The rapid growth of lipid nanoparticle (LNP)-based RNA therapeutics demands predictive tools to accelerate formulation and lipid design, yet development remains limited by complex delivery mechanisms and scarce high-quality data. We investigated few-shot meta-learning (FSL) as a strategy for early stage, data-limited LNP development using a published data set. Meta-learning tasks were constructed from data provenance and formulation conditions, and several FSL methods were benchmarked against supervised baselines using fingerprint- and graph-based representations. In a stringent extrapolation setting, all siRNA data were excluded from meta-training and reserved for testing. Under this protocol, model-agnostic meta-learning (MAML) outperformed both supervised and transfer-learning baselines, achieving an average R(2) of 0.38 +/- 0.049 on the siRNA holdout task, whereas non-meta-learning models performed near zero. In retrospective active-learning simulations and experimental validation with 15 newly synthesized ionizable lipids across multiple cell lines and RNA cargos, MAML consistently surpassed random forest (RF), supporting FSL as a promising framework for data-scarce RNA delivery design tasks.
Keywords
