Forecasting nanoparticle toxicity using nonlinear predictive regressor learning systems

Nanoparticle (NP) toxicity is determined by a vast number of topological, sterical, physico-chemical as well as biological properties, rendering a priori evaluation of the effect of NP on biological tissue as arduous as it is necessary and urgent. We aimed at mining the HORIZON 2020 MODENA COST NP cytotoxicity database through nonlinear predictive regressor learning systems in order to assess the power of available NP descriptors and assay characteristics in predicting NP toxicity. Specifically, we assessed the results of cytotoxicity assays performed on 57 NP and trained two different nonlinear regressors (Support Vector Regressors [SVR] with polynomical kernels and Radial Basis Function [RBF] regressors) within a nested-cross validation scheme for parameter optimization to predict toxicity as quantified by EC25, EC50 and slope while using the regressional ReliefF algorithm (RReliefF) for feature selection. Available NP attributes were material, coating, cell type, dispersion protocol, shape, 1st and 2nd dimension, aspect ratio, surface area, zeta potential and size in situ. In most regressor learning systems, after feature selection with the RReliefF algorithm, the correlation between real and estimated toxicity endpoint values increased monotonically with the number of included features, reaching values above 0.90. The best performance was obtained with RBF regressors, and the most informative features in predicting toxicity endpoints were related to nanoparticle structure. These trends did not change significantly between toxicity endpoints. In conclusion, EC25, EC50 and slope can be predicted with high correlation using purely data-driven, machine learning methods in Adenosine triphosphate (ATP)-based NP cytotoxicity assays