Applicability domain estimation for classification QSARs on example of Ames test and CYP450 inhibition

Iurii Sushko, Sergii Novotarskyi, Robert Koerner, Ahmed Abdelaziz, Wolfram Teetz and Igor Tetko, eADMET GmbH, Germany

In QSAR research, it is of crucial importance to determine the compounds where models give reliable predictions, that is the applicability domain of QSAR models. One of the approaches that has been shown to provide good results for regression QSARs is based on so called distances to models (DMs) - special metrics that estimate the prediction accuracy of QSAR models. This work generalizes this approach to classification QSARs and shows the its successful application to the prediction of mutagenicity and CYP450 inhibition potential of chemical compounds.
For both the predicted properties, our approaches could identify highly accurate predictions, which have the accuracy close to that of experimental measurements (90-95%). Precisely these predictions are most useful and should be used to substitute experimental measurements and, therefore, save significant efforts and costs. On the contrary, it was also possible to identify very inaccurate predictions with the accuracy close to random (50%). The use of such predictions, naturally, is infeasible and should be avoided. Thus, we prove that DMs can be successfully used to estimate the prediction accuracy and, therefore, to estimate applicability domain no only for regression, but also for classification problems.

(presenting author: Iurii Sushko)