A physics-informed machine learning framework for predicting and mitigating doxorubicin nanocarrier toxicity in...

Area of Science:

• Biomedical Engineering
• Computational Chemistry
• Nanotechnology

Background:

• Doxorubicin (DOX) efficacy is limited by dose-dependent systemic toxicity, particularly cardiotoxicity.
• Nanocarrier systems offer solutions but optimizing their complex physicochemical properties for biological outcomes is challenging.

Purpose of the Study:

• To develop a hybrid computational framework integrating classical and physics-informed machine learning (PIML) to predict and optimize the cytotoxicity of DOX-loaded nanocarriers toward normal cells.
• To identify key physicochemical properties governing nanocarrier safety and establish an optimal design space for reduced toxicity.

Main Methods:

• Compiled an extensive dataset of 77 nanocomposite systems with physicochemical and biological data.
• Trained and compared various machine learning (ML) models, including a Physics-Informed Neural Network (PINN) incorporating drug release kinetics, colloidal stability, and diffusion.
• Utilized SHAP analysis for feature importance and Bayesian optimization for design space exploration.

Main Results:

• The PINN model demonstrated superior predictive capability (R² = 0.89, RMSE = 0.14) over conventional ML methods.
• Zeta potential and particle size were identified as the most critical factors influencing cytotoxicity.
• An optimal design space was determined: sizes of 120-150 nm, zeta potentials of -25 to -35 mV, loading efficiency of 5-10%, and encapsulation efficiency >85%.

Conclusions:

• The proposed PIML framework provides a robust and interpretable method for rational nanocarrier design.
• This approach significantly enhances the development of safer chemotherapeutic delivery systems by minimizing empirical optimization.
• Experimental validation confirmed the model's accuracy, with prediction errors below 3%.