Artificial intelligence-based diagnostic model for identifying PTPRZ1-MET fusion in a clinically defined secondary glioblastoma cohort

Glioblastoma (GBM) is an aggressive and highly lethal brain tumor. Secondary glioblastoma (sGBM), which arises through malignant progression from lower-grade diffuse glioma, represents a clinically and biologically distinct subset of GBM. In this disease context, the protein tyrosine phosphatase receptor type Z1-mesenchymal-epithelial transition factor (PTPRZ1-MET; ZM) fusion has emerged as a recurrent oncogenic driver associated with adverse clinical outcomes. In this study, we analyzed 159 patients with sGBM, including 15 ZM-positive and 144 ZM-negative cases, and confirmed that ZM-positive tumors were associated with significantly shorter overall and progression-free survival. Transcriptomic profiling identified 359 genes upregulated in ZM-positive tumors, with enrichment in cell-cycle regulation and mitotic spindle-related pathways. To explore surrogate biomarkers associated with ZM fusion status, we benchmarked eight machine-learning classifiers and retained XGBoost as the primary feature-prioritization model. MET, PCDHGA3, and FAM3C emerged as the most informative biomarkers, and the fixed three-gene panel showed stable discriminative performance across cross-validation, nested evaluation, feature-pool sensitivity analyses, repeated random seeds, and class-weighted modeling. Protein-level validation in an independent formalin-fixed, paraffin-embedded (FFPE) cohort using multiplex and conventional chromogenic immunohistochemistry supported the pathology-compatible detection of elevated MET, PCDHGA3, and FAM3C expression in ZM-positive tumors. Collectively, these findings support a robust three-gene molecular signature that may facilitate the identification and stratification of ZM fusion-positive sGBM and provide biological insights with potential translational relevance for precision glioma diagnosis.