Optimization of Predicted Spatial Restraints on a Coarse- Grained Protein Model
1 – Dept. of Computer Science, 2 – Dept. of Biological Sciences, Purdue University, West Lafayette, IN, USA
It has been shown that in many cases the recent generation of fold recognition methods can capture at least structural fragment information even when the global structure can not be reliably predicted1. Assembling structure fragments detected by a fold recognition method is one of the common ways for ab initio/de novo protein structure prediction1-3. In the CASP5, it was reported that several consensus methods or meta-server approaches4,5 showed high performance6. Based on these two observations, our approach developed for CASP6 is an optimization of predicted spatial restraints calculated by various servers, including servers participating in CAFASP4 using a coarse-grained protein model.
A protein is represented by a simplified model which explicitly specifies positions of alpha carbons in the main chain7. A conformation of this Ca model is defined by a set of rotational and hinge angles between adjacent alpha carbons. Information of predicted structures of a target protein by various methods is used in the following way: (1) The predicted structures are clustered globally and locally. (2) The distribution of inter-residue distances and angles are calculated and subsequently used as soft restraints8 in the next refinement step. Starting from several initial structures, the conformation of the model is refined so that is satisfies these soft restraints by a Monte Carlo optimization with the Metropolis criteria. Consensus prediction of the secondary structures is also used. Usually a conformation converges relatively quickly since the method uses a large number of restraints.
During the course of the development, we phased in statistics of the structure preference of known structures in PDB as penalty terms of spatial restraints to avoid “non-protein-like” conformations7. These terms include minimum distance between Ca - Ca , peptide bond - peptide bond, and Ca - peptide bond distances as well as hinge angle restraints.
Suggestions from our function prediction team were often a great help in the final model selection. Our three teams, the structure, function (B213-207Func), and domain prediction (B213-207Dom) teams worked in a coordinated manner. Although this method is still in an early stage of the development, the performance will surely improve as additional scoring terms are incorporated.
1. Kihara,D., Lu,H., Kolinski,A. & Skolnick,J. TOUCHSTONE: an ab initio protein structure prediction method that uses threading-based tertiary restraints. Proc Natl Acad Sci U S A 98, 10125-30 (2001).
2. Jones,D.T. Predicting novel protein folds by using FRAGFOLD. Proteins Suppl 5, 127-32 (2001).
3. Bonneau,R. et al. De Novo Prediction of Three-dimensional Structures for Major Protein Families. J Mol Biol 322, 65. (2002).
4. Fischer,D. 3D-SHOTGUN: A novel, cooperative, fold-recognition meta-predictor. Proteins 51, 434-41 (2003).
5. Ginalski,K. & Rychlewski,L. Detection of reliable and unexpected protein fold predictions using 3D-Jury. Nucleic Acids Res. 31, 3291-3292 (2003).
6. Kinch,L.N. et al. CASP5 assessment of fold recognition target predictions. Proteins 53 Suppl 6, 395-409 (2003).
7. Kolinski,A. Protein modeling and structure prediction with a reduced representation. Acta Biochim. Pol. 51, 349-371 (2004).
8. Sali,A. & Blundell,T.L. Comparative protein modelling by satisfaction of spatial restraints. J Mol Biol 234, 779-815 (1993).