Machine learning-driven exosome-mimetic lipid nanoparticles for tumor-specific targeting

Exosome-mimetic lipid nanoparticles (ENPs) are a promising alternative to PEGylated lipid nanoparticles (LNPs) for targeted cancer therapy, offering improved biocompatibility and reduced immune clearance. However, the rational design of these biomimetic particles is challenging due to complex lipid composition requirements. We developed a hybrid algorithm to optimize exosome-mimetic formulations by predicting key nanoparticle properties (size, zeta potential, and polydispersity index). The algorithm was trained on an expanded dataset of 17,800 lipid compositions generated by augmenting experimental and publicly available data using the LipidGAN generative model, incorporating physicochemical modeling and feature extraction. It identified optimal formulations, which were validated in vitro across three cancer cell lines (HeLa, H1975, and MCF-7). Cytotoxicity assays confirmed minimal toxicity (cell viability > 90%), and uptake studies demonstrated efficient, cell-type-specific internalization (91 ~ 95%). These results highlight the potential of artificial intelligence (AI)-driven lipid design to emulate the functionality of natural exosomes and advance the development of safe, effective, and personalized cancer nanomedicines.