ScholarWorks@Y-Scholar Hub

초록

Exosome-mimetic lipid nanoparticles (ENPs) represent a promising alternative to PEGylated lipid nanoparticles (LNPs) and conventional chemotherapies, which suffer from rapid systemic clearance and off-target toxicity. We developed a regressionbased hybrid algorithm to rationally design ENPs by formulating them from five critical material attributes (CMAs)—cholesterol (CHOL), sphingomyelin (SM), phosphatidylcholine (PC), phosphatidylethanolamine (PE), and phosphatidylserine (PS)—that represent the dominant lipid composition of natural exosomes. To enhance predictive performance, we employed a lipid generative adversarial network (LipidGAN) to generate 17,800 synthetic ENP compositions based on experimentally validated and publicly available datasets. This augmented dataset improved property prediction accuracy (R² > 0.95) for three physicochemical traits. The hybrid algorithm then ranked candidate formulations and recommended the top ten optimal ENP compositions for validation. These candidates were experimentally tested via three critical quality attributes (CQAs)—particle size, zeta potential, and polydispersity index (PDI)—followed by in vitro assessment across three cancer cell lines (HeLa, H1975, MCF-7). The resulting ENPs exhibited high biocompatibility (>90% viability) and strong cellular uptake (91–95%), with PC identified as the key uptake driver. This hybrid algorithm provides a PEG-free, rational design platform for precision nanomedicine.