A new empirically based method for predicting favourable interaction regions within the binding sites of proteins is presented. The method uses spatial distributions of atomic contact preferences derived from a non-homologous dataset of 83 high-resolution protein structures. The contact preferences are obtained for 26 different atom types relative to 163 different types of three-atom fragments. Each fragment consists of a triplet of bonded atoms, 1-2-3, which defines a reference frame for the three-dimensional distributions. In this way, directional, as well as distance, information is retained. Once derived, the distribution can be applied in a predictive manner. Given a protein's binding site, each distribution is transformed on to the three-atom fragments of the constituent residues and, when combined, can identify the favourable interaction regions for each different atom type. These predicted regions can then form the basis either for the modification of known inhibitors or for the search and design of new ones. Five known protein-ligand complexes are used to demonstrate the validity and usefulness of the approach. The results show that the method provides a powerful tool both in understanding how a given ligand exploits the interactions available to it in an active site and in helping to design improved, or novel, protein ligands.