Mechanism-informed machine learning for interpretable drug release prediction and virtual formulation intervention in polymeric long-acting injectables.

The correct time-course forecasting of the release of a drug contained in a drug-polymer matrix is difficult because drug dissolution is a complex process involving all formulation parameters, which are interdependent and have nonlinear effects. Although the predictive accuracy of machine learning (ML) models is high, classical methods generally cannot differentiate between predictive features that are only correlated and actually useful ones that are relevant in a formulation design. This work presents a framework for mechanism-informed ML that brings together pharmaceutical knowledge, a knowledge-driven graph, explainability of artificial intelligence (AI), intervention-aligned effect estimation, and computer-assisted formulation experiments. The XGBoost model, which used data on physicochemical properties, composition, and time, showed good results (R2 = 0.9774, RMSE = 0.0491; MAE = 0.0339). Feature analysis showed that correlation and SHAP importance agree strongly (r = 0.761), while both correlation and SHAP importance for effects under intervention are very poor (r = 0.132 and 0.021), indicating that predictive and actionable variables differ substantially. This new platform facilitates evaluating formulation changes thoroughly and can help decide which experiments to prioritize, resulting in LAI development of polymeric material being transparent and based on mechanism.