IL-FormAI: A machine learning-guided framework for screening mRNA-LNP formulations toward improved in vivo expression

Abstract

The mRNA delivery efficiency of lipid nanoparticles (LNPs) is governed by the chemical structure of ionizable lipid (IL) and formulation. Appropriate formulation optimization may help realize the delivery potential of IL, but current formulation screening methods for newly designed ILs still rely on wet-lab experiments by trial and error. Here, we present IL-FormAI, a machine learning-guided formulation optimization framework that rapidly generates candidate LNP formulations for ionizable lipids with diverse chemical structures. Briefly, we first set up an expanding database containing IL structures, LNP formulations and their corresponding in vivo mRNA expression efficiency. Then, a graph neural network was employed to encode the molecular structures of ILs into machine-readable feature vectors. These features, together with formulation parameters, were used as input features to construct IL-FormAI with mRNA expression efficiency as the prediction target. Through the analysis of complex nonlinear relationships between mRNA expression and LNP formulations, an iterative screening optimization framework was further introduced in IL-FormAI to achieve high accuracy and enhanced interpretability for LNP formulation from limited data. Experiments demonstrated that even for ILs with substantially different chemical structures, LNPs prepared according to IL-FormAI-generated formulations exhibited significantly enhanced in vivo delivery efficiency, compared with those prepared using benchmark formulation. In summary, IL-FormAI provides a rapid and effective method for the formulation optimization of newly designed ILs.