Abstract
Lipid nanoparticles (LNPs) have gained significant attention thanks to their ability to encapsulate and deliver mRNA. Exploring a variety of lipid compositions and different preparation processes is essential for a better understanding of the mRNA encapsulation and the LNPs assembly. In this study, we investigated the development of mRNA-LNPs through microfluidic PFVs production followed by a postencapsulation approach. A library of preformed vesicles (PFVs) was produced by microfluidics using a full factorial design by varying 4 formulation and process parameters: main lipid type, sterol type, flow rate ratio, and chip design. The Size and polydispersity index (PDI) of PFVs were measured and compiled into a data set. A subset of formulations was subsequently postencapsulated with mRNA, after which the size, PDI, and EE% of the resulting mRNA-LNPs were quantified to generate a labeled data set. The results showed the effects of chip design and lipid composition on the size and PDI of PFVs, with smaller PFVs obtained with the chip design that provides a higher mixing efficiency. Postencapsulated formulations showed consistent increases in nanoparticle size and decreases in PDI values, compared to the corresponding PFVs. An XGBoost model was trained and validated on the labeled data set for predicting size and EE%, and was further optimized through semisupervised learning by incorporating pseudolabeled data derived from the PFVs data set. The model demonstrated an ability to predict the size and the EE% of LNPs based on the composition, the process parameters, and the physicochemical properties of PFVs. The use of microfluidics and machine learning allowed the obtaining of relevant information with limited resources and time. Integrating machine learning and advanced data analysis in nanomedicine research could unveil its full potential and accelerate the development of innovative formulation strategies.