Title : High throughput optimization of cell-specific lipid nanoparticles through learning to rank
Abstract:
Lipid nanoparticles (LNPs) have gained traction in recent years due to their use in the COVID-19 vaccine. They are able to deliver nucleic acids, enabling cells to transiently acquire novel therapeutic functions. Yet, one persistent problem remains: the inability to achieve cell-type specificity. Most LNPs exhibit off-target transfection, with significant accumulation in the liver – limiting their application in precision medicine. To address this, we developed a high-throughput computational pipeline that explores the vast compositional space of LNPs to identify cell-specific formulations.
Methodology: We leverage a Learning to Rank (LTR) machine learning framework. Unlike traditional regression models that predict absolute transfection – values that often fluctuate significantly between experimental runs – LTR predicts the relative performance of LNPs. While absolute transfection may vary, the performance hierarchy of different LNP formulations remains consistent across experiments. Input features for the model included formulation compositional parameters and UniMol molecular embeddings to capture the structural and spatial properties of helper and ionizable lipids varied throughout the study. Our initial dataset, created with high-throughput screening, included novel ionizable lipid chemistries. The ranking logic was then implemented alongside the Non-dominated Sorting Genetic Algorithm II (NSGA-II). This allows for multi-objective optimization, where the model is used to maximize transfection in a target cell type whileminimizing off-target transfection in another.
Results: The model demonstrated high predictive accuracy, achieving a validation Normalized Discounted Cumulative Gain (NDCG) of 0.84 and a test NDCG of 0.87. Both the validation and test sets consisted of held-out ionizable lipid structures. Thus, the high performance suggests the model generalizes to novel chemistries. In addition, SHAP (SHapley Additive exPlanations) analysis identified compositional parameters and specific features within the molecular embeddings that drive model predictions. These SHAP values were then used to generate t-SNE plots, allowing for the identification of helper and ionizable lipid pairs that form high-transfection clusters. Furthermore, the NSGA-II algorithm successfully generated a Pareto front of LNPs capable of potent transfection in B16 melanoma cells while avoiding Jurkat T-cells.
Future Directions: Ongoing work focuses on improving this pipeline by introducing a custom loss function in the LTR architecture to increase ranking accuracy. We’re also working to increase interpretability of the
molecular embeddings, elucidating the chemical features of helper and ionizable lipids that lead to higher transfection. This framework will be expanded to directly optimize lipid chemistries. We’ll also apply it to optimize poly-beta amino ester (PBAE) nanoparticles, fine tuning polymer-based delivery systems. Ultimately, we aim to continue developing a scalable, interpretable approach toward engineering next-generation drug delivery vehicles.
