Decoding silica nanoparticle toxicity: integrating machine learning, feature importance, rule extraction, and adsorption-uptake processes

The complexity of nanoparticle toxicity necessitates applying machine learning to large toxicological datasets to identify predictive features and toxicity rules. This study investigates the relationships between the physicochemical properties of silica nanoparticles (SiNPs), external experimental parameters, and toxicity under in vitro conditions. Data from the literature, databases, and in-house experiments were balanced using our Balanced Fitted Dose–Response approach, yielding nearly equal numbers of concentration data points across toxicity levels. The CatBoost algorithm outperformed Decision Tree models in predicting SiNP toxicity. The Catboost Shapley value analysis revealed the following sequence of feature importance: Mass > Total Surface Area > Serum > Total NP Number ∼ Primary Size > Exposure Time > Chemical Surface Modification > Cell Age > Cell Culture > Cell Disease. Rule extraction from a transparent decision-tree model revealed detailed information on the interrelationships among the features of surface-modified SiNPs and their toxicity. It showed that surface-unmodified SiNPs–OH are more toxic than surface-modified SiNPs–NH2 and SiNPs–COOH. Adsorption and uptake measurements by imaging flow cytometry, employing alveolar macrophages, revealed significantly higher adsorption for SiNPs–OH compared with modified SiNPs, while uptake of SiNPs–OH was lower. The integration of computational and experimental findings suggests that adsorption of SiNPs to the macrophage membrane contributes to their toxicity, in addition to internalization, providing valuable insight into the mechanisms underlying nanoparticle safety.