The goal of this project is to develop computational methods for evaluation of molecular models. Evaluation is based on the statistical distribution of non-local, nonbonded interactions in the model tertiary structure. This distribution is compared to that of correct structures in a crystallographic data base, using a Boltzmann-like probability law. We interpret evaluation functions derived in this way as potentials of mean force, or empirical free energy functions. The approximate validity of the Boltzmann-like probability law has been demonstrated by an analysis ion-pair frequency in protein crystal structures. Estimated electrostatic potential as a function of inter-atomic distance agrees well with expectations from physical principle and experiment. We have used this probability law to derive an energy function for evaluation of the initial models produced in homology modeling. These initial models assign a new amino acid sequence to the known polypeptide backbone structure of a suspected homolog, but do not specify locations of side-chain atoms. In the energy function we represent amino acid residues by mean coordinates for peptide and side chain groups, and we categorize nonbonded interactions by residue pair type and distance interval. Energies for contacts of each type are derived by analysis of a crystal structure data base, using a statistical model appropriatefor categorical data. We have shown that this energy function can identify correct models among millions of incorrect, """"""""misfolded"""""""" alternatives. Models based on the structures of related proteins can provide valuable insight into the biological function of proteins where only sequence is known. Existing computational methods cannot confirm their correctness, however. The importance of this work is in providing quantitative criteria for model evaluation.
