Data-Driven Development of Predictive Models for Sustained Drug Release.

Mathematical modeling of drug release can aid in the design and development of sustained delivery systems, but the parameter estimation of such models is challenging owing to the nonlinear mathematical structure and complexity and interdependency of the physical processes considered. Highly parameterized models often lead to overfitting, strong parameter correlations, and as a consequence, inaccurate model predictions for systems not explicitly part of the fitting database. Here, we show that an efficient stochastic optimization algorithm can be used not only to find robust estimates of global minima to such complex problems but also to generate metadata that allow quantitative evaluation of parameter sensitivity and correlation, which can be used for further model refinement and development. A practical methodology is described through the analysis of a predictive drug release model on published experimental data sets. The model is then used to design a zeroth-order release profile in an experimental system consisting of an antibody fragment in a poly(lactic-co-glycolic acid) solvent depot, which is validated experimentally. This approach allows rational decision-making when developing new models, selecting models for a specific application, or designing formulations for experimental trials.