Portal de Periódicos da CAPES

Folding a Protein in the Computer

2005; Elsevier BV; Volume: 13; Issue: 4 Linguagem: Inglês

10.1016/j.str.2005.03.005

1878-4186

Angel E. Garcı́a, José N. Onuchic,

RNA and protein synthesis mechanisms

In this issue of Structure, Herges and Wenzel, 2005Herges T. Wenzel W. Phys. Rev. Lett. 2005; 94: 018101-018104Crossref Scopus (35) Google Scholar describe a structure-based force field that, when combined with a stochastic optimization method, a modified basin hopping method, can fold α-helical proteins. Although limited to α-helical structures, this approach further supports that predicting protein structures in a computer is becoming a reality. In this issue of Structure, Herges and Wenzel, 2005Herges T. Wenzel W. Phys. Rev. Lett. 2005; 94: 018101-018104Crossref Scopus (35) Google Scholar describe a structure-based force field that, when combined with a stochastic optimization method, a modified basin hopping method, can fold α-helical proteins. Although limited to α-helical structures, this approach further supports that predicting protein structures in a computer is becoming a reality. Developing a protein model able to predict protein structures has been a challenge in theoretical structural biology. The challenge becomes even larger if one is also interested in understanding the entire folding mechanism and protein landscape. A report in this issue of Structure by Herges and Wenzel works toward this goal. Their computation method is based on the thermodynamic assumption that the folded structure is determined by the global energy minimum. From these simulations, they claim that, in addition to predicting the folding conformation, they are able the characterize the entire protein folding funnel (Onuchic and Wolynes, 2004Onuchic J.N. Wolynes P.G. Curr. Opin. Struct. Biol. 2004; 14: 70-75Crossref PubMed Scopus (952) Google Scholar). In the particular case presented in the manuscript, they fold the 36 amino acid villin headpiece, and compare the complexity of its folding to that of a similar three-helix bundle, the 40 amino acid HIV accessory protein ( 1F4I ). Previous calculations by the same authors (Herges and Wenzel, 2005Herges T. Wenzel W. Phys. Rev. Lett. 2005; 94: 018101-018104Crossref Scopus (35) Google Scholar) and others have been able to fold other α-helical structures. In this article the authors show that their potential energy function and sampling method can predict the structure of α-helical structures. All atom simulations using explicit solvent have been able to describe kinetics of folding (e.g., Pande and collaborators [Snow et al., 2004Snow C.D. Gai F. Hagen S.J. Pande, V.S. Proc. Natl. Acad. Sci. 2004; 101: 4077-4082Crossref PubMed Scopus (177) Google Scholar]), unfolding (e.g., Daggett and collaborators [Mayor et al., 2003Mayor U. Guydosh N.R. Johnson C.M. Grossmann J.G. Sato S. Jas G.S. Freund S.M.V. Alonso D.O.V. Daggett V. Fersht A.R. Nature. 2003; 421: 863-867Crossref PubMed Scopus (412) Google Scholar]), and folding/unfolding equilibrium of proteins, but have correctly claimed not to have the ability to predict. One exception has been the prediction of the structure of a designed trp cage mini protein by Simmerling et al., 2002Simmerling C. Strockbine B. Roitberg, A. J. Am. Chem. Soc. 2002; 124: 11258-11259Crossref PubMed Scopus (523) Google Scholar. The advances in the development of energy functions with predictive power and the development of efficient sampling methods have opened the door for physically based protein structure prediction—in contrast to knowledge-based prediction methods which have dominated the field of structure prediction (Bradley et al., 2003Bradley P. Chivian D. Meiler J. Misura K.M. Rohl C.A. Schief W.R. Wedemeyer W.J. Schueler-Furman O. Murphy P. et al.Proteins. 2003; 53: 457-468Crossref PubMed Scopus (151) Google Scholar). Physically based methods may have the advantage of being able to predict protein structure as a function of solvent conditions, such as pH, ionic strength, cosolvents, temperature, and pressure. Testing of physically based models require extensive sampling and cannot be extended easily to larger systems. In addition, information about the entire landscape of the protein is needed in order to properly evaluate protein potentials! Solving the protein folding problem from physically based methods depends on the development of a force field that can fold proteins from sequence information, and a robust sampling method that enables the sampling of the energy landscape. These two effects cannot be solved independently. On one hand, we would like to represent all the details of the atomic interactions–including explicit solvent (Garcia and Onuchic, 2003Garcia A.E. Onuchic J.N. Proc. Natl. Acad. Sci. USA. 2003; 100: 13898-13903Crossref PubMed Scopus (302) Google Scholar), polarizable force fields (Ponder and Case, 2003Ponder J.W. Case D.A. Adv. Protein Chem. 2003; 66: 27-85Crossref PubMed Scopus (1364) Google Scholar), or ab initio. On the other hand, we want to be able to sample quickly—which implies using implicit solvent models, unified residue representation, constrained degrees of freedom, etc. Herges and Wenzel opted for simplifying the force field representation, while maintaining atomic detail. Their optimization method and model is orders of magnitude faster than direct simulation. Their potential energy function is physically based, although there is some knowledge-based bias in the sampling, since the backbone dihedrals are partially biased to sample configurations in the Ramchandran map found in crystal structures. This approximation might be crucial for the success of the method, since it has been shown that dihedral potential energy terms in most force fields are not well parameterized. Also, Takada, 2001Takada S. Proteins. 2001; 42: 85-98Crossref PubMed Scopus (47) Google Scholar was able to fold many α-helical proteins by adding a Ramachandran bias potential to the knowledge-based potential energy function used by Wolynes' group (Eastwood et al., 2003Eastwood M.P. Hardin C. Luthey-Schulten Z. Wolynes P.G. J. Chem. Phys. 2003; 118: 8500-8512Crossref Scopus (13) Google Scholar)—and could not fold without this potential. Attempts to parameterize similar biased potential in all atom simulations have been done by Feig et al., 2003Feig M. MacKerell Jr., A.D. Brooks III, C.L. J. Phys. Chem. B. 2003; 107: 2831-2836Crossref Scopus (180) Google Scholar. Searching for the global energy minimum is a difficult task, since the number of configurations is extremely large and there is no way to determine that an incomplete search has found a global minimum, except in special cases. For example, there are ways to determine lower bounds of the free energy—if a minimum found has the same energy as the lowest bound, then a global minimum has been found. However, this does not exclude the existence of other minima with the same low energy. Exhaustive searches have been conducted for short peptides. An implicit assumption in these searches is that an energy function of atomic interactions can represent a free energy. This assumption excludes the contribution from configurational entropy (at best, within the harmonic approximation, it assumes that all energy wells have the same curvature). It also excludes the possibility of degeneracy—the density of states also contributes to the entropy. So, at best, the employed energy function deals with a reduced free energy function that includes, implicitly, solvent entropy effects. Such assumptions can be pathological and will limit the outcome of such an approach. One such limitation has been encountered by the authors when the tree search cannot resolve highly populated trees within an energy threshold. The optimized potential (or free) energy function employed has been used for many small α-helical proteins. The authors point out that the potential might be able to fold larger α-helical proteins, but the sampling would require enormous sampling not accessible to them currently. The emphasis on the fact that their potential energy function might be able to fold any α-helical protein highlights what is not said—the potential energy function fails to fold β sheet structures. This problem is also present in all atom simulations and other knowledge-based potential energy functions. Attempts to solve this problem have developed multibody cummulant expansions of the interaction between amino acids in a chain (Liwo et al., 2001Liwo A. Czaplewski C. Pillardy J. Scheraga H.A. J. Chem. Phys. 2001; 115: 2323-2347Crossref Scopus (197) Google Scholar, Eastwood et al., 2003Eastwood M.P. Hardin C. Luthey-Schulten Z. Wolynes P.G. J. Chem. Phys. 2003; 118: 8500-8512Crossref Scopus (13) Google Scholar). The origin of this multibody interaction potential could be found in solvent-mediated interactions (Papoian et al., 2004Papoian G.A. Ulander J. Eastwood M.P. Luthey-Schulten Z. Wolynes P.G. Proc. Natl. Acad. Sci. USA. 2004; 101: 3352-3357Crossref PubMed Scopus (251) Google Scholar). Explicit solvent simulations of protein folding equilibrium on protein A (Garcia and Onuchic, 2003Garcia A.E. Onuchic J.N. Proc. Natl. Acad. Sci. USA. 2003; 100: 13898-13903Crossref PubMed Scopus (302) Google Scholar) showed that protein desolvation, helix formation, and folding occur cooperatively and in synchronization. Although challenges remain, clear progress has been made toward the ultimate goal of protein folding prediction.