New similarity measures for ligand-based virtual screening

Abstract

The process of drug discovery using virtual screening techniques relies on “molecular similarity principle” which states that structurally similar molecules tend to have similar physicochemical and biological properties in comparison to other dissimilar molecules. Most of the existing virtual screening methods use similarity measures such as the standard Tanimoto coefficient. However, these conventional similarity measures are inadequate, and their results are not satisfactory to researchers. This research investigated new similarity measures. It developed a novel similarity measure and molecules ranking method to retrieve molecules more efficiently. Firstly, a new similarity measure was derived from existing similarity measures, besides focusing on preferred similarity concepts. Secondly, new similarity measures were developed by reweighting some bit-strings, where features present in the compared molecules, and features not present in both compared molecules were given strong consideration. The final approach investigated ranking methods to develop a substitutional ranking method. The study compared the similarity measures and ranking methods with benchmark coefficients such as Tanimoto, Cosine, Dice, and Simple Matching (SM). The approaches were tested using standard data sets such as MDL Drug Data Report (MDDR), Directory of Useful Decoys (DUD) and Maximum Unbiased Validation (MUV). The overall results of this research showed that the new similarity measures and ranking methods outperformed the conventional industry- standard Tanimoto-based similarity search approach. The similarity measures are thus likely to support lead optimization and lead identification process better than methods based on Tanimoto coefficients.