Predicting Cytotoxicity of Nanoparticles: A Meta-Analysis Using Machine Learning

Cytotoxicity evaluation of nanoparticles (NPs) is regarded as a crucial step for their successful application in the biomedical industry. However, conventional experimental methodologies for cytotoxicity measurements are often expensive, time-consuming, and demand intense training in cell culture. In this study, we developed generalized machine learning (ML) models for both qualitative and quantitative prediction of cytotoxicity across a wide variety of NPs. In particular, a meta-analysis of cytotoxicity data was conducted from published literature on metallic, metal oxide, polymer, and carbon-based NPs, leading to the development of random forest-based regression and classification models for predicting cell viability from physicochemical properties of NPs, cellular attributes, and testing conditions. Our feature importance analysis showed that accurately predicting the cytotoxicity of NPs using the regression model requires knowledge of their composition, concentration, zeta potential, and size, as well as exposure time, toxicity assay, and tissue type. Interestingly, among these attributes, the information about composition of NPs or tissue type was not needed for achieving high accuracy in the qualitative prediction of cytotoxicity using the classification model, indicating its superior robustness compared to the regression model. These findings may encourage future researchers to employ ML models more effectively and frequently to reliably assess the safety of NPs.