Many different measures of structural similarity have been suggested for matching chemical structures, each such measure focusing upon some particular type of molecular characteristic. The multi-faceted nature of biological activity suggests that an appropriate similarity measure should encompass many different types of characteristic, and this paper discusses the use of data fusion methods to combine the results of searches based on multiple similarity measures. Experiments with several different types of dataset and activity suggest that data fusion provides a simple, but effective, approach to the combination of individual similarity measures. The best results were generally obtained with a fusion rule that sums the rank positions achieved by each molecule in searches using individual measures.

Many different measures of structural similarity have been suggested for matching chemical structures, each such measure focusing upon some particular type of molecular characteristic. The multi-faceted nature of biological activity suggests that an appropriate similarity measure should encompass many different types of characteristic, and this paper discusses the use of data fusion methods to combine the results of searches based on multiple similarity measures. Experiments with several different types of dataset and activity suggest that data fusion provides a simple, but effective, approach to the combination of individual similarity measures. The best results were generally obtained with a fusion rule that sums the rank positions achieved by each molecule in searches using individual measures.

EVA is a new molecular descriptor that provides a concise summary of the fundamental frequency components of a molecule's infrared range vibrational spectrum in a vector format. Target structures from the Starlist database are used to demonstrate the effectiveness of the descriptor for similarity searching and its difference from a conventional similarity measure based on the matching of two-dimensional (2D) fingerprints. The use of data fusion on the rankings resulting from the EVA-based and the 2D-based similarity measures results in a combined ranking that can be more effective in simulated property prediction experiments than either of the individual rankings.

WILLETT, P.; FERGUSON, A. M.; HERITAGE, T. W.; J. Chem. Inf.