Machine Learning Elucidates Design Features of Lipid Nanoparticles for Cell Type-Preferential Transfection

Introduction: Lipid nanoparticles (LNPs) carrying nucleic acid cargos are potent delivery vehicles for gene therapy. LNPs are generally composed of five components including an ionizable lipid, helper lipid, cholesterol, PEGylated lipid, and 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 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 chemistry and the molar ratios and percentages of each nucleic acid and lipid component. Here, we developed a machine learning (ML) pipeline to elucidate underlying design rules governing cell type-preferential LNP transfection in vitro. Using this pipeline, we aimed to develop a quantitative understanding of LNP design rules and uncover critical features for optimization of transfection efficiency of the LNPs.

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 (murine melanoma), HepG2 (human hepatocellular carcinoma), HEK293 (human embryonic kidney), PC3 (human prostatic adenocarcinoma), ARPE-19 (human retinal pigment epithelia), and N2a (murine neuroblast) cell lines. Transfection efficiency was recorded as luciferase expression at 48 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, neural networks, and decision tree models implemented from the sklearn, XGboost, and LightGBM python libraries. Randomized grid search for hyperparameters and nested cross-validation was used to identify the optimized model architecture for each algorithm. The optimized algorithm with the lowest mean absolute error (MAE) when evaluated with a hold-out validation set was selected as the best performing model. SHapley Additive exPlanations (SHAP) were used to give insight into the machine learning model predictions. [2]

Results, Conclusions, and Discussions: LightGBM (LGBM), a decision-tree based model, exhibited the lowest percent error between predicted and experimental transfection results of a hold-out validation set compared to other tested models for B16F10 cells (Fig. 1A). Predictions displayed strong linear correlations with experimental transfection, validating the model’s predictive ability and strong understanding of key design rules impacting LNP transfection (Fig. 1B). High performing models were similarly optimized for all 6 cell types (Fig. 1C). SHAP quantifies key formulation design rules governing cell type-preferential LNP transfection efficiency in the B16F10 cell type (Fig. 1D). Visualizing the location of feature values of the most important features in 2-dimensional space by t-distributed stochastic neighbors embedding dimensional reduction provides deeper insight into new design spaces for improved LNP formulations (Fig. 1E). SHAP analysis was repeated for all 6 cell types. Notably, we identified an understudied, yet critical helper lipid feature that impacts cell-type tropism, namely helper lipid hydrophobicity (“cLogP”). We are the first to report that cLogP plays a crucial role in LNPs transfecting B16F10 and PC3, a moderately impactful role in HEK293, HepG2, ARPE19, and is a bystander variable in N2a (Fig. 1F). Furthermore, while ML confirmed known LNP design heuristics such as moderately low PEGylated lipid content (“PEG/(Chol+PEG)%”), moderately high molar percentage of ionizable lipid and helper lipid i.e. (“(IL+HL)%”), high N/P ratios, and cationic helper lipids, our pipeline is the first to formally quantify the contributions of each features in a cell type dependent manner. We also uncovered distinct and mutually exclusive sets of formulation parameters for each cell type to optimize cell type-preferential transfection (Fig. 1G).

In conclusion, we have established an ML-based formal quantitative approach to characterize the contributions of specific LNP design features and identified key design features and optimal values governing LNP-preferential transfection in different cell types. Our proof-of-concept pipeline is now ideally positioned to investigate expanded parameter spaces revealed by new lipid chemistries, the addition of more lipid components (such as SORT lipids), and new payload types including mRNA and other RNA therapeutics in future studies.

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

References:
(1) Zhu, Y., et al. Nat. Commun, 2022; 13(4282), 1–14.
(2) Lundberg, S. M. & Lee, S.-I. Neural Inf. Process. Syst. 2017.