The Origin of Consistent Protein Structure Refinement from Structural Averaging

Abstract

Recent studies have shown that explicit solvent molecular dynamics (MD) simulation followed by structural averaging can consistently improve protein structure models. In this study, we investigate the origin of improvements from averaging. We first show that improvement upon averaging is not limited to explicit water MD simulation, as consistent improvements are also observed for more efficient implicit solvent MD or Monte Carlo minimization simulations. We next examine the changes in model accuracy brought about by averaging at the individual residue level, and find that these changes are correlated with the extent of residue fluctuation. Residues undergoing large fluctuations tend to diverge, and these deviations are dampened by averaging. Residues undergoing medium sized fluctuations often improve significantly, and averaging amplifies these improvements. These observations are consistent with an energy landscape model in which the magnitude of the energy gradient towards the native structure decreases with increasing distance from the native state.

High‐accuracy refinement using Rosetta in CASP13

“…Model 1 from our human prediction has a GDT‐HA of 93 and Cα‐RMSD of 0.48 å. Post analysis on this target indicated a complementary role of Rosetta and explicit water MD as observed previously: while Rosetta successfully recovered a critical error at the C‐terminus and side‐chain orientations around it, MD successfully brought the positions of backbones and side‐chains even closer to the crystal structure. Our MD refinement alone on the input model did not result in this level of accuracy (GDT‐HA of 83 and RMSD 1.6 å).…”

CASP11 refinement experiments with ROSETTA

“…Core‐refinement is where recent progress has been found from MD‐based approaches . As we recently described, core‐refinement methods primarily improve residues that are nearly correct, but do not substantially improve regions with significant errors. Rosetta sampling methods could be useful for refining along the two other axes and so complement the limitations in core‐refinement approaches.…”

Protein structure refinement via molecular‐dynamics simulations: What works and what does not?

“…As described in detail previously, this success was attributed to a number of factors: extensive sampling with 30 × 20 ns = 600 ns per target under restraints to prevent large deviations from the initial models, a recently refined version of the CHARMM force field, generation of ensemble averages rather than selection of a single structure, and the use of quality assessment filters to remove decoy sets where scoring was likely not going to discriminate native‐like from non‐native structures. The averaging was especially important since it reproduces the ensemble averaging in experiment but also amplifies recurring native‐like features in a large set of structures over non‐native elements as discussed in detail recently . Achieving consistency in refining template‐based models was a significant milestone, but the extent of refinement remained rather modest with an average of 2.6 GDT‐HA units and a maximum improvement by 6.5 GDT‐HA units for model 1 submissions.…”