Delivering an mRNA drug is not just about reaching the address. The package has to be handed to the recipient inside.

Lipid nanoparticles (LNPs) are the delivery vans; ionizable lipids are the components that determine how well the van performs. The persistent failure mode: mRNA gets stuck in the endosome, the cell's sorting depot, and never reaches the cytoplasm where it must work. Finding materials that release cargo efficiently has traditionally meant synthesizing and testing molecule after molecule.

A different order of operations

A team led by Prof. Yan Hao and Prof. Kang Feiyu at Tsinghua Shenzhen International Graduate School flipped the sequence: use experimental data to train a model, then let the model search a much larger space.

Step one: modularly combine 40 amine heads with 13 degradable tails, synthesize 520 lipids, and measure actual mRNA expression after delivery in cell assays.

Step two: feed that data to an XGBoost model and expand the virtual screening space to 10,920 lipid structures.

Step three: synthesize and test the model's top picks to confirm.

The result

Among the screened candidates, a lipid designated PA97MD101 stood out: in mouse studies it produced roughly four times the hepatic mRNA functional expression of SM-102, a commonly used comparator.

The detail that matters: PA97MD101 and SM-102 formed particles with broadly similar overall liver accumulation, yet the former drove markedly stronger functional expression.

That contrast reframes the problem. Delivery performance is not "how many particles arrive at the liver" but "how much mRNA gets released inside cells." Many candidates fail not at the door but at the exit.

Why release is better

The team probed the mechanism with cell imaging, membrane interaction assays and molecular simulation. The evidence points to membrane perturbation behavior consistent with an acidic environment, lower lysosome colocalization, and molecular geometry plus lipid-RNA interactions that favor intracellular release.

Multiple complementary lines of evidence are more convincing than a single metric.

Beyond one cargo

The advantage is not limited to a reporter gene. PA97MD101 also delivered erythropoietin mRNA and CRISPR/Cas9 payloads, achieving TTR expression knockdown in mice.

Do not over-read it

One sentence in the paper deserves emphasis: the model learned lipid transfection activity in vitro, not whether a lipid targets the liver precisely. Actual hepatic delivery performance had to be validated by independent animal experiments.

AI here is a search-space reducer, not an oracle that replaces experiments.

The value of this study is that it closed the loop: small-scale experiments, model-expanded search, synthesis and validation, mechanistic explanation. Four steps completed is what makes the conclusion stand. For materials and formulation R&D, that methodology transfers better than any single molecule.