Use of the UNRES force field in template-assisted prediction of protein structures and the refinement of server models: Test with CASP12 targets

Highlights

• Hybrid (physics- and template-based) approach to protein structure prediction.

• Molecular dynamics with coarse-grained UNRES model, restraints from templates.

• Restraints derived from similar fragments of multiple templates.

• Method able filter out information from poor templates (70% confidence).

• Outstanding prediction obtained for some targets.

Abstract

Knowledge-based methods are, at present, the most effective ones for the prediction of protein structures; however, their results heavily depend on the similarity of a target sequence to those of proteins with known structures. On the other hand, the physics-based methods, although still less accurate and more expensive to execute, are independent of databases and give reasonable results where the knowledge-based methods fail because of weak sequence similarity. Therefore, a plausible approach seems to be the use of knowledge-based methods to determine the sections of the structures that correspond to sufficient sequence similarity and physics-based methods to determine the remaining structure. By participating in the 12th Community Wide Experiment on the Critical Assessment of Techniques for Protein Structure Prediction (CASP12) as the KIAS-Gdansk group, we tested our recently developed hybrid approach, in which protein-structure prediction is carried out by using the physics-based UNRES coarse-grained energy function, with restraints derived from the server models. Best predictions among all groups were obtained for 2 targets and 80% of our models were in the upper 50% of the models submitted to CASP. Our method was also able to exclude, with about 70% confidence, the information from the servers that performed poorly on a given target. Moreover, the method resulted in the best models of 2 refinement targets and performed remarkably well on oligomeric targets.

Introduction

Prediction of protein structure from amino-acid sequence still remains a major challenge of structural biology, because of still insufficient supply of experimental structures and because the methods developed do not provide sufficiently accurate structures in all situations [1]. While knowledge-based methods, which are largely based on sequence-structure similarity, perform well when good templates with sufficiently similar sequences can be found in structural databases, their reliability decreases dramatically with decreasing sequence similarity. The biannual Critical Assessment of Techniques for Protein Structure Prediction (CASP) experiments [2] demonstrate that there are about 10% of targets for which none of the methods available provides a structure close enough to the experimental structure. The physics-based approaches, which are based on finding the structures with the lowest free energy given an appropriate force field and conformational-search method and are, therefore, not database-dependent, could be expected to handle the cases that are not covered by knowledge-based approaches but the existing force fields are not yet sufficiently accurate.

In our recent work [3,4] we proposed a hybrid approach, which combines the best features of the knowledge-based and physics-based methodology. The approach starts from selecting top models from selected servers, which are then converted into distance- and angle-restraints and, subsequently, large-scale multiplexed replica-exchange molecular dynamics (MREMD) [5] simulations with the physics-based coarse-grained UNRES force field (ref 6 and references therein) are carried out. UNRES applied in ab initio mode has already proved to be good in predicting protein structures, scoring considerable success in the past CASP exercises [7,8]. In particular, it was found that UNRES can predict correct domain packing, resulting in models better in overall topology than those found by template-based methods even for targets of the template-based category [7]. Therefore, in our recent work [4], we proposed an approach in which restraints are imposed only on those fragments of the structure that are similar in all server models; in particular, domain packing is unrestrained in this procedure. This fragment-based approach is new compared to those applied in, e.g., MODELLER [9] or MULTICOM [10], in which restraints derived from whole models are imposed. When the models are diverse in general but share common fragments, it can be expected that the structures of these fragments are predicted with much higher confidence in general than the arrangement of these fragments and the rest of the structure. Therefore, it seems reasonable to impose restraints on the common fragments only. Use of a coarse-grained force field is of advantage in this regard because of lower cost of energy and force evaluation and more extensive search of the conformational space, owing to the reduction of the number of degrees of freedom, even though details of the structure are lost. The test results with CASP11 targets suggested that such an approach gives much better results compared to imposing restraints on whole structures. In this work, we tested the approach in the CASP12 experiment, obtaining a number of very good predictions, two of which (for targets T0892 and T0942) were first-ranked and 80% of models being in the upper 50% of models. Apart from the regular prediction (“T0” type targets) category, our KIAS-Gdansk group also participated in the refinement, oligomer, and data-assisted predictions. Our performance in the data-assisted prediction category has been published as a separate paper [11], while this paper reports our results obtained in the regular, refinement, and oligomer prediction category.