Machine learning-based models to predict aquatic ecological risk for engineered nanoparticles: using hazard concentration for 5% of species as an endpoint.

Abstract

Assessment and prediction for the ecotoxicity of engineered nanoparticles (ENPs) at the community or ecosystem levels represents a critical step toward a comprehensive understanding of the ecological risks of ENPs. Current studies on predicting the ecotoxicity of ENPs primarily focus on the cellular and individual levels, with limited exploration at the community or ecosystem levels. Herein, we present the first of the reports for the direct prediction of aquatic ecological risk for ENPs at the community level using machine learning (ML) approaches in the field of computational toxicology. Specifically, we extensively collected the threshold concentrations of twelve ENPs including metal- and carbon-based nanoparticles for aquatic species, i.e., hazardous concentrations at which 5% of species are harmed (HC5), established by a species sensitivity distribution. Afterwards, we used eight supervised ML methods including Adaboost, artificial neural network, C4.5 decision tree, K-nearest neighbor, logistic regression, Naive Bayes, random forest, and support vector machine to develop nine classification models and four regression models, respectively, for the qualitative and quantitative prediction of HC5. The evaluation of model performance yielded the internal validation accuracy of all classification models ranging from 71.4 to 100%, and the determination coefficient of regression models ranging from 0.702 to 0.999, indicating that the developed models showed good performance. By using a cross-validation method and an application domain characterization, the selected models were further validated to have powerful predictive ability. Furthermore, the incorporation of three nanostructural descriptors (metal oxide sublimation enthalpy, zeta potential, and specific surface area) linked to toxicity mechanisms (the release of metal ions, the stability of dispersions of particles in aqueous suspensions, and the surface properties of the material) effectively enhanced the prediction power and mechanistic interpretability of the selected models. These findings would not only be beneficial in the screening of ENPs with potential high ecological risks that need to be tested as a priority but also contribute to the development of environmental regulations and standards for ENPs.