On the Analysis of Compressed Chemical Fingerprints

Abstract

Chemical fingerprints are binary strings used to represent the distinctive features of molecules in order to efficiently support similarity search of chemical data. In large repositories, chemical fingerprints are conveniently stored in compressed format, although the lossy compression process may introduce a systematic error on similarity measures. Simple correction formulae have proposed by Swamidass and Baldi in [13] to compensate for such an error and, thus, to improve the similarity-based retrieval. Correction is based on deriving estimates for the weight (i.e., number of bits set to 1) of fingerprints before compression from their compressed values. Although the proposed correction has been substantiated by satisfactory experimental results, the way in which such estimates have been derived and the approximations applied in [13] are not fully convincing and, thus, deserve further investigation. In this direction, the contribution of this work is to provide some deeper insight on the fingerprint generation and compression process, which could constitute a more solid theoretical underpinning for the Swamidass and Baldi correction formulae.