Genetic algorithm optimization of tree-based models to predict cargo- and carrier-related factors affecting drug release from liposomes

Abstract

Liposomal drug delivery systems have shown promising potential to improve drug delivery in several aspects. Precise in vitro characterization of formulated liposomes is vital to achieve proper in vivo function. In particular, in vitro release testing of liposomes offers crucial insights for predicting in vivo drug release behavior, guiding the development of more effective liposomal formulations. Herein, random forest (RF)- and XGboost-genetic algorithms were implemented to establish a model to predict release profiles of liposomes based on critical characteristics of carriers, cargoes, and release media. The models were trained with a dataset consisting of release parameters of 203 different liposomal formulations extracted from the literature, divided into three categories: all cargoes, therapeutics with log P < 1, and log P > 1. The SHapley additive exPlanations (SHAP) approach was used to determine how data contributes to the model prediction. A mean squared error of 1.648 highlighted the capability of the genetic algorithm to optimize the hyperparameters of the tree-based models. Furthermore, the developed models demonstrated high accuracy as evidenced by loss functions and statistical metrics such as root mean squared error, mean absolute error, and R-squared. Our results shed light on the distinct influence of phase transition temperature, drug concentration, log P, water solubility, molecular weight, vesicle size, cholesterol to phospholipid molar ratio, and surfactant concentration in the release medium on drug release at different phases of the release profiles. The predictive models developed in this study can facilitate the design of an ideal liposomal formulation for any desired release profile.