Self-Driving Lab Finds Brominated Lipids Outperform Moderna's COVID Vaccine Lipid in mRNA Delivery

Drug delivery for mRNA therapeutics has a bottleneck. Only three lipid nanoparticle (LNP) formulations have received FDA approval despite years of intensive research, and finding new ones that safely and efficiently carry mRNA into target cells has remained difficult. The experimental space is vast, the data is sparse, and traditional trial-and-error approaches are too slow to map it systematically. A University of Toronto team has now demonstrated an AI-robotics platform that largely automates the discovery process - and in doing so, found a class of lipid that no one had previously linked to mRNA delivery.

The system, called LUMI-lab (Large-scale Unsupervised Modeling followed by Iterative experiments), was published in Cell by researchers at the Leslie Dan Faculty of Pharmacy, University of Toronto. Across ten active-learning cycles, LUMI-lab synthesized and tested more than 1,700 new lipid nanoparticles. Among the top-performing candidates, one chemical feature dominated: brominated lipid tails, which accounted for only 8 percent of the chemical library LUMI-lab drew from but represented over half of the highest-performing compounds.

Why autonomous discovery matters for mRNA delivery

mRNA therapeutics are among the fastest-growing classes of drugs. They offer a way to instruct cells to produce therapeutic proteins without altering DNA, which has obvious implications for vaccines, cancer immunotherapy, and protein-replacement disorders. But the mRNA molecule itself is fragile and cannot reach target cells without protection. LNPs - spherical lipid particles that encapsulate mRNA and fuse with cell membranes - provide that protection, but only a tiny fraction of candidate lipid structures actually work well enough to be therapeutically useful.

Machine learning approaches to LNP discovery face a particular challenge: the historical datasets needed to train models are small. Unlike drug databases that contain millions of entries accumulated over decades, mRNA-LNP data is sparse and recent. The LUMI-lab team addressed this by using a foundation model pretrained on more than 28 million molecular structures of all types before being applied to the specific task of predicting LNP performance. The broad pretraining gave the model a rich chemical vocabulary that it could then apply to the relatively small mRNA delivery dataset.

Ten learning cycles, one unexpected discovery

Active learning is an iterative process: the model predicts which compounds are most likely to perform well, the robotic synthesis system makes them, the results come back, the model updates, and the cycle repeats. LUMI-lab ran this loop ten times, building on each round of experimental results.

The bromination finding emerged from this process without any researcher hypothesis guiding it. Brominated lipid tails - a chemical modification in which bromine atoms replace hydrogen atoms along the lipid's carbon chain - had not previously been connected to mRNA delivery performance. LUMI-lab identified the structural feature as predictive of high transfection efficiency (the ability to get mRNA into cells) and the subsequent experiments confirmed it.

"The key advance of this AI-driven system is that it independently identified bromination as an important, meaningful design feature without prior hypothesis or researchers telling it to look for it first," said Bowen Li, GSK Chair in Pharmaceutics and Drug Delivery at the Leslie Dan Faculty of Pharmacy.

Preclinical performance and safety comparison

Tested in preclinical models, some of the brominated LNP candidates outperformed SM-102, the lipid used in Moderna's COVID-19 mRNA vaccine - a compound that has been through rigorous clinical testing and is one of the best-characterized delivery lipids available. The brominated candidates also showed safety profiles similar to established clinical lipids in the same preclinical systems, supporting the idea that they warrant further development.

Preclinical results in cell and animal models do not guarantee human performance. The compounds that outperformed SM-102 in laboratory tests will require extensive additional safety and efficacy evaluation before any clinical application is possible. The researchers position the finding as evidence that the design space for LNPs is substantially larger than the approved set suggests, and that AI-driven exploration can navigate it more systematically than human hypothesis-generation alone.

Next steps for LUMI-lab

"Next, we're expanding LUMI-lab to optimize multiple clinically relevant properties at once - not just delivery potency but also safety, tolerability, and tissue selectivity," said Li. Multi-property optimization is substantially harder than single-metric optimization, and whether active learning platforms can balance competing objectives across a large chemical space remains an open question. The team plans to continue publishing results as the platform evolves.