No abstract

No abstract

No abstract

The contribution of individual fingerprint bit positions to similarity search performance is systematically evaluated. A method is introduced to determine bit significance on the basis of Kullback-Leibler divergence analysis of bit distributions in active and database compounds. Bit divergence analysis and Bayesian compound screening share a common methodological foundation. Hence, given the significance ranking of all individual bit positions comprising a fingerprint, subsets of bits are evaluated in the context of Bayesian screening, and minimal fingerprint representations are determined that meet or exceed the search performance of unmodified fingerprints. For fingerprints of different design evaluated on many compound activity classes, we consistently find that subsets of fingerprint bit positions are responsible for search performance. In part, these subsets are very small and contain in some cases only a few fingerprint bit positions. Structural or pharmacophore patterns captured by preferred bit positions can often be directly associated with characteristic features of active compounds. In some cases, reduced fingerprint representations clearly exceed the search performance of the original fingerprints. Thus, fingerprint reduction likely represents a promising approach for practical applications.

Molecular fingerprints have a long history in computational medicinal chemistry and continue to be popular tools for similarity searching. Over the years, a variety of fingerprint types have been introduced. We report an approach to identify preferred bit subsets in fingerprints of different design and "recombine" these bit segments into "hybrid fingerprints". These compound class-directed fingerprint representations are found to increase the similarity search performance of their parental fingerprints, which can be rationalized by the often complementary nature of distinct fingerprint features.