The selection of diverse molecules'subsets is a very important stake in the pharmaceutical research. Indeed, the effective discovery of a drug will depend of the quality of this selection. Several methods exist to address this problem. Some of them are based on the creation of groups of molecules, the others on the principle of dissimilarity between chemical compounds. In this work, we propose a new technique, between these two concepts, which allows to obtain subsets, at the same time, diverse in the space and representative from the initial set which they are extracted. First of all, to create this selection method, we defined and formalized mathematically a diversity criterion, then we used heuristics known in machine learning to conceive our algorithm. This one was compared with the other types of diversity selections usually used in chemoinformatic such as k-medoïds, Maximum-Dissimilarity, Sphere-Exclusion. The formalization of the diversity criterion finally allowed us to propose a new criterion of evaluation of the quality of the selections. The algorithm and the criterion presented in this work give diverse and representative samples of a chemical space.
