INTERPRETABLE ENSEMBLE LEARNING PREDICTS GLIOBLASTOMA SENSITIVITY TO NATURAL COMPOUNDS

Abstract

Glioblastoma is a belligerent heterogeneous type of brain tumor, inherently difficult to treat with a dismal prognosis. Conventional therapies are rendered ineffectual by various limitations associated with the blood-brain barrier (BBB), tumor microenvironment, and adaptability manifested by the tumors. This study developed an interpretable machine learning framework to predict drug sensitivity in GBM and identify potential new therapeutic candidates. An XGBoost regression model was trained on a curated dataset integrating drug response data (from GDSC with baseline transcriptomic profiles Drug features included 1024-bit Morgan fingerprints and 9 key physicochemical/ADME properties, while cell line features comprised 100 gene expression markers selected via Recursive Feature Elimination. The final model demonstrated strong predictive performance, achieving a mean R² ~ 0.833 and a mean RMSE ~ 1.060 across five repeated train-test splits. SHAP analysis provided crucial insights into model predictions, identifying key drivers of model learning and the expression levels of GBM-relevant genes like ZEB2 and ABCB6. The model was subsequently used to screen a filtered subset of the COCONUT natural product database, identifying several compounds predicted to exhibit high potency. CNP0152293.3 was the top identified compound.