Predictive Modeling of Drug Release from Liposomal Nano-Anticancer Drugs Using Advanced Machine Learning Techniques and TreeSHAP for Interpretability

Background: Accurately predicting drug release from liposomal nano-anticancer drugs is crucial for optimizing therapeutic efficacy and patient compliance. Traditional release kinetics models often oversimplify the complex mechanisms of drug release, necessitating the application of advanced machine learning (ML) approaches for improved predictive performance and deeper insights. However, existing models lack generalizability and fail to capture the intricate interplay of formulation parameters and environmental factors, highlighting a critical research gap. Methods: This study utilized nine ML models to predict fractional drug release. Input features included molecular descriptors and experimental conditions. Nested cross-validation ensured robust model evaluation, while SHAP provided interpretability for understanding feature importance. The NGBRegressor model, integrated with TreeExplainer, was identified as the best-performing model. Results: The NGBRegressor achieved the lowest mean absolute error of 5.11, outperforming other models in handling nonlinearity and uncertainty. SHAP analysis revealed key features influencing drug release, offering actionable insights into formulation parameters. The study also demonstrated that ML approaches provide superior predictive accuracy compared to traditional mathematical models, highlighting their ability to model complex, nonlinear relationships. Conclusion: This study underscores the potential of ML models, particularly NGBRegressor, in predicting drug release from liposomal formulations. By combining robust predictive performance with interpretability tools, the findings provide a data-driven framework for the design of liposomal drug delivery systems.