[Nuclear magnetic resonance and protein structure]

NMR provides a wealth of structural information about proteins in solution, but does not, by itself, permit an unambiguous determination of a unique structure. A rigorous interpretation of NMR data to obtain the entire family of structures compatible with a given data set requires extensive, systematic and unbiased sampling of the conformational space of the polypeptide chain. Methods of sampling based on the exclusion paradigm--i. e. those that generate structures, check constraints and accept or reject members of the family on that basis, avoid the problem of generating erroneous structures by converging on local minima, which is a common pitfall of methods based on the optimization paradigm. Their much higher computational cost can be reduced by solving the structure in stages, using abstract representations of partial structures, and guiding the computation by control heuristics. The heuristic refinement method developed at Stanford and encoded in the expert system PROTEAN yields more or less extensive families of structures, depending on the size of the NMR data set, and defines the "allowed volume" in which each atom (or other substructure) may lie, with all experimental constraints satisfied. The allowed volume is a measure of the uncertainty of our knowledge of the structure, to which both the limitations of the data and the uncertainty of position resulting from molecular motion may contribute. Prediction of the experimental NMR spectra by solving the generalized Bloch equations (or the Redfield density matrix) for the protein, using atomic coordinates that lie within the allowed atomic volume, provides the final test for the correctness of the proposed structure.(ABSTRACT TRUNCATED AT 250 WORDS)