Evaluation of descriptors and mini-fingerprints for the identification of molecules with similar activity

Combinations of 65 preferred 1D/2D molecular descriptors and 143 single str uctural keys were evaluated for their performance in compound classificatio n focused on biological activity. The analysis was based on principal compo nent analysis of descriptor combinations and facilitated by use of a geneti c algorithm and different scoring functions. In these calculations, several descriptor combinations with greater than 95% prediction accuracy were ide ntified. A set of 40 preferred structural keys was incorporated into a smal l binary fingerprint designed to search databases for compounds with biolog ical activity similar to query molecules. The performance of mini-fingerpri nts was tested by systematic similarity search calculations in a database c onsisting of compounds belonging to:seven biological activity classes, whic h had not been used to select effective descriptors. In these blind test ca lculations, mini-fingerprints correctly identified approximately 54% of com pounds sharing similar biological activity. and with 1% false positives. Th us, although the design of mini-fingerprints is conceptually simple, they p erform well inactivity-oriented similarity searching.