Developing random forest based QSAR models for predicting the mixture toxicity of TiO2 based nano-mixtures to Daphnia magna

Abstract

During emission, TiO2 nanoparticles (NPs) might meet various chemicals, including metal ions and organic compounds in aquatic environments (e.g., surface water, sediments). At environmentally safe concentrations, combinations of both TiO2 NPs and those chemicals might cause cocktail effects (i.e., mixture toxicity) to aquatic organisms. Previous models such as concentration addition and independent action require dose-response curves of single components in the mixtures to predict the mixture toxicity. Structure-activity relationship (QSAR) models might predict the toxicity of nano-mixtures without dose-response curves of single components in the mixtures. However, current quantitative structure-activity relationship (QSAR) models are mainly focused on predicting cytotoxicity (i.e., cell viability) of heterogeneous metallic TiO2 nanoparticles (NPs) or mixtures of TiO2 NPs and four metal ions (Cu2+, Cd2+, Ni2+, and Zn2+). To minimize the experimental cost of nano-mixture risk assessment, in this study, we developed novel nano-mixture QSAR models to predict i) EC50 of 76 nano-mixtures containing TiO2 NPs and one of eight inorganic/organic compounds (i.e., AgNO3, Cd(NO3)2, Cu(NO3)2, CuSO4, Na2HAsO4, NaAsO2, Benzylparaben and Benzophenone-3), to Daphnia magna(D. magna), and ii) immobilization of D. magna exposed to one of 98 mixtures containing TiO2 NPs and one of eleven inorganic/organic compounds (i.e., AgNO3, Cd(NO3)2, Cu(NO3)2, CuSO4, Na2HAsO4, NaAsO2, Benzylparaben Benzophenone-3, Pirimicarb, Pentabromodiphenyl Ether and Triton X-100). The nano-mixture QSAR models were developed with mixture descriptors (Dmix) combing quantum descriptors of mixture components (e.g., TiO2 NPs and its partners) by using different machine learning techniques (i.e., random forest, neural network, support vector machine, and multiple linear regression). Nano-mixture QSAR models built with the random forest algorithm and proposed mixture descriptors exhibited good performance for predicting logEC50 (Adj.R2test = 0.955 ± 0.003, RMSEtest = 0.016 ± 0.002, and MAEtest = 0.008 ± 0.001) and immobilization (Adj.R2test = 0.888 ± 0.011, RMSEtest = 11.327 ± 0.730, and MAEtest = 5.933 ± 0.442). The models developed in this study were implemented in a user-friendly application for assessing the aquatic toxicity of TiO2 based nano-mixtures.