Validation status of QSAR models for exposure- and effects-related parameters

Abstract

The fact that QSARs are local models (unless the respective endpoint can be traced back to a single uniform rate-limiting process) inevitably causes varying degrees of validity and reliability for their predictive applications. Depending on the chemical class and the endpoint concerned, the estimates may range between satisfactory to very rough approximations. Corresponding to the variability in the underlying experimental data, the estimates may be exact by ± 10% or may scatter by a factor of 10 or more. Furthermore, the accuracy of the QSAR predictions depends on the selection of the appropriate model on a case-by-case basis. A traditional illusion is that physico-chemical measurements are more exact than biological tests, so it is not surprising that severe problems remain for predicting such apparently simple properties as water solubility or melting points, whilst complex endpoints such as aquatic toxicity can be calculated with a rather high degree of confidence. The apparent success of QSARs for selected endpoints strongly depends on the (experimental) efforts devoted to the respective parameters, as also influenced by the demands of the regulating authorities for certain summarizing parameters.