Hamiltonians for protein tertiary structure prediction based on three-dimensional environment principles.

We describe a computational approach to protein tertiary structure prediction that combines ideas from the three-dimensional (3D) profile method of Bowie, Lüthy and Eisenberg and the associative memory Hamiltonians of Friedrichs and Wolynes. The ultimate goal of our work is to extend and generalize the capabilities of these heuristics so as to be able to predict novel structures that might be found in nature or designed proteins. In our approach we approximate the interactions between residues through a pseudo-potential function similar to an associative memory Hamiltonian. This function is constructed based on 3D environment principles. Favorable inter-residue contacts for each residue in a target protein are inferred by using 3D environment propensities of the residues and a collection of 3D environment templates derived from a dataset of protein crystal structures. A Hamiltonian encoding this information is used to guide an optimization phase via molecular dynamics with annealing, which then leads to the folded structure. With our algorithm we can recover the structure of dataset proteins and have also succeeded in constructing the fold for a protein with little sequence similarity to any dataset protein.