A Machine Learning Assisted Tool and Numerical Model for Analyzing Lipid Nanoparticles

The transfection potency and biological fate of gene-loaded lipid nanoparticles (LNPs) are often determined by their morphological and physicochemical properties. Cryogenic-electron microscopy (cryo-EM) remains the most effective tool to analyze LNP morphology and internal structures in their native state, but analysis of cryo-EM micrographs is time-consuming and inefficient due to the diversity in size, shape, and structure of LNPs. In this study, we developed the Lipid Nanoparticle Morphology and Object Detector (LNP-MOD) pipeline. We adopted a modular design by using the You Only Look Once (YOLO) model for object detection and the Segmentation Anything model 2 (SAM 2) for LNP compartmental segmentation. We trained the model and demonstrated that LNP-MOD can effectively identify and segment different classes of LNPs and their corresponding internal structures with ∼80% accuracy. We further compared the image analysis data with mathematical modeling of LNPs containing water and mRNA (Liposomal LNPs and Bleb LNPs) according to the phase preferences of the lipids, and showed correspondence between LNP-MOD output, modeling results, and experimental data. Our approach of combining single-particle cryo-EM imaging with LNP-MOD is complementary to other analytical techniques. It allows for rapid identification and segmentation of a variety of LNP-nucleic acid morphologies and presents a powerful tool to inform the design of next-generation LNPs.