A stepwise approach for determining the model applicability domain is proposed. Four stages are applied to account for the diversity and complexity of the current SAR/QSAR models, reflecting their mechanistic rationality (including metabolic activation of chemicals) and transparency. General parametric requirements are imposed in the first stage, specifying in the domain only those chemicals that fall in the range of variation of the physicochemical properties of the chemicals in the training set. The second stage defines the structural similarity between chemicals that are correctly predicted by the model. The structural neighborhood of atom-centered fragments is used to determine this similarity. The third stage in defining the domain is based on a mechanistic understanding of the modeled phenomenon. Here, the model domain combines the reliability of specific reactive groups hypothesized to cause the effect and the domain of explanatory variables determining the parametric requirements in order for functional groups to elicit their reactivity. Finally, the reliability of simulated metabolism (metabolites, pathways, and maps) is taken into account in assessing the reliability of predictions, if metabolic activation of chemicals is a part of the (Q)SAR model. Some of the stages of the proposed approach for defining the model domain can be eliminated depending on the availability and quality of the experimental data used to derive the model, the specificity of (Q)SARs, and the goals of their ultimate application. The performance of the proposed definition of the model domain is tested using several examples of (Q)SARs that have been externally validated, including models for predicting acute toxicity, skin sensitization, and biodegradation. The results clearly showed that credibility in predictions of QSAR models for chemicals belonging to their domain is much higher than for chemicals outside this domain.
Journal of Chemical Information and Modeling, 2005, Vol 45, Issue 4, p. 839-849