Comparison between molecular dynamic based and knowledge based potentials for protein side chains

Parameterizing protein coarse-grained models from atomic-level force fields is a relatively new and promising approach in protein modeling. Here, dihedral angle potentials for the amino acid side chains are derived using molecular dynamic simulations. These are compared to those obtained using the traditional knowledge based approach, where the potentials are obtained from known protein structures. Side chain potentials consist of two- or three-dimensional dihedral angle histograms with a 20 degrees resolution. The simulations for the amino acids are carried out in explicit water using variants of the united atom Gromos force field, in the all-atom OPLS-AA/L force field, and in implicit solvent using variants of the Amber force field. It was found that the knowledge-based and molecular dynamic potentials are significantly correlated, with correlation coefficients in the upper 0.70. Nevertheless, in energy minimization tests performed on a group of proteins keeping the backbones fixed, the knowledge-based potentials generate angles that correspond closer to the angles in native structures (about 20% closer), for either buried and solvent exposed residues. Furthermore, in tests using high-resolution proteins, the prediction accuracy for buried residues reached 88%. Among the molecular dynamic-based potentials, the one derived using the G43A2 force field resulted in the highest prediction accuracy.