From prediction to functional delivery: evaluating artificial intelligence in lipid nanoparticle design

Artificial intelligence (AI) is expanding the chemical and formulation space explored for nucleic acid lipid nanoparticles (LNPs), but accurate assay prediction does not by itself establish useful delivery in vivo . This critical narrative review examines how evidence links molecular design, complete formulation identity, and biological context to cell-resolved function, with a focus on mRNA. We distinguish tissue accumulation from functional cargo delivery, membrane damage from productive release, and target-cell coverage from panel-restricted on-target precision. Four purposively selected cases illustrate distinct inference problems: AGILE’s selected-candidate association and reformulation, Hamilton’s non-AI accumulation screen and functional follow-up, LANTERN’s benchmark using AGILE-derived data, and Wang’s delivery-classification and pKa-regression workflow with changes between screening rounds. We propose a claim-specific evidence audit and a prospective comparison whose primary outcome is the number of confirmed qualifying formulations found within a declared resource ceiling. Manufacturing, storage, and tolerability are considered where they constrain the application. Neither the audit nor the comparison design has been empirically validated. They identify testable requirements for evaluating whether AI improves a declared delivery decision.