Development of a Compound Class-Directed Similarity Coefficient That Accounts for Molecular Complexity Effects in Fingerprint Searching

In chemoinformatics, fingerprints are defined as bit string representations of molecular structure and properties. The evaluation of fingerprint similarity, that is, quantification of fingerprint overlap, is known to be biased by differences in molecular complexity and size. For example, similarity searching using optimized lead compounds that are typically more complex and larger than average database compounds often leads to the artificial selection of large molecules. A modified version of the Tversky coefficient has been introduced to balance such complexity effects. In addition, compound class-directed fingerprint bit position-dependent weight vectors have been designed to aid in the assessment of Tanimoto similarity. We show that by merging weight vectors with the modified Tversky coefficient, a class-directed similarity metric is obtained that effectively balances complexity effects and further improves the accuracy of fingerprint search calculations.