Physics informed machine learning for predictive toxicology and optimization of curcumin nanocarriers

Curcumin's clinical utility is limited by poor bioavailability and dose-dependent toxicity. Although nano-encapsulation can address these shortcomings, rationally optimizing nanocarrier biosafety remains challenging due to the highly multidimensional design space. Here, we develop an interpretable Physics-Informed Machine Learning (PIML) framework that integrates experimental data from 75 curcumin nanocarriers with DLVO stability theory and drug-release kinetics to predict and optimize cytotoxicity. Among the evaluated models, XGBoost attained the highest statistical performance (R² = 0.89), although the PIML model provided physically coherent predictions with similar accuracy (R² = 0.86). SHAP research indicated a moderate negative zeta potential (-30 to -40 mV), chitosan-based coatings, and particle sizes of 150-250 nm as the principal factors contributing to decreased cytotoxicity. Multi-objective Bayesian optimization delineated a Pareto-optimal design space, facilitating approximately 82% toxicity reduction compared to free curcumin, while preserving around 70% loading efficiency. The study develops a proven, generalizable computational methodology that converts intricate nanocarrier design interactions into practical guidelines for the fabrication of safer curcumin-based nanotherapeutics.