA cohorts

To gain a comprehensive view of metabolic alterations between high- and low-risk patients across the three datasets, we compared KEGG metabolic pathway scores (Fig. 4G). High-risk patients showed upregulation of pathways primarily related to sugar metabolism, amino acid metabolism, and glycan biosynthesis and degradation, reflecting enhanced energy production and macromolecule turnover. Representative pathways included amino sugar and nucleotide sugar metabolism, cysteine and methionine metabolism, galactose metabolism, glutathione metabolism, and various types of N-glycan biosynthesis. In contrast, low-risk patients exhibited enrichment in pathways associated with lipid metabolism, fatty acid synthesis, and secondary metabolite biosynthesis, suggesting a metabolic profile favoring lipid processing and signaling. Typical pathways included alanine, aspartate and glutamate metabolism, fatty acid biosynthesis, glycerolipid metabolism, steroid hormone biosynthesis, and terpenoid backbone biosynthesis. These results highlight a strong association between IPM, tumor cell proliferation, the inflammatory microenvironment, and distinct metabolic reprogramming within tumors.

Lasso risk score characterizes recurrence enhancement and macrophage activation in the LGG cohort

We further investigated the association between IPM risk score and other clinical and immunological features of LGG. Notably, the risk score was significantly lower in the responding and stable LGG groups than that in the progressive phase (p < 0.0001; Fig. 5A). Notably, IPM risk scores were higher in recurrent tumors than those in primary tumors for both TCGA and CGGA datasets (Figs. 5B–D), suggesting that IPM may have played a more important role initially than it did in current treatment. Finally, we performed a joint analysis of the IPM risk score and immune microenvironment to further explore the immune factors associated with IPM. Notably, the IPM risk score showed a consistent positive relationship with macrophages, DCs, monocytes, mast cells, CD4 + memory T cells, Tgd cells, Th2 cells, and B cells, and negatively correlated with regulatory T cells (Tregs) (Fig. 5E). These results suggested that IPM affects LGG progression in multiple ways.

Fig. 5.

The risk score predicted by Lasso model indicates the treatment response and tumor recurrence, and is positive correlated to monocyte/macrophage/DC infiltration. A The boxplot of risk score in differential response to first treatment. B–D The boxplot of risk score in primary and recurrent tumors in TCGA (B), CGGA (C) and CGGA (D). E The heatmap of spearman correlation of risk score with immune infiltration in TCGA and CGGA datasets. * P < 0.05, ** P < 0.01, *** P < 0.001, **** P < 0.0001

Discussion

Our study represents the initial thorough exploration of the prognostic importance and functional consequences of IPM in LGG. We thoroughly examined three transcriptomic datasets obtained from the TCGA and CGGA LGG cohorts and determined the predictive worthiness of IPM in LGG, created a prognostic model utilizing IPM, and delineated the molecular and tumor microenvironmental variations between patient groups with high and low risk. A 33-gene signature (pIPM) was identified using survival analysis that highlighted the prognostic relevance of IPM in LGG and showed a significant association with patients’ survival.

Among the pIPM genes that were screened, PLCG1, the most malignant gene, has been reported to promote a growth and metastasis of a variety of tumors [21] and lead to reduced survival in patients with LGG [22]. It could be attributed to PLCG1’s diverse biological functions and its capacity to promote cell growth in various cancers through interactions with a range of molecules and second messengers, ultimately leading to a positive impact on cell growth across multiple cancer types [23–26]. Recent studies have shown that reduced PLCG1 expression can inhibit tumor proliferation, migration, and invasion and promote apoptosis by disrupting essential PLCG1 functions such as protein secretion, MYC targets, mitotic spindle, G2/M checkpoint, and E2F targets [22]. Consistent with these reports, we observed enhanced inflammatory responses in the high-risk group, including activation of type I/II interferon signaling, interleukin-6 (IL6), and tumor necrosis factor-α (TNF-α). Collectively, these findings suggest that PLCG1 serves as a pivotal regulator of multiple physiological processes and contributes to the malignant progression of LGG.

Beyond PLCG1, other pIPM genes such as PIK3CG and PIK3CD have been implicated in promoting natural killer cell chemotaxis, as well as macrophage and tumor cell migration and colony formation [27–29]. Moreover, our results revealed enhanced glycolysis and altered cholesterol metabolism in high-risk patients, further emphasizing the significance of IPM-associated metabolic reprogramming in LGG pathogenesis.

Although our study is primarily based on bioinformatic analyses, the consistent identification of PLCG1 as a key driver gene underscores its potential mechanistic relevance in LGG. Given its established role in intracellular calcium signaling and downstream MAPK and NF-κB activation, PLCG1 dysregulation may facilitate tumor proliferation, survival, and immune modulation. Notably, exploring the relationship between PLCG1 expression and well-defined molecular subtypes—such as IDH mutation status and 1p/19q co-deletion—could further refine its prognostic and biological interpretation. Future in vitro and in vivo investigations focusing on PLCG1 knockdown or overexpression will be crucial to validate its role in malignant phenotypes, thereby supporting its potential as both a biomarker and a therapeutic target in LGG.

Our data revealed a link between IPM risk scores and the reshaping of tumor microenvironment. We found that low risk scores were often seen in patients with a good response to initial treatment, whereas high scores were associated with tumor recurrence. Inflammatory responses such as gamma interferon (IFN-γ) and transforming growth factor beta (TGF-β) have been reported to be associated with poor prognosis in low-grade gliomas (LGGs) and to correspond to specific mutational patterns [30–32], aligning with our observations. Functional analyses showed that pathways and inflammatory factors (TNF, IL6, TGF-β, and IFN) associated with immune response, cell proliferation, and inflammatory signaling were more active in patients at high risk for LGG than in those at low risk. We also observed an association between the inflammatory response and EMT, hypoxia, and angiogenesis. Furthermore, a robust relationship was observed between IPM risk scores and inflammatory phenotypes (M1 macrophages and cDCs) as well as immunomodulatory phenotypes (Th2 cells and Tregs). There was a correlation indicating an increase in the infiltration of specific immune cell subsets, such as macrophages, dendritic cells (DCs), and memory T cells, and a negative association with Tregs. It is believed that macrophages and DCs undergo reprogramming within a glioma microenvironment, which may contribute to the inhibition of immune responses and acceleration of tumor progression [33–35]. The possible immunosuppressive function of Tregs in the tumor merits additional exploration [36]. These discoveries indicate that IPM is tightly linked to LGG metabolism and the immune microenvironment, suggesting its potential as a metabolic checkpoint in molecular pathology and clinical applications. This opens up avenues for identifying potential therapeutic targets and conducting further research.

Several survival models based on pIPM features have shown strong discriminative abilities in multiple datasets. In all of these ML models, the pIPM features showed high accuracy and sensitivity, supporting an idea that IPM is a key feature in LGG prognosis. Additionally, the simplicity and excellent performance of the Enet model highlight the simplicity and predictive accuracy of the IPM and its potential clinical use in predicting the prognosis of patients with LGG.

Limitations

Notably, this study has a few limitations. Firstly, while we pinpointed prognostically significant IPM features, further validation through functional experiments both in vitro and in vivo is necessary to clarify a detailed mechanism of connection between IPM dysregulation and the LGG advancement. Secondly, to ascertain the cellular origin of IPM, it is essential to employ cutting-edge research techniques, like single-cell RNA sequencing or experiments. Finally, the model validation cohort used was rather limited and needs to be validated in a large and diverse patient cohort to ensure accuracy and generalizability.

Conclusions

To summarize, our research thoroughly uncovered the prognostic importance and functional influence of IPM in LGG. By integrating multi-omics data and applying machine learning techniques, we identified the impact of phosphatidylinositol metabolism on tumor biology and patient prognosis, offering insights for future investigations and potential therapeutic strategies in the treatment of LGG.

Author contributions

This original study was designed by Chunfa Qian and Yihao Zhu. Data collection, technical support, and conceptual advice were provided by Yihao Zhu, Jindou Wei and Yang Xu. Bioinformatics statistical analysis was done by Yihao Zhu and Fuduo Bo. All authors are consistent to publish this article.

Funding

This study was funded by the Youth Talent Project of Nanjing Brain Hospital (Grant No. 23-25-2R7) and the General Project of Nanjing Medical Science and Technology Development (Grant No. YKK22142).

Data availability

This study incorporated data from publicly accessible websites, as detailed in the “Materials and Methods” section. The Genomic Data Commons (GDC) data portal was obtained from https://gdc.cancer.gov/about-data/publications/pancanatlas. Two transcriptome datasets could be accessed from the TCGA website (http://www.cgga.org.cn). The single-sample gene set enrichment analysis (ssGSEA) was conducted in the website (https://www.gsea-msigdb.org/gsea/index.jsp).Additional datasets used and analyzed during the current study are available from the corresponding author on reasonable request.

Declarations

Ethics approval and consent to participate

Not applicable.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.