Recent Advances in Machine Learning Models for Predicting Toxicity of Inorganic Nanoparticles

Nanoparticles (NPs) such as engineered inorganic NPs (metals, metal oxides, carbon materials, etc.) can induce cytotoxicity in normal biological systems when used for disease treatment or when exposed to the environment, which has raised widespread concerns about their safety in biomedicine, environmental chemistry, and other application fields. Therefore, developing efficient strategies for the hazard and risk assessment of NPs is extremely important to ensure their safety and sustainable development in above applications. Herein, we provide a systematic and comprehensive review that includes the following sections: (i) mechanisms and influencing factors of nanotoxicity, (ii) the classical statistical cytotoxicity prediction models such as nano-quantitative structure-activity relationship (nanoQSAR), physiologically based pharmacokinetic (PBPK), and meta-analysis (MA) models, (iii) the ML-accelerated development of the above three types of models, and (iv) some important nanotoxicity databases. The ML-accelerated nanoQSAR, PBPK, and MA models are mainly focused, in which the ML algorithms, advantages, and schemes for model development are described, and also the prediction performance and key features that influence the cytotoxicity for the developed models are discussed in detail. In addition, future opportunities and challenges in promoting the development of highly efficient, robust, and interpretable ML models for predicting the cytotoxicity of NPs are also highlighted.