Selection of representative structures from large biomolecular ensembles

Abstract

Despite the incredible progress of experimental techniques, protein structure determination still remains a challenging task. Due to the rapid improvements of computer technology, simulations are often used to complement or interpret experimental data, particularly for sparse or low-resolution data. Many such in silico methods allow us to obtain highly accurate models of a protein structure either de novo or via refinement of a physical model with experimental restraints. One crucial question is how to select a representative member or ensemble out of the vast number of computationally generated structures. Here, we introduce such a method. As a representative task, we add co-evolutionary contact pairs as distance restraints to a physical force field and want to select a good characterization of the resulting native-like ensemble. To generate large ensembles, we run replica-exchange molecular dynamics (REMD) on five mid-sized test proteins and over a wide temperature range. High temperatures allow overcoming energetic barriers while low temperatures perform local searches of native-like conformations. The integrated bias is based on co-evolutionary contact pairs derived from a deep residual neural network to guide the simulation toward native-like conformations. We shortly compare and discuss the achieved model precision of contact-guided REMD for mid-sized proteins. Finally, we discuss four robust ensemble-selection algorithms in great detail, which are capable to extract the representative structure models with a high certainty. To assess the performance of the selection algorithms, we exemplarily mimic a "blind scenario,"i.e., where the target structure is unknown, and select a representative structural ensemble of native-like folds.