Identification of diverse database subsets using property-based and fragment-based molecular descriptions

This paper reports a comparison of calculated molecular properties and of 2D fragment bit-strings when used for the selection of structurally diverse subsets of a file of 44295 compounds. MaxMin dissimilarity-based selection and k-means cluster-based selection are used to select subsets containing between 1% and 20% of the file. Investigation of the numbers of bioactive molecules in the selected subsets suggest: that the MaxMin subsets are noticeably superior to the k-means subsets; that the property-based descriptors are marginally superior to the fragment-based descriptors; and that both approaches are noticeably superior to random selection.