The predictivity of QSARs for toxicity: Recommendations for improving model performance

Quantitative structure–activity relationships (QSARs) are invaluable computational tools for the prediction of the biological effects and physico-chemical properties of molecules. For chemical safety assessment they are used frequently to make predictions of toxic or adverse effects, as well as other activities related to toxicokinetics. QSARs and their predictions can be assessed against a number of criteria for their potential use as surrogates for animal, or other, tests. A recent exercise by the Division of Genetics and Mutagenesis, National Institute of Health Sciences, Japan, assessed QSARs to predict the outcome of the Ames test. The predictive performance of models was scrutinised with full disclosure of results. The authors of this publication developed one such model, which had disappointing performance in this predictive exercise. In order to understand why the QSAR had poor performance metrics, this paper reflects on factors that affect a QSAR model. There is no one reason for poor performance of a QSAR model, rather it is likely to be a combination of factors. Reasons for poor performance included inadequate consideration of the underlying data quality, consistency and relevance; lack of appropriate descriptors relating to the endpoint and mechanism of action; not selecting a model correctly in terms of its structure (i.e., complexity) and number of descriptors; not addressing metabolism adequately in the modelling process; ill-defined assessment of the uncertainties within a model; and not ensuring predictions are within the applicability domain of the model. Whilst this paper draws on examples for the prediction of mutagenicity, the findings are applicable to all toxicological activities and physico-chemical properties.