Folding core predictions from network models of proteins

Abstract

Two different computational methods are employed to predict protein folding nuclei from native state structures, one based on an elastic network (EN) model and the other on a constraint network model of freely rotating rods. Three sets of folding cores are predicted with these models, and their correlation against the slow exchange folding cores identified by native state hydrogen–deuterium exchange (HX) experiments is used to test each method. These three folding core predictions rely on differences in the underlying models and relative importance of global or local motions for protein unfolding/folding reactions. For non-specific residue interactions, we use the Gaussian Network Model (GNM) to identify folding cores in the limits of two classes of motions, shortly referred to as global and local. The global mode minima from GNM represent the residues with the greatest potential for coordinating collective motions and are explored as potential folding nuclei. Additionally, the fast mode peaks that have previously been labeled as the kinetically hot residues are identified as a second folding core set dependent on local interactions. Finally, a third folding core set is defined by the most stable residues in a simulated thermal denaturation procedure of the FIRST software. This method uses an all-atomic analysis of the rigidity and flexibility of protein structures, which includes specific hydrophobic, polar and charged interactions. Comparison of the three folding core sets to HX data indicate that the fast mode peak residues determined by the GNM and the rigid folding cores of FIRST provide statistically significant enhancements over random correlation. The role of specific interactions in protein folding is also investigated by contrasting the differences between these two network-based computational methods.