A Ten-Gene Transcriptomic Biomarker Panel for Glioma Classification and Prognosis Identified via Integrative Hypergraph and Rough Set Analysis

Background: Clinically actionable biomarkers that reliably distinguish glioblastoma (GBM) from lower-grade glioma (LGG) across expression platforms remain an unmet need. Existing transcriptomic signatures are frequently confounded by batch effects, platform heterogeneity, and the inability to translate to single-patient clinical workflows. Methods: We developed a topology-aware biomarker discovery framework in which analysis-of-variance ranking defines a candidate gene pool, hypergraph co-expression analysis at correlation threshold τ=0.75 identifies densely connected hubs within this pool, rough set reduct optimisation selects a minimal sufficient subset of these hubs, and a Random Forest classifier with stratified cross-validation performs the final discrimination. The pipeline was trained exclusively on GSE16011, a single-platform single-institution Affymetrix microarray cohort free from batch-class confound, and validated on two independent RNA-sequencing cohorts (CGGA-325 and CGGA-693). Robustness was further assessed through bootstrap optimism correction, DeLong cross-cohort equivalence testing, leave-one-gene-out analysis, and a sensitivity analysis under WHO CNS5 (2021) class definitions. Results: The pipeline identified a ten-gene biomarker panel (CSMD3, CHI3L1, PLP2, FRY, FCHSD2, ADM, MCUB, ANXA1, DUSP26, and HK2), achieving a fivefold cross-validation AUROC of 0.906±0.029 and a held-out AUROC of 0.831. External validation yielded AUROC =0.838 in CGGA-325 and AUROC =0.836 in CGGA-693. The biomarker-derived risk score demonstrated independent prognostic value in CGGA-693 (multivariate Cox hazard ratio =9.195; p<0.001) after adjustment for WHO histological grade, with Kaplan–Meier analysis confirming highly significant survival separation (log-rank p=4.60×10−37). Class definitions in the present work follow the histology-based pre-2021 WHO classification used in the source datasets and do not directly incorporate WHO CNS5 (2021) molecular criteria, such as IDH mutation status, that distinguish IDH-wild-type glioblastoma from IDH-mutant grade-IV astrocytoma. After excluding IDH-mutant grade-IV cases from the CGGA cohorts, the classification AUROCs increased to 0.906 in CGGA-325 and 0.872 in CGGA-693, with a Cox risk-score hazard ratio of 8.57 (p=1.4×10−13) and log-rank p=1.4×10−32 retained on the CNS5-aligned cohort. Conclusions: The methodological contributions introduced in this study, namely, the topology-aware hypergraph candidate pool construction, the rough set combinatorial reduct selection, the fixed-reference single-sample normalisation protocol, and the nested validation regime combining bootstrap optimism correction with cross-platform DeLong testing, are platform agnostic and directly applicable to future CNS5-aligned cohorts as such resources become publicly available, supporting the prospective re-derivation of molecularly defined glioma signatures within the integrated histopathological and molecular frameworks of contemporary neuro-oncology.