Motivation: Protein dynamics is believed to influence protein function through a variety of mechanisms, some of which are not fully understood. Thus, prediction of protein flexibility from sequence or structural characteristics would assist in comprehension of the ways dynamics is linked to function, and would be important in protein modeling and design. In particular, quantitative description of side-chain dynamics would allow us to understand the role of side-chain flexibility in different functional processes, such as protein-ligand and protein-protein interactions.
Results: Using a dataset of 18 proteins, we trained a neural network for the prediction of methyl-bearing side-chain dynamics as described by the methyl side-chain generalized order parameters (S(2)) inferred from NMR data. The network uses 10 input parameters extracted from 3D structures. The average correlation coefficient between the experimental and predicted generalized order parameters is r = 0.71 +/- 0.029. Further analysis revealed that the order parameter depends more strongly on the methyl carbon packing density, the methyl carbon distance to the C(alpha) atom, and the knowledge-based pair-wise contact potential between the methyl carbon and neighboring amino acids. In general, we observed an improvement in the prediction of methyl order parameters by our network in comparison with molecular dynamics simulations. The sensitivity of the predictions to minor structural changes was illustrated in two examples (calmodulin and barnase) by comparing the S(2) predictions for the unbound and ligand-bound structures. The method was able to correctly predict most of the significant changes in side-chain dynamics upon ligand binding, and identified some residues involved in long-range communications or protein-ligand binding.
Availability: http://epigenomique.genopole.fr/ approximately carbonell