Identification of a ligand binding site on a protein is pivotal to drug discovery. To date, no reliable and computationally feasible general approach to this problem has been published. Here we present an automated efficient method for determining binding sites on proteins for potential ligands without any a priori knowledge. Our method is based upon the multiscale concept where we deal with a hierarchy of models generated using a k-means clustering algorithm for the potential ligand. This is done in a simple approach whereby a potential ligand is represented by a growing number of feature points. At each increasing level of detail, a pruning of potential binding site is performed. A nonbonding energy function is used to score the interactions between molecules at each step. The technique was successfully employed to seven protein-ligand complexes. In the current paper we show that the algorithm considerably reduces the computational effort required to solve this problem. This approach offers real opportunities for exploiting the large number of structures that will evolve from structural genomics.