A machine learning and physics-informed framework for predictive modeling of gemcitabine nanocomposite cytotoxicity.

BACKGROUND

Gemcitabine is an important chemotherapeutic agent that suffers from poor bioavailability and drug resistance. Although the nanocomposite delivery systems are highly promising, their optimization for efficient therapy is quite challenging because of complex design parameters involved.

METHODS

A dataset consisting of more than fifty gemcitabine nanocomposites is used in the present work. We developed prediction models using well-known machine learning algorithms, such as Random Forest and XGBoost, besides a new physics-informed machine learning approach that incorporated drug release and cellular uptake equations.

RESULTS

The developed models reached a high predictive accuracy in cytotoxicity and encapsulation efficiency, with R squared values higher than 0.85. Encapsulation efficiency, nanoparticle size, and zeta potential emerged as the most critical design parameters according to the feature importance analysis. The physics-informed model further improved the predictive performance for extrapolation tasks by 16% compared to conventional machine learning.

CONCLUSION

The present work sets up a strong computational framework integrating data-driven modeling with physical principles that can guide the rational design of high-efficacy gemcitabine nanocarriers and potentially accelerate their development.