A Two-Pronged Method for the Identification of Highly Biocompatible Nanomaterials

Overview

Nanoparticle (NP) toxicity analysis is critical for minimizing their potential harm to the biological system through non-specific uptake. To this end, this study introduces a machine learning pipeline focused on model explainability for predicting NP toxic ity, utilizing an in-vitro dataset ( N = 10,856 samples) characterized by 20 physicochemical properties and experimental conditions. It leverages feature selection algorithms and SHapley Additive exPlanations (SHAP) to pinpoint critical toxicity determinants, employing the Gradient Boosting Classifier, which outperformed 24 models with accuracy and recall of over 90%, to identify seven key parameters influencing toxicity. The model's robustness was confirmed through external verification (N = 135 samples) and application to an in vivo dataset ( N =126 samples), both showing over 85% accuracy, indicating transferability. The second half of the study explores the effect of lipid subcomponents on Lipid Nanoparticle (LNP) cytotoxicity in Prostate Cancer and Human Embryonic Kidney (HEK) cells, marking the first comprehensive assessment of lipid subcomponent impact on toxicity. It reveals the necessary helper lipid quantity to mitigate the toxic effects of highly positive NPs in HEK Cells, prompting a reevaluation of charge impact on NP toxicity. PEG and Cholesterol were deemed non -toxic, with all LNPs found non-toxic to Prostate Cancer Cells. This dual approach aims to enhance nanoparticle identification and elimination in drug development, offering potential cost and time efficiencies. Tennessee