Ionizable lipids are a key component of lipid nanoparticles (LNPs), a leading nonviral messenger RNA (mRNA) delivery technology. Here, we introduce Lipid Optimization using Neural networks (LiON), a deep-learning strategy for designing ionizable lipids. To train LiON, we generated a dataset of over 9,000 lipid nanoparticle activity measurements and fed this data into a directed message-passing neural network to predict nucleic acid delivery across diverse lipid structures.
Lipid optimization using LiON successfully predicted RNA delivery in both in vitro and in vivo held-out test sets and extrapolated to structures distinct from the training set. Next, we evaluated 1.6 million lipids in silico and identified two structures, FO-32 and FO-35, which demonstrated state-of-the-art local mRNA delivery to mouse muscle and nasal mucosa. FO-32 also matched the state of the art for nebulized mRNA delivery to the mouse lung, while both FO-32 and FO-35 efficiently delivered mRNA to ferret lungs—representing the first published example of mRNA delivery to ferret conducting airways.
Overall, this work highlights the potential of deep learning to enhance nanoparticle delivery and introduces LNPs with promising activity for pulmonary gene therapy.