Machine Learning Accelerates Discovery of High-Performance Corrole Photosensitizers for Optical Imaging Diagnosis and Photodynamic Therapeutics of Nasopharyngeal Carcinoma.

This study centers on corrole, an emerging photosensitizer with great application potential, and innovatively develops an intelligent machine learning-based screening strategy. Through integrating molecular descriptor generation, feature engineering, multiple predictive models, and SHapley Additive exPlanations (SHAP)-based feature analysis, we have predicted two key performances of corrole photosensitizers, involving absorption peak wavelength and singlet-triplet intersystem crossing rate (kISC). Among 10 tested models, XGBoost outperforms other models. After introducing descriptors such as electronic structure and excited-state characteristics and conducting dimensionality reduction, its coefficient of determination is up to 0.87 for kISC prediction. SHAP analysis clarifies core design principles such as flat molecular structure, HOMO localization/LUMO delocalization and low surface electrostatic potential. The screened corrole with peripheral pyridyls, central nonmetallic P, and hydroxyls has excellent red light/NIR optical performance. Upon light irradiations, it exhibits a robust capacity for generating reactive oxygen species and enables the complete inactivation of cancer cells in vitro and in vivo. By leveraging fluorescence lifetime imaging, this corrole enables effective discrimination between normal cells and cancer cells. The intelligent screening strategy offers a novel paradigm for the efficient development of high-performance photosensitizers with distinct clinical translation potential.