Knowledge-based potentials have been found useful in a variety of biophysical studies of macromolecules. Recently, it has also been shown in self-consistent studies that it is possible to extract quantities consistent with pair potentials from model structural databases. In this study, we attempt to extend the results obtained from these self-consistent studies toward the extraction of realistic pair potentials from the Protein Data Bank (PDB). The new method utilizes a clustering approach to define atom types within the PDB consistent with the optimal effective pairwise potential. The method has been integrated into the SMoG drug design package, resulting in an improved approach for the rapid and accurate estimation of binding affinities from structural information. Using this approach, it is possible to generate simple knowledge-based potentials that correlate (R = 0.61) with experimental binding affinities in a database of 118 diverse complexes. Furthermore, predictions performed on a random 1/3 of the database consistently show an average unsigned error of 1.5 log Ki units. It is also possible to generate specialized knowledge-based potentials, targeted to specific protein families. This approach is capable of generating potentials that correlate strongly with experimental binding affinities within these families (R = 0.8-0.9). Predictions on 1/3 of these family databases yield average unsigned errors ranging from 1.1 to 1.3 log Ki units. In summary, we describe a physically motivated approach to optimizing knowledge-based potentials for binding energy prediction that can be integrated into a variety of stages within a lead discovery protocol.