Integrated in silico approaches for the prediction of Ames test mutagenicity

Abstract

The bacterial reverse mutation assay (Ames test) is a biological assay used to assess the mutagenic potential of chemical compounds. In this paper approaches for the development of an in silico mutagenicity screening tool are described. Three individual in silico models, which cover both structure activity relationship methods (SARs) and quantitative structure activity relationship methods (QSARs), were built using three different modelling techniques: (1) an in-house alert model: which uses SAR approach where alerts are generated based on experts judgements; (2) a kNN approach (k-Nearest Neighbours), which is a QSAR model where a prediction is given based on outcomes of its k chemical neighbours; (3) a naive Bayesian model (NB), which is another QSAR model, where a prediction is derived using a Bayesian formula through preselected identified informative chemical features (e.g., physico-chemical, structural descriptors). These in silico models, were compared against two well-known alert models (DEREK and ToxTree) and also against three different consensus approaches (Categorical Bayesian Integration Approach (CBI), Partial Least Squares Discriminate Analysis (PLS-DA) and simple majority vote approach). By applying these integration methods on the validation sets it was shown that both integration models (PLS-DA and CBI) achieved better performance than any of the individual models or consensus obtained by simple majority rule. In conclusion, the recommendation of this paper is that when obtaining consensus predictions for Ames mutagenicity, approaches like PLS-DA or CBI should be the first choice for the integration as compared to a simple majority vote approach.