AI-Derived Prognostic Model Identifies High-Risk Gene Signatures in Pediatric Gliomas

Abstract

Background: Pediatric gliomas, comprising both low-grade (LGGs) and high-grade gliomas (HGGs), exhibit significant molecular and clinical heterogeneity. While LGGs generally have a favorable prognosis, HGGs are associated with poor long-term survival despite aggressive treatment. Advances in molecular profiling have enabled targeted therapies, but treatment resistance and tumor heterogeneity remain major challenges. The integration of artificial intelligence (AI) and transcriptomic data holds promise for refining prognostic models and guiding personalized treatment strategies, yet its application in pediatric gliomas remains underexplored. Method: We applied the Artificial Intelligence-Derived Prognostic Index (AIDPI) model to analyze transcriptomic data from pediatric glioma patients. Differentially expressed genes (DEGs) were identified and incorporated into a machine learning-based prognostic model. Single-cell RNA-seq data were also integrated to assess cellular heterogeneity within the tumor microenvironment. Kaplan-Meier survival analysis, Cox regression, and receiver operating characteristic (ROC) curve analysis were performed to evaluate the model ' s predictive power. Functional enrichment analysis was conducted to explore potential therapeutic targets. Results: The AIDPI model identified nine key genes (GRIA1, ZNF165, TM9SF2, PRKAR2A, PSMD6, H1F0, CDC25B, HIST1H2AE, and NCAPD2) that were consistently associated with prognosis across multiple pediatric glioma datasets. These genes were used to construct a machine learning-based prognostic model, which demonstrated superior predictive performance with a C-index > 0.85. High AIDPI scores correlated with poorer survival outcomes, as confirmed by Kaplan-Meier survival analysis and time-dependent ROC curves. The AIDPI model outperformed 30 other glioma prognostic models, highlighting its potential for precision prognosis. Functional analysis of the AIDPI-related genes revealed involvement in immune suppression and cell adhesion pathways. Single-cell analysis identified TM9SF2 and H1F0 as key prognostic genes, with high H1F0 expression being associated with poor prognosis in pediatric gliomas. Conclusions: Our findings highlight the potential of AI-driven transcriptomic analysis in improving pediatric glioma prognosis. The identified gene signatures may serve as biomarkers for risk stratification and personalized treatment strategies, advancing precision oncology in pediatric neuro-oncology.