The present study introduces a new strategy of selection of a maximum diversity sample of n compounds from N available in a molecular database. This strategy can be useful in pharmacological screening, combinatorial chemistry or parallel synthesis planning. It consists of first describing the compounds by means of parameters derived from quantum mechanical computations (water solvation ΔG, benzene solvation ΔG, octanol solvation ΔG, dipolar moment), as well as standard molecular parameters such as solvent-accessible surface area and molecular weight. Solvation parameters are used because of the importance of this phenomenon in the pharmacological behaviour. Redundant information in the description of the compounds is eliminated by using principal components (PC) instead of the original descriptors. Based on the similarity between the N compounds in the PC space, they are classified into n groups by k-means cluster analysis. The compounds that are nearest to the centroid of each cluster co nstituted the maximum diversity sample. When practical difficulties exist for the use of one of the proposed compounds, another also close to the cluster centroid can substitute for it. This strategy has been tested in the selection of a sample of 50 amines from the 923 available in the Aldrich catalogue. The results have been contrasted with those obtained from an optimal, distance-based experimental design, resulting in an 86 percent of agreement between both approaches. An R²-like diversity coefficient has been used to assess the quality of the proposed solutions.