Optimising the production of PLGA nanoparticles by combining design of experiment and machine learning.
Journal:
International journal of pharmaceutics
PMID:
39491656
Abstract
Poly(lactic-co-glycolic acid) (PLGA) is a widely used biodegradable polymer in drug delivery and nanoparticle (NP) formulation due to its controlled drug release properties and safety profiles. Among the methods available for NP production, nanoprecipitation is distinguished by its simplicity and scalability. However, it requires careful optimisation to achieve the desired NP characteristics, making the process potentially lengthy and costly. This study aimed to assess and compare the predictive performance of Design of Experiments (DOE) and Machine Learning (ML) models for the optimisation of PLGA nanoparticle size and zeta potential produced by nanoprecipitation. Various ML methods were employed to predict particle size, with Extreme Gradient Boosting (XGBoost) identified as the best performing. The key finding is that integrating ML with DOE provides deeper insights into the dataset than either method alone. While ML outperformed DOE in predictive performance, as evidenced by lower root mean squared error values and higher coefficients of determination, both methods struggled to accurately predict zeta potential, generating models with high errors. However, ML proved more effective in identifying the parameters that most significantly influence NP size, even with a smaller DOE dataset. Combining DOE datasets with ML for parameter importance was particularly advantageous in situations where data is limited, offering superior predictive power and the potential to streamline experimental design and optimisation. These results suggest that the synergistic use of ML and DOE can lead to more robust feature analysis and improved optimisation outcomes, particularly for NP size. This integrated approach can enhance the accuracy of predictions and supports more efficient experimental design, streamlining nanoparticle production processes, especially under resource-constrained conditions.