In silico prediction of protein structures and ensembles

Abstract

Determination of a protein’s structural equilibrium constitution remains a challenge. Experimental techniques like X-ray crystallography or nuclear magnetic resonance spectroscopy either are only able to determine single snapshots of the protein or are not applicable due to the protein's size or dynamics. Orthogonal techniques like electron paramagnetic resonance (EPR) spectroscopy are able to capture all significant populations of the protein but the obtainable data are typically too sparse to unambiguously determine a structural ensemble. Computational methods on the other hand, suffer from necessary simplifications of the structure sampling and free energy evaluation. In order to solve these existing problems, I developed a computational prediction pipeline for protein structures and ensembles that supports incorporation of limited experimental data from EPR spectroscopy and chemical cross-linking. The pipeline encompasses coarse-grained Monte Carlo Metropolis sampling using BCL::Fold, high-resolution refinement using Rosetta, and stability evaluations using molecular dynamics simulations. Novel methods were developed to incorporate the experimental data into the pipeline. Both types of experimental data significantly improved the average accuracy of the sampled models as well as the discrimination between accurate and inaccurate models. In addition, a novel loop sampling algorithm consisting of conformation hashing and cyclic coordinate descent was developed. The algorithm is substantially faster than other available algorithms and samples the conformation of the protein’s major population in 94 % of all cases. The developed methods were applied to determine the structure and dynamics of the Bcl-2-associated X protein (BAX), exotoxin U (ExoU), and the efflux-multidrug resistance protein (EmrE) in conjunction with structural data obtained through EPR spectroscopy.