Evaluation of Distance Metrics for Ligand‐Based Similarity Searching

Ligand-based similarity metrics are frequently and successfully employed for diversity analysis and the selection of activity-enriched subsets in early-phase virtual screening and compoundlibrary design. As they come in many varieties, it is not trivial to choose the most appropriate concept for the task at hand. Fundamentally, these methods rely on representative reference structures (also termed TMquery∫ or TMseed∫ structures), molecular descriptors that are correlated with biological activity, and an appropriate similarity metric. TMRetrospective screening∫ provides a means of evaluating these factors. The basic idea is to select a subset from a large pool of compounds (typically a compound database or a virtual library) and try to maximize the number of known actives in the subset, thereby forming a TMfocused library∫. Subset selection is based on the pairwise similarity between the query structure and each molecule in the pool. The result of this calculation is a list ranked by similarity. Such a retrospective screening experiment can be rated by the enrichment factor, ef [Eq. (1)] . 6] A method that is superior to a random selection of compounds returns an ef>1.