Overcoming the prediction limitation of small datasets of mRNA lipid nanoparticles: Leveraging the power of language models and multi-task learning

Traditional screening of ionizable lipids of mRNA lipid nanoparticles (mRNA-LNPs) primarily relies on time-consuming and resource-intensive trial-and-error experiments. While machine learning approaches offer promise in predicting the behavior of mRNA-LNPs, their applications are limited by small datasets. In this study, we developed FormulationLNP, a novel multi-task learning model to predict two key properties of mRNA-LNPs: the delivery efficiency and apparent p K a . By integrating the training strategies of in-domain pre-training, multi-task learning, and data augmentation, FormulationLNP effectively addressed the small dataset challenge. FormulationLNP showed great prediction performance on both tasks, achieving the receiver operating characteristic curve (ROC-AUC) scores of 0.862 and 0.867, respectively. Furthermore, important substructures of ionizable lipids were identified to help gain insight into the relationship between lipid structure and in vivo behavior. The model demonstrated great generalization with a prediction accuracy of 0.829 on the external delivery efficiency test set and a prediction accuracy of 0.656 on the external apparent p K a test set. In conclusion, this work provides a powerful tool to predict mRNA-LNPs’ behavior, which will significantly accelerate the design and optimization of LNP delivery systems. A novel architecture named FormulationLNP was developed by combining the language model and multi-task learning to predict the in vivo delivery efficiency and apparent pKa of mRNA-lipid nanoparticles.