Interpretable Two-Stage Machine Learning for Early and Full Drug Release Prediction in PLGA Microspheres.

Area of Science:

• Materials Science
• Pharmaceutical Sciences
• Computational Biology

Background:

• Poly(lactic-co-glycolic acid) (PLGA) microspheres are crucial for long-acting injectable (LAI) formulations due to biocompatibility and controlled degradation.
• Current optimization of PLGA microspheres relies heavily on time-intensive in vitro testing.

Purpose of the Study:

• To develop a predictive machine learning model for estimating drug release profiles from PLGA microspheres.
• To accelerate the early-stage development and optimization of LAI formulations.

Main Methods:

• A classification model was developed to identify slow-release behavior (≤20% release within 3 days) using a dataset of 321 release profiles from 89 drugs.
• A regression model integrated early-release probability to predict cumulative drug release over time, with XGBoost showing optimal performance.
• SHapley Additive exPlanations (SHAP) were used to identify key influencing factors on drug release.

Main Results:

• The XGBoost model achieved a low Mean Absolute Error (MAE) of 0.126 and a high Pearson correlation coefficient (r) of 0.831 on the training dataset.
• SHAP analysis identified drug and polymer molecular weight, predicted slow-release probability, and polymer concentration as significant factors affecting release.
• The framework demonstrated low MAE values (0.096 and 0.068) when validated with external datasets for olaparib and semaglutide microspheres.

Conclusions:

• The developed machine learning framework effectively predicts in vitro drug release from PLGA microspheres.
• This predictive capability can significantly streamline and enhance the optimization process for PLGA-based LAI formulations.