Neighborhood behavior: a useful concept for validation of "molecular diversity" descriptors

When searching for new leads, testing molecules that are too "similar" is wasteful, but when investigating a lead, testing molecules that are "similar" to the lead is efficient. Two questions then arise. Which are the molecular descriptors that should be "similar"? How much "similarity" is enough? These questions are answered by demonstrating that, if a molecular descriptor is to be a valid and useful measure of "similarity" in drug discovery, a plot of differences in its values vs differences in biological activities for a set of related molecules will exhibit a characteristic trapezoidal distribution enhancement, revealing a "neighborhood behavior" for the descriptor. Applying this finding to 20 datasets allows 11 molecular diversity descriptors to be ranked by their validity for compound library design. In order of increasing frequency of usefulness, these are random numbers = log P = MR = strain energy < connectivity indices < 2D fingerprints (whole molecule) = atom pairs = autocorrelation indices < steric CoMFA fields = 2D fingerprints (side chain only) = H-bonding CoMFA fields.