Abstract
Metal-oxide nanoparticles find widespread applications in mundane life today, and cost-effective evaluation of their cytotoxicity and ecotoxicity is essential for sustainable progress. Machine learning models use existing experimental data, and learn the relationship of various features to nanoparticle cytotoxicity to generate predictive models. In this work, we adopted a principled approach to this problem by formulating a feature space based on intrinsic and extrinsic physico-chemical properties, but exclusive of any in vitro characteristics such as cell line, cell type, and assay method. A minimal set of features was developed by applying variance inflation analysis to the correlation structure of the feature space. Using a balanced dataset, a mapping was then obtained from the normalized feature space to the toxicity class using various hyperparameter-tuned machine learning models. Evaluation on an unseen test set yielded > 96% balanced accuracy for both the random forest model, and neural network with one hidden layer model. The obtained cytotoxicity models are parsimonious, with intelligible inputs, and include an applicability check. Interpretability investigations of the models yielded the key predictor variables of metal-oxide nanoparticle cytotoxicity. Our models could be applied on new, untested oxides, using a majority-voting ensemble classifier, NanoTox, that incorporates the neural network, random forest, support vector machine, and logistic regression models. NanoTox is the very first predictive nanotoxicology pipeline made freely available under the GNU General Public License (https://github.com/NanoTox).
Competing Interest Statement
The authors have declared no competing interest.
