A Generative Adversarial Network Approach to Predict Nanoparticle Size in Microfluidics

To achieve precise control over the properties and performance
of nanoparticles (NPs) in a microfluidic setting, a profound understanding
of the influential parameters governing the NP size is crucial. This
study specifically delves into poly(lactic-co-glycolic
acid) (PLGA)-based NPs synthesized through microfluidics that have
been extensively explored as drug delivery systems (DDS). A comprehensive
database, containing more than 11 hundred data points, is curated
through an extensive literature review, identifying potential effective
features. Initially, we employed a tabular generative adversarial
network (TGAN) to enhance data sets, increasing the reliability of
the obtained results and elevating prediction accuracy. Subsequently,
NP size prediction was performed using different machine learning
(ML) techniques including decision tree (DT), random forest (RF),
deep neural networks (DNN), linear regression (LR), support vector
regression (SVR), and gradient boosting (GB). Among these ensembles,
DT emerges as the most accurate algorithm, yielding an average prediction
error of 8%. Further simulations underscore the pivotal role of the
synthesis method, poly(vinyl alcohol) (PVA) concentration, and lactide-to-glycolide
(LA/GA) ratio of PLGA copolymers as the primary determinants influencing
NP size.