Combining molecular dynamics and machine learning to improve protein function recognition.

As structural genomics efforts succeed in solving protein structures with novel folds, the number of proteins with known structures but unknown functions increases. Although experimental assays can determine the functions of some of these molecules, they can be expensive and time consuming. Computational approaches can assist in identifying potential functions of these molecules. Possible functions can be predicted based on sequence similarity, genomic context, expression patterns, structure similarity, and combinations of these. We investigated whether simulations of protein dynamics can expose functional sites that are not apparent to the structure-based function prediction methods in static crystal structures. Focusing on Ca2+ binding, we coupled a machine learning tool that recognizes functional sites, FEATURE, with Molecular Dynamics (MD) simulations. Treating molecules as dynamic entities can improve the ability of structure-based function prediction methods to annotate possible functional sites.