Four Association Coefficients for Relating Molecular Similarity Measures

Four association coefficients are defined for assessing the relatedness between an arbitrary pair of molecular similarity measures. The utility of these coefficients is illustrated by assessing the relatedness of two molecular similarity measures in use at Pharmacia & Upjohn, one based on topological indices and the other on the occurence of structural fragments. The high coaggregation and discrimination coefficients imply that the two similarity measures preserve neighborhood relationships. Consequently, the two similarity measures should perform comparably with respect to property prediction if one used a locally-defined predictive method such as nearest neighbor prediction. The small density correlation coefficient indicates that the two measures could lead to quite different choices with regard to compound acquisition programs designed to fill in the “sparse” regions of structure space. The small distance correlation coefficient indicates that the two methods would perform differently when used for dissimilarity selection of subsets of diverse compounds. These coefficients replace our cruder assessments of the relatedness of two molecular similarity measures based on a series of similarity searches with more precise and statistically-defined statements regarding the manner in which they are related.