Relevance of Feature Combinations for Similarity Searching Using General or Activity Class-Directed Molecular Fingerprints

Abstract

We introduce a methodology for the systematic identification of feature combinations derived from fingerprints of bioactive compounds. Structural features were organized in co-occurrence networks from which reference set-based feature cliques were extracted. A similarity search strategy is presented that is based on frequency ranking of cliques. Three types of fingerprints have been compared that are either general in their design or incorporate, to a different degree, compound class information. Taking control calculations into account, search performance was overall best for a molecule-specific extended connectivity fingerprint. For compound class-directed fingerprints, reference set-derived feature cliques occurring with low frequency in a screening database were found to consistently enrich active compounds in database selection sets, even if the features are not highly conserved among reference compounds. Thus, in contrast to general fingerprints, feature combinations play a crucial role in similarity searching using activity-class-directed fingerprints.