A knowledge-based method for protein structure refinement and prediction

The native conformation of a protein, in a given environment, is determined entirely by the various interatomic interactions dictated by the amino acid sequence (1-3). We describe here a knowledge-based approach for protein structure assessment and prediction. Using a well-defined set of high-resolution protein structures, we have derived statistical potentials, in the form of atom-pairwise distance probability density functions. These provide a description of pairwise interatomic interactions of native proteins. When applied to highly randomized and noisy structures of proteins distinct from the basis set, native-like structures were obtained to very high precision (< or = 2A). The examples tested include proteins of all sizes (from 38 up to 461 amino acids long) and diverse topological structures (alpha, beta and alpha-beta classes). The potentials appear to be sensitive enough to recognize subtle distortions from a native packing structure and in optimization of structures drive them consistently to a higher probability. Therefore they provide a powerful tool for refinement of X-ray and NMR derived structures at arbitrary degrees of initial precision.