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

Area of Science:

• Nanomedicine
• Computational Chemistry
• Biomaterials Science

Background:

• Curcumin's therapeutic potential is hindered by low bioavailability and toxicity.
• Nano-encapsulation offers a solution, but optimizing nanocarrier safety is complex.
• A multidimensional design space complicates rational optimization of nanocarrier biosafety.

Purpose of the Study:

• To develop an interpretable Physics-Informed Machine Learning (PIML) framework for predicting and optimizing curcumin nanocarrier cytotoxicity.
• To integrate experimental data with DLVO stability theory and drug-release kinetics.
• To establish practical guidelines for fabricating safer curcumin-based nanotherapeutics.

Main Methods:

• Developed a PIML framework integrating experimental data from 75 curcumin nanocarriers.
• Incorporated DLVO stability theory and drug-release kinetics into the PIML model.
• Utilized XGBoost and SHAP analysis for model evaluation and feature importance.

Main Results:

• The PIML model achieved high accuracy (R² = 0.86) in predicting cytotoxicity.
• Key factors for reduced cytotoxicity include negative zeta potential (-30 to -40 mV), chitosan coatings, and particle sizes of 150-250 nm.
• Multi-objective Bayesian optimization identified a design space reducing toxicity by ~82% while maintaining ~70% loading efficiency.

Conclusions:

• The developed PIML framework provides a generalizable computational methodology for nanocarrier design.
• This approach translates complex design interactions into actionable guidelines for safer nanotherapeutics.
• The study facilitates the development of improved curcumin-based nanomedicines with reduced toxicity and preserved efficacy.