Glioblastoma diagnostic models and therapeutic drug discovery based on GEO data and machine learning methods

Glioblastoma (GBM) remains lethal due to high molecular heterogeneity and treatment resistance. While previous studies have proposed various biomarkers, a critical research gap exists: the lack of robust algorithmic validation and systematic linkage to drug discovery. Existing research predominantly relies on single machine learning models or traditional statistics, which often fail to provide stable results across diverse clinical datasets. To address this, we developed a high-dimensional pipeline that compares and ensembles 175 machine learning algorithm combinations. Unlike conventional single-model workflows, this approach ensures superior target identification stability and utilizes SHAP-based explainability and molecular dynamics to bridge the gap between biomarker discovery and precision therapeutics.

DEGs from GEO datasets were refined via PPI and functional analyses. The 175-algorithm ensemble identified core genes, with clinical utility validated via survival analysis. A drug discovery pipeline incorporating virtual screening, ADMET, and molecular dynamics (MD) was then implemented to evaluate compounds targeting the identified core genes.

From 771 DEGs, 34 key genes were identified, with LOX validated as the core therapeutic target. The optimal predictive model achieved a robust AUC of 0.953, while survival analysis underscored the significant prognostic value of LOX. Following systematic screening, the most outstanding compound was prioritized via MD simulations, exhibiting exceptional binding stability, favorable pharmacokinetics, and minimal toxicity risk.

This integrated pipeline provides a robust framework for identifying precision targets and potent candidate compounds, offering a novel strategy for overcoming GBM treatment barriers.