An evaluation of global QSAR models for the prediction of the toxicity of phenols to Tetrahymena pyriformis

This study presents an analysis of the ability of a two-parameter response surface, a multiple linear regression and a neural network model to produce global quantitative structure-activity relationships (QSARs) to predict the toxic potency of phenols to Tetrahymena pyriformis. The phenolic toxicity data set analysed is characterised by multiple mechanisms of toxic action. The study aimed to evaluate the confidence that can be applied to the modelling of the differing mechanisms of action. Assessment of confidence was decided in terms of whether the statistics for the global models reflect the ability of the QSARs to model the individual mechanisms of toxic action present in the data set. The results showed that the global statistics only reflected the ability of models to predict the two non-covalent mechanisms (polar narcosis and respiratory uncoupling), with the metabolically transformed and electrophilic mechanism (pre-electrophiles and soft electrophiles) being modelled poorly by all three model building methods. The results confirm the difficulty in modelling electrophilic mechanisms of toxic action. The results also highlight the fact that this poor predictivity is often 'hidden' in good statistical fit of some global models. In particular these results emphasise that for practical predictive purposes the mechanistic applicability domain is required to give confidence to estimated toxicity values.