TMET-09. Peripheral Blood Metabolomics and Machine Learning Reveal Diagnostic Biomarkers for Glioma

BACKGROUND Gliomas, the most common malignant brain tumors, require non-invasive diagnostic tools due to limitations of current invasive methods. This study explored peripheral blood metabolomics combined with machine learning(ML) to identify metabolic biomarkers for glioma diagnosis and subtyping. METHODS Serum untargeted metabolomics (UPLC-QTOF/MS) was performed on 152 individuals (mixed cohort). Transcriptomic/clinical data for an additional 361 subjects were sourced from Beijing Tiantan hospital. Patients (n=429) were categorized: Glioblastoma (GBM), Astrocytoma, and Oligodendroglioma; healthy controls (n=84 total). Cohorts were split into discovery (n=101) and validation (n=51) sets. Differential metabolite analysis identified significant features (p<0.05, |log2FC|>1). Pathway enrichment (KEGG) was performed. A two-step ML framework employed multiple algorithms for biomarker panel selection (PCA, oPLS-DA, RF, etc.) and classification models (SVM, PLS-DA, LR, RF), assessed via cross-validated AUC. RESULTS Analysis revealed 1,145 unique metabolites. In Glioma vs Normal (906 differential metabolites), PCA and oPLS-DA-derived panels achieved high validation AUCs (0.964 and 0.952, respectively). GBM vs Normal and LGG vs Normal comparisons yielded outstanding discrimination (AUCs > 0.97). GBM vs LGG separation was moderate (best AUC 0.726). Oligodendroglioma vs Astrocytoma discrimination proved challenging. Key dysregulated pathways included amino acid and fatty acid metabolism. Pyruvic acid and sphingosine-1-phosphate emerged as robust discriminators. CONCLUSION Blood metabolomics integrated with ML effectively distinguishes glioma patients from healthy individuals and differentiates major subtypes, achieving high diagnostic accuracy. This approach identifies robust metabolic biomarker panels supporting the feasibility of non-invasive, early glioma detection.