BALANCING REPRESENTATIVENESS AGAINST DIVERSITY USING OPTIMIZABLE K-DISSIMILARITY AND HIERARCHICAL-CLUSTERING

When assessing the pharmacological potential of large libraries of com pounds, it is often useful to start by determining the biochemical act ivities of some subset thereof. This is so whether the compounds in qu estion have in fact already been synthesized or exist solely as virtua l libraries. A suitable subset for this task must be structurally dive rse, so as to minimize redundant testing, but must also be representat ive, so that valuable subgroups do not get overlooked. These two needs are intrinsically in conflict, with gains in one necessarily coming a t the expense of the other. Results obtained using optimizable K-dissi milarity selection and clustering are described and compared with thos e obtained using more traditional agglomerative hierarchical clusterin g techniques.