Applying quantitative structure-activity relationships approaches to nanotoxicology: Current status and future potentials

Application of in silico methods involves several steps. Firstly, chemical or structural properties of nanomaterials are represented by mathematical objects called descriptors, many of which can be calculated rather than measured. Examples of descriptors suitable for nanomaterials include particle size, shape and surface area, ionization potentials of metals, heats of formation of metal oxide clusters6, zeta potentials, and physicochemical properties (e.g. lipophilicity, hydrogen bond donor or acceptor strength) of molecules covalently bound to nanoparticles surfaces. Secondly, using additional mathematical techniques, subsets of descriptors are chosen that are most likely to relate to the biological property (e.g. cell apoptosis, metabolism, or signalling pathway modulation) being modelled. Statistical modelling or machine learning methods, often employing neural networks, generate mathematical models linking descriptors to biological activity. Finally, the model’s robustness and ability to predict properties of new materials is assessed by statistical cross-validation techniques, or by predicting properties of materials in a test set not used to develop the model.