Domain applicability is the art of assigning an error bar to a QSAR prediction of a given molecule M on a given QSAR (activity) model. There are many ways address this issue and there is no consensus in the field. A number of “DA metrics” of various flavors (distance to training set, bagged variance, etc.) have been proposed that might correlate with prediction error.  In our laboratory we prefer two-tiered modeling approach to domain applicability. That is, we produce a set of prediction errors and DA metrics by cross-validation, and then use a QSAR method (in our case random forest), to generate a second model (“error” model”), which predicts the mean absolute error from the DA metrics. QSAR has well-established methods for determining the most important parameters in a given model and for determining whether a model is statistically significant, so we can tell whether the error model is significant and which DA metrics are most important. Generally, if the training set for the QSAR model is diverse, and random forest was used to produce the activity model, the most important DA metrics for predicting are TREE_SD (the variation of predictions among random forest trees), and PREDICTED, the predicted value itself.  If the training set is very narrow, SIMILARITYNEAREST1, similarity of M to the nearest training set compound, becomes most important.  Generally speaking, though, DA metrics that have to do with “cliffiness” of the training set molecules around M (e.g. the DA metric CONFIVE) do not appear very important in predicting errors.  On the other hand, compounds on activity cliffs are notoriously hard to predict, whatever the QSAR method or descriptor, and thus DA metrics having to do with cliffiness would be expected to be important. This presents a seeming paradox.<br/>

Error model approach to domain applicability: Is there an activity cliff paradox?

251st National Meeting (2016)

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.