Protein structure determination by non-parametric regression and knowledge-based constraints.

We have devised a non-parametric regression-based approach for the estimation of small- and medium-range inter-residual three-dimensional (3d) distances in a protein using only the primary sequence as input. A multivariate analysis of variance technique is used to identify the attributes of the primary sequence that is most effective in determining the tertiary structure. Certain compactness and hydrophobic core building heuristics are used along with the estimated distances in a distance geometry program to predict the 3d-structure (tertiary fold). Our method is found to predict correctly the native topologies of small proteins having up to 150 residues. The sensitivity of the structures to long-range distance constraints is studied by incorporating a small number of NMR distance restraints. In terms of modularity, precision, accuracy and computational efficiency our method is found to be better in comparison with current computational methods like X-PLOR and DRAGON on the sample that was reported in the literature for the comparison of these two methods.