How similar are those molecules after all? Use two descriptors and you will have three different answers

Molecular similarity searching (ligand-based virtual screening) is one of the routine computational techniques used in drug discovery and pharmacological research. However, while a large number of descriptors exist, there are no general guidelines whatsoever which descriptors work better and which descriptors should be used in the different cases.

This review provides a brief overview of current molecular descriptors and databases used for their evaluation, followed by a critical discussion of their differences.

After reading this review, the reader will be aware of how very differently molecular descriptors assess similarities of molecules, and the performance that can be realistically expected from them.

Molecular descriptors come in a variety of forms, and they show vast differences in assessing the similarity between molecules. Virtual screening performance of many descriptors is often lower than expected, compared to 'dumb' descriptors while some simple methods such as circular fingerprints offer surprisingly good performance in many cases. The choice of the right benchmark library is crucial, many of which are summarized in this review.