Poster X4 - Machine Learning-Guided Iterative Optimization Pipeline for mRNA Lipid Nanoparticles

Introduction: Lipid nanoparticles (LNPs) have emerged as promising non-viral gene delivery vectors. However, they require extensive optimization across a vast and complex design space – a task where manually-directed grid-search approaches have proven inefficient. Here, we build and deploy a machine learning (ML) guided iterative screening platform to accelerate the optimization of LNP composition. From a base library screening, we select, train, and refine high performing ML models that predict transfection outcomes as a function of LNP design features. Then, based on these predictions, we leverage Bayesian optimization to search for de novo test formulations, focusing on areas with the greatest transfection potential. Data obtained from each screening iteration is used for model retraining, establishing an efficient data-driven experimental loop. Our in silico guided approach improves the efficiency of LNP formulation search for optimized transfection outcomes.

Materials and

Methods: B lymphocytes, a canonically hard-to-transfect cell type, was chosen as the target for ML-guided optimization. To generate a diverse base library for initial ML training, we used DLin-MC3-DMA as the ionizable lipid, cholesterol, DMG-PEG2000, and one of six helper phospholipids with varying charge (DOTAP, DDAB, DOPE, DSPC, 14PA, and 18PG). We varied four parameters: (1) combined molar percentage of DLin-MC3-DMA and helper lipid; (2) weight ratio of cholesterol to DMG-PEG2000; (3) weight ratio of DLin-MC3-DMA to helper lipid; and (4) molar ratio of amine and phosphate. In total, 486 LNP formulations were screened for in vitro transfection efficiency in Ramos B cells using luciferase reporter mRNA. The sci-kit learn python package (1.2.2) was used to conduct model training and prediction on a panel of tree-based, linear, and partial least squares regression models. Helper lipids were parameterized using select chemical features, and data underwent 5-fold cross-validation with an 80-20 train-test split, optimized through randomized grid search to minimize mean absolute error (MAE). For each cycle of iterative optimization, gaussian process-based Bayesian Optimization via Scikit-Optimize (0.9.0) was used to suggest 12 de novo LNP formulations to test per helper lipid, using expected improvement (EI) to trade-off exploration and exploitation.

Results, Conclusions, and Discussions: From initial screening across 6 helper lipids (n = 486), DSPC formulations displayed the highest overall transfection and were selected for subsequent compositional optimization (Fig. 1A-B). XGBoost (XGB) demonstrated the best predictive power of models tested, achieving impressive Pearson’s correlations (r = 0.91), and Spearman’s rank (r = 0.89) on hold-out datasets (Fig. 1C). Feature importance analysis, using SHapley Additive exPlanations (SHAP), reveals structure-function relationships learned by our trained model (Fig. 1D). Through three cycles of iterative optimization, a progressive shift towards improved transfection was observed, converging by the third iteration (Fig. 1E). Side-by-side experimental validation of ML-optimized formulations showed significantly higher transfection efficiency compared to grid-screened formulations, with the top ML-optimized formulation achieving a ~1.4-fold increase in luciferase signal (Fig. 1F) and ~1.8-fold increase in mCherry+ percentage (Fig. 1G) over the top gird-searched formulations. Our platform efficiently tunes LNP formulation composition, achieving significant improvement in transfection efficiency over conventional screening. This ML-guided LNP engineering platform for screening LNP compositional ratios can be adopted to optimize delivery for diverse payloads and cell types.

Acknowledgements (Optional):