Deducing molecular similarity using Voronoi binding sites.

We have devised a new measure of molecular similarity with respect to given simple partitions of space into regions. The similarity is determined by numerical integration of the difference in the optimal interaction between the two molecules and the regions over a large range of interaction parameter values. Compounds differing in empirical formula are differentiated by a single infinite region; cis/trans or ortho/meta/para isomers are distinguishable by two adjacent regions that are half-spaces; and stereoisomers require five regions. This can be viewed as a natural classification of isomers. The concept can also be applied to drug binding studies to determine which molecules may bind alike in a given biological receptor and to elucidate a necessary starting geometry when a binding site is modeled for inhibitors whose experimental binding energies are different.