SIMILARITY SEARCHING AND CLUSTERING OF CHEMICAL-STRUCTURE DATABASES USING MOLECULAR PROPERTY DATA

Previous work on the clustering of chemical-structure databases has focused on the use of intermolecular similarity measures that are based on structural features of various kinds. In this paper, we report nearest-neighbor searching and clustering experiments with a set of 5982 molecules, each of which is characterized by 13 calculated global molecular properties. The nearest-neighbor algorithm is an upperbound procedure that uses the triangle inequality to minimize the number of distance calculations that need to be carried out when searching for nearest neighbors in metric spaces. Our experiments suggest that it performs well when small numbers of nearest neighbors are required, but that the basic ''brute-force'' procedure is best when large numbers are needed, such as when clustering is to be carried out. The clustering methods tested are the Ward and group-average hierarchic agglomerative methods, the minimum-diameter polythetic hierarchic divisive method, and the Jarvis-Patrick nearest-neighbor method. Our experiments suggest that the first three methods, which gave similar results, are the best methods for clustering molecules characterized by property data. The Jarvis-Patrick method, which has been extensively used for clustering molecules characterized by structural fragments, was not as effective as these other methods.