Simulations can predict small protein structures

Proteins are the small, underappreciated and occasionally orange Oompa-loompas of all biological systems. These nano-sized linear chains of amino acids "fold" up in complex, globular shapes by way of poorly understood mechanisms to perform almost every biological job imaginable, sometimes individually and sometimes in groups or complexes.

And just as Willy Wonka would only employ Oompa-loompas because of the risk of industrial espionage and used them for each candy-making process in his factory, you utilize proteins for all of your cellular processes and they function to protect you from diseases, metabolize your food and manage everything else your body does.

But the process of protein folding remains, in many ways, frustratingly mysterious. For decades, researchers have explored the problem in a variety of ways, primarily computationally due to the difficulty of studying such small-scale systems experimentally. How can it be that any linear chain of amino acids in a specific sequence always folds autonomously into the same unique, three-dimensional (3D) shape?

The problem is like watching hundreds of sheets of paper spontaneously assemble into cranes, and nothing else. While the driving forces behind protein folding are well-studied (and paper, by contrast, doesn't fold itself under any known conditions), given the infinite number of shapes that a chain of amino acids could assume, it is unclear why each protein has an ideal conformation into which it always folds, or a "native state." Understanding this is integral to a full understanding of how proteins work.

One of the most recent advances in understanding protein folding mechanisms is the large-scale simulation study conceived by Vincent Voelz, Scott Shell and Ken Dill at the University of California, San Francisco, and described in PLoS Computational Biology as "Predicting Peptide Structures in Native Proteins from Physical Simulations of Fragments." This study examines the extent to which conformations of peptide fragments, or short amino acid sequences lacking stable 3D structure, can predict conformations of proteins.

"An unfolded polypeptide has?an enormous conformational search problem?to solve, yet many proteins can fold in only a few?milliseconds, and we are?still trying to understand how this happens," Voelz said.

The simulation serves primarily to show the role of local structure in protein folding, without the effects of the various forces believed to drive higher-order assembly of proteins.

"Our work sheds light on how much [of the] conformational search problem can be solved locally, without long-range cooperative interactions. By performing fragment simulations in water, we can decouple the local structuring of peptide chains from the long-range tertiary interactions necessary for folding," Voelz added.

Voelz and his colleagues went about this goal by performing replica exchange molecular dynamics (REMD) simulations of several hundred peptide fragments of three different sizes from 13 different proteins using force fields. A force field is a large set of parameters for describing the potential energy of a system; all-atom sampling provides information for each atom in a system. REMD simulations have been used for all-atom sampling in the past, for identifying regions of protein structures about which the folding might occur.

They used the Zipping and Assembly Method (ZAM) for highly efficient searching, which was previously developed to investigate important parts of a conformational space at a high speed. The ZAM strategy of protein folding has also been shown to fold local structures using a physics-based simulation and then combine those structures to form stable assemblies.

The REMD simulations using ZAM yielded results that showed that small peptide fragments adopt conformations in solution that are extremely similar to the conformations that they adopt in their native proteins. But there are limitations in using all-atom force fields for predicting protein structures, most significantly that they require a lot of computational power.

"Currently physics-based simulation is too computationally expensive to compete with the best strategies for?ab initio??protein folding in the biannual CASP assessment, which involve getting good sequence alignments to known template structures. However, it is generally thought that all-atom physics-based models will be increasingly necessary for modeling the details of protein conformation, dynamics and ligand binding," Voelz said.

For very large data sets, such as the one in this study, the use of an all-atom force field shows how predictive the simulation really is for predicting local protein structures.

"Local structuring is very interesting in terms of folding mechanisms, of course, but also, the results of our analysis suggest that perhaps we can devise clever algorithms that exploit the local nature of folding," Voelz said.

The researchers analyzed the simulations by computing a variety of metrics based on contact within the protein structures and applied statistical classifier methods for determining which metastable contacts yielded from the simulations were present in the native structures of the proteins.

"We chose to use contact-based metrics, because much previous work has shown the importance of contact formation in folding. For example, protein folding rates are strongly correlated to the topology of the native state, which suggests that finding native-like long-range contact formation is the rate-limiting step in folding," Voelz said.

"Also, we had good results in the past using these contact-based metrics to predict structures from simulations. They are natural coordinates for capturing hydrophobic interactions, for instance, which are very important in determining protein structure and folding pathways."

The metrics applied by the team took into account many aspects of peptide folding. The metric of contact probability, which they found to be the most predictive of a peptide's native structure, can be described as "the equilibrium probability of a given contact." Contact probability was calculated as the fraction of sampled states that had inter-residue distances less than some cutoff distance (in this study, eight angstroms).

Other metrics considered were distance profile score, which was designed to get information about the interaction of different residues as a function of distance; mutual stability score and mutual cooperativity score, which were designed to characterize the average extent of cooperative interactions for any contact; and mesoentropy score, which was related to the peptide backbone entropy, or the diversity of conformations of the peptide at equilibrium.

Interestingly, the model was less predictive when combinations of metrics were used, indicating that metrics other than contact probability did not play a major role in predicting local peptide conformation.

"Cooperativity between multiple contacts, at least for the small fragments we examined, does not provide very much predictive power beyond that of single contacts," Voelz said.

Could there be other metrics that were not considered that are more predictive? Voelz claims that although this is possible, they have put great effort into eliminating noise in the data that is irrelevant to the physics of protein folding.

"Our starting point is to consider the results of our molecular simulations at face value to probe how predictive local structuring is in predicting native structures. Using [molecular dynamics] simulation, which use transferrable forcefields developed over decades of research, we find that the structuring we see is somewhat predictive," Voelz said.

But there is some noise. The results show that this methodology still produces many false positives and false negatives, which is inevitable when using physics-based simulation models with limited sampling.

Hard-to-get experimental data would be required to ultimately verify this and other simulation studies. "It would be great to get atomic-resolution experimental data for small peptide fragments to compare with our simulations - this would be the first step in assessing the quality of our simulations," Voelz said.