Machine Learning Informatics for Targeted Glioblastoma Photodynamic Therapy using optimized Yb/Er doped Upconversion Nanomaterials: A Multi-Photosensitizer Theranostic Framework

Description

Abstract (English)

Glioblastoma Multiforme (GBM) remains a lethal primary brain tumor due to micro-invasive infiltration and poor light penetration in cranial tissue (<2-3 mm). In the scope of nanomedicine and nanobiotechnology, lanthanide-doped upconversion luminescent nanomaterials offer an unprecedented non-invasive therapeutic platform. These smart functional biomaterials convert deep-penetrating 980 nm near-infrared (NIR) light into visible/UV emissions in situ, activating bound photosensitizers to generate cytotoxic reactive oxygen species (ROS) [1]. However, optimizing host lattice composition, dopant ratios, and particle size for diverse photosensitizer Q-bands remains unstandardized. This study establishes a machine learning informatics framework and suitability matrix for tailoring smart upconversion nanostructures across 9 biomedical photosensitizers (Chlorin e6, PPIX, ZnPc, Methylene Blue, Rose Bengal, MC540, Hypericin, ICG, and TiO2) for targeted brain cancer therapy. We constructed a dataset of 1,710 formulation-photosensitizer pairs evaluated across a published 190 UCNP host lattices dataset[2]. Five domain-physics features were engineered: Spectral Overlap Integral (J_PS), Blood-Brain Barrier (BBB) Transcytosis Size Penalty (W_BBB), Effective Atomic Number (Z_eff), and Cross-Relaxation Index (CR_idx). The machine learning regressors were benchmarked using 80/20 train/test splits with SHAP explainability. Extra Trees Regressor achieved top predictive accuracy (Test R2 = 0.9485, RMSE = 0.6842), followed by Gradient Boosting (R2 = 0.9272) and XGBoost (R2 = 0.9104). This work provides a validated nanobiotechnology blueprint for designing BBB-permeable smart functional nanomaterials, advancing precision nanomedicine for deep-seated brain tumor therapy.