A difficulty for those working on developing virtual screening methods is how to evaluate them. It is now widely recognised that many evaluation studies carried out in the past were flawed due to poor statistical techniques, biased data sets, congeneric actives and other problems (see the JCAMD issue 'Evaluation of Computational Methods, issue 22 numbers 3-4).A number of authors have recently produced data sets for use in validating virtual screening methods, such as the Directory of Useful Decoys (DUD) and the Maximum Unbiased Validation (MUV) data set, or techniques for generating such data sets. A common feature of many algorithms used to design these data sets is concern about including unknown actives in the decoys. Having such unknown actives can lead to the �false false positive� problem, where a VS method correctly flags a decoy molecule as active, but the analysis penalises this as the molecule is marked as inactive.One common solution to this perceived problem is to filter the data set by removing decoys which are �too close� to the actives in some sense. However, we present analyses that show that, under reasonable assumptions of the frequency of actives in the decoy population, this filtering is not necessary. In typical virtual screening scenarios, the �false false positive� problem will not affect the validity or outcome of a retrospective virtual screening experiment. Moreover, the efforts to work around the perceived problem can cause more harm than good by adding additional bias, rendering invalid any comparisons of 3D and 2D methods.

False false positives - problem or red herring?

245th National Meeting (2013)

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