Bit silencing in fingerprints enables the derivation of compound class-directed similarity metrics

Abstract

Fingerprints are molecular bit string representations and are among the most popular descriptors for similarity searching. In key-type fingerprints, each bit position monitors the presence or absence of a prespecified chemical or structural feature. In contrast to hashed fingerprints, this keyed design makes it possible to evaluate individual bit positions and the associated structural features during similarity searching. Bit silencing is introduced as a systematic approach to assess the contribution of each bit in a fingerprint to similarity search performance. From the resulting bit contribution profile, a bit position-dependent weight vector is derived that determines the relative weight of each bit on the basis of its individual contribution. By merging this weight vector with the Tanimoto coefficient, compound class-directed similarity metrics are obtained that further increase fingerprint search calculations compared to conventional calculations of Tanimoto similarity.