(G-262) Machine Learning Elucidates Design Features of Lipid Nanoparticles for Gene Delivery

Introduction:: The development of lipid nanoparticles (LNPs) carrying plasmid DNA or mRNA for gene therapy is an exponentially growing field. LNPs are generally composed of five components 1) ionizable lipid, 2) helper lipid, 3) cholesterol, 4) PEGylated lipid, and 5) nucleic acid payload. Each component can be engineered with varying chemical features and combined in different molar ratios and molar percentages to create LNPs with distinct properties. However, important design parameters of LNP formulations to optimize transfection in vitro and in vivo remain poorly understood. We have previously established a multi-step high-throughput screening approach to efficiently select plasmid DNA-carrying LNP formulations from a 1080-LNP library with high transfection efficiency in vitro and in vivo.¹ Specifically, this library focuses on LNP optimization by varying the helper lipid used and the molar ratios and percentages of each nucleic acid and lipid component. Here, we developed a machine learning pipeline to analyze the library screening results and elucidate underlying design rules governing LNP transfection in vitro. These insights enable researchers to better understand and rationally design LNPs for gene delivery.

Materials and Methods:: Data was collected through in vitro transfection screening of 1080 distinct LNP formulations prepared at varying molar ratios and percentages of luciferase plasmid DNA, ionizable lipid (Dlin-MC3-DMA), helper lipid (DSPC, DOPE, DOTAP, DDAB, 18PG, or 14PA),¹ cholesterol, and PEGylated lipid (DMG-PEG 2 kDa) in B16F10 melanoma cells. Transfection efficiency was recorded as luciferase expression at 24 h after transfection. The machine learning training data consisted of molar ratios and percentages of each LNP component and chemical features of the helper lipid used. The machine learning pipeline evaluated multiple regression algorithms including linear regression, nearest neighbor, and decision tree models implemented from the sklearn, XGboost, and LGBM python libraries. Randomized grid search was used to optimize hyperparameters. Nested cross-validation was used to evaluate the best performing algorithm and corresponding model architecture. SHAPley Additive Explanations were used to give insight into the machine learning model predictions.²


Results, Conclusions, and Discussions:: Results and Discussion: Decision-tree based algorithms (XGBoost, LightGBM, and Random Forest) exhibited the lowest MAE between predicted and experimental transfection results of a hold-out validation set (Fig. 1A). XGBoost model predictions of hold-out LNP formulations display a strong linear correlation with the experimental transfection, validating the model’s predictive ability and strong understanding of key design rules impacting LNP transfection (Fig. 1B). SHAPley Explanations uncover key formulation design rules governing LNP transfection efficiency (Fig. 1C). Formulations with 1) moderately low ratios of cholesterol to DMG-PEG, 2) moderately high molar percentage of Dlin-MC3-DMA and helper lipid, and 3) high N/P ratios are associated with increased transfection efficiency. Interestingly, no obvious association between molar ratios of Dlin-MC3-DMA to the helper lipid and transfection can be identified, indicating more complex relationships may govern this parameters effect on transfection efficiency. Additionally, certain helper lipid chemical features are associated with increasing transfection efficiency such as high cLogP, high cTPSA, and higher numbers of positively charged centers. Overall, for this library of LNP formulations the molar ratios and percentages of each component played the most important role in defining LNPs transfection efficiency when compared to helper lipid used.


Conclusions: Decision Tree-based machine learning algorithms can be used to accurately predict the in vitro transfection efficiency of LNPs despite relatively small training set size of 1080 formulations. Importantly, the algorithm can be used to elucidate key design parameters for LNPs with high transfection efficiency and provide a platform for more rigorous analysis of LNP formulation libraries. In the future, the addition of more training data from more diverse parameter spaces will continue to improve the accuracy and reliability of this model to precisely understand the complex mechanisms dictating LNP transfection.


Acknowledgements (Optional): : This study is partially supported by a grant U01AI155313 from the National Institute of Allergy and Infectious Diseases.

References (Optional): : ¹Zhu, Y., et al. Nat. Commun, 2022; 13(4282), 1–14.

²Lundberg, S. M. & Lee, S.-I. Neural Inf. Process. Syst. 2017.