Overview of refinement protocol

The overall method for protein structure refinement used in CASP13 is based on our earlier refinement strategy of averaging MD-generated structure ensembles^{14,16,19,20,25,26} but the protocol has been updated by incorporating new features. As outlined in Figure 1, the CASP13 refinement protocol consisted of three stages: pre-sampling, sampling, and post-sampling. In the pre-sampling stage, initial models were subjected to locPREFMD²⁷, our local refinement method, to quickly remedy stereochemical errors that can potentially cause abnormal MD sampling afterward. We conducted protein refinement with ligands if there were putative binding ligands. Therefore, an additional pre-sampling step involved the prediction and docking of ligand conformations where appropriate. Starting from the processed initial model, two protocols were followed in the sampling stage. The iterative protocol consisted of three iterations of MD simulations with flat-bottom harmonic restraints, followed by clustering, and selection of new initial models for subsequent rounds of sampling. Altneratively, we also applied a conservative protocol analogous to our CASP12 refinement method^20,26, but with reduced simulation time. The refined model from the iterative protocol was submitted as “Model 1” when we had confidence that it was successful. Otherwise, the refined model from the conservative protocol replaced it, and the iterative protocol was terminated after the first iteration. Based on benchmark tests on CASP10–12 targets, we did not have confidence that our new, more aggressive protocol would succeed in the case of bigger targets (with a radius of gyration of the initial model greater than 17 Å), targets with highly unstable initial models, or with targets involving bound ligands. The initial model stability was determined from the ratio of snapshots deviating less than 1 Å Cα-RMSD from the initial model with the conservative protocol. A model was considered highly unstable if the fraction of structures remaining close to the initial model was less than 0.1%. After the sampling stage, the generated conformations were evaluated with the Rosetta score (ref2015).²⁴ Based on the score, a subset of selected structures were selected for each protocol and subjected to averaging and final stereochemical refinement. We submitted five models for each target including local error estimations as described below. The refined models from the main and the conservative protocol were submitted as either “Model 1” or “2” as described above, and ensemble averaged structures of the three largest clusters from the iterative protocol comprised the remainder of the submitted models. Further details of the refinement protocol are described in the following.

Overall refinement protocol applied by FEIGLAB group in CASP13.