GalaxyRefine: protein structure refinement driven by side-chain repacking

1The structure of a protein can be predicted accurately from its sequence by template-based modeling when the sequence identity is sufficiently high (e.g >30%). However, even at a high sequence identity, side-chain structure may be less accurate than the backbone structure, whereas at a lower sequence identity, predicted structures may have significant errors in both side-chain and backbone structures. Although ab initio protein structure predictions from sequences are notoriously difficult, ab initio refinement starting from a reasonable initial model structure is expected to be less difficult. Successful refinement can increase the applicability range of template-based models by providing more precise structures for functional study, molecular design or experimental structure determination.

2Since 2008, various refinement methods have been tested in the refinement category of the community-wide protein structure prediction experiment Critical Assessment of techniques for protein Structure Prediction (CASP). Several methods were shown to improve the initial model structures. Consistent improvements in such refinement experiments is more difficult than the typical refinement tests performed on lower quality initial structures, as the initial structures are selected from the best models submitted by CASP predictors, which have been already refined by other prediction methods.

3In this article, we present a new model structure refinement web server called GalaxyRefine that has shown consistent improvement in CASP10, the most recent CASP held in 2012. GalaxyRefine first rebuilds all side-chain conformations and repeatedly relaxes the structure by short molecular dynamics simulations after side-chain repacking perturbations. Interestingly, this method can improve global and local structure quality. The method can improve global and local structure accuracy as well as physical correctness in 59, 67 and 79% of the CASP10 refinement category targets when measured by GDT-HA, GDC-SC and MolProbity score. This method has been assessed to be more successful in refining the local structure and side-chain quality than any other methods tested in CASP10. GalaxyRefine also provides four additional models generated by relaxation simulations after larger perturbations on secondary structure elements and loops, resulting in larger changes from the initial model structure. GalaxyRefine can improve the models generated by state-of-the-art structure prediction servers such as I-TASSER and ROSETTA when tested on the server models submitted in CASP10.

4GalaxyRefine first rebuilds all side-chains by placing the highest-probability rotamers, starting from the core and then extending to the surface layer by layer. On detecting steric clashes, rotamers of the next highest probabilities are attached. After attaching all side chains, the number of neighboring Cβ atoms is counted around each side chain, and the initial side-chain conformation is recovered if the number deviates from the canonical distribution for the amino acid under the same degree of surface exposure.

5The model with the rebuilt side chains is then refined by two relaxation methods, a mild relaxation and an aggressive one. The lowest energy model of 32 models generated by the mild relaxation is returned as model 1, and four additional models closest to the four largest clusters of 32 models generated by aggressive relaxation are returned as models 2–5. Both of the methods are based on repetitive relaxations (22 and 17 for mild and aggressive relaxations, respectively) by short molecular dynamics simulations (0.6 and 0.8 ps for mild and aggressive relaxations, respectively) with 4 fs time step after structure perturbations. Structure perturbations are applied only to clusters of side chains in the mild refinement, whereas more forceful perturbations to secondary structure elements and loops are applied in the aggressive refinement. The triaxial loop closure method is used to avoid breaks in model structures caused by perturbations to internal torsion angles.

6The energy functions used for the two relaxation methods are linear combinations of a physics-based energy function complemented by database-derived terms and a harmonic restraint energy derived from the given initial model structure. The relative weight of the restraint energy to the physics-based energy for the mild relaxation is five times larger than that for the aggressive relaxation. The physics-based energy function contains CHARMM22-based molecular-mechanics bonded energy terms, Lennard–Jones interaction energy, Coulomb potential energy, FACTS solvation free energy and solvent accessible surface area energy, whereas the database-derived energy function contains hydrogen bond energy, dipolar-DFIRE potential energy and side-chain and backbone torsion angle energy.

7The GalaxyRefine method has been extensively tested on (i) the refinement category targets of CASP8, CASP9 and CASP10 (53 proteins), (ii) Zhang-server (I-TASSER) models (84 proteins) and (iii) ROSETTA server models (69 proteins) for CASP10 template-based modeling targets and (iv) FG-MD benchmark set targets (147 proteins). The test results in terms of improvement of model 1 (and the best refined model out of model 1–5) over initial input models for backbone structure accuracy measured by GDT-HA, side-chain structure accuracy measured by GDC-SC and physical correctness measured by MolProbity score are summarized in Table 1. The GalaxyRefine server shows average improvement in all test cases except for the MolProbity score of ROSETTA models, which have exceptionally good MolProbity scores. Although GalaxyRefine can improve GDT-HA and GDC-SC for all test sets, the average improvements are small (<1 and <3%, respectively), suggesting the necessity for further improvement in this field. Improvement in MolProbity score is relatively larger with an average improvement of 0.6 (from 2.58 to 1.96). Typical MolProbity scores for experimental structures are in the range of 1–2. A successful refinement example is illustrated in Figure 1.

8The GalaxyRefine server runs on a cluster of 4 Linux servers of 2.33 GHz Intel Xeon 8-core processors. The web application uses Python and the MySQL database. The refinement method implemented in the GALAXY program package is written in Fortran 90. The Java viewer JMol (http://www.jmol.org) is used for visualization of predicted structures.

9The only required input is a single-chain protein structure without internal gap in the PDB format. The expected run time is generally 1–2 h. Five refined models can be viewed and downloaded from the website (Figure 2). Information on structural changes obtained by the refinement of the input structure is provided in terms of GDT-HA, RMSD and MolProbity score in a separate table.

10GalaxyRefine is a web server for protein model structure refinement that is particularly successful in improving local structure quality as demonstrated by the tests on CASP refinement category targets and CASP10 server models. On average, it shows moderate improvement in backbone structure quality. The server may be used to refine model structures obtained from available structure prediction methods, including the current best template-based modeling servers.

11National Research Foundation of Korea funded by the Ministry of Education, Science and Technology [2012-0001641, 2011-0012456 and 2012M3C1A6035362]. Funding for open access charge: Seoul National University.

12Conflict of interest statement. None declared.