An evaluation of the suitability of a number of structural descriptors and clustering methods for use in diversity selection is presented. The methods are compared by their success in simulated biological activity predictions. The results suggest that simple 2D structural descriptors are particularly effective and that hierarchical clustering methods are superior to the non-hierarchical methods traditionally used for diversity related tasks. Results are presented which suggest that the difference in the utility of the descriptors can be accounted for by the different extent to which each encodes information relevant to the forces of ligand-receptor binding.