The computational prediction of accurate protein structures from the amino acid sequence remains a big challenge. In particular, it has been very difficult to refine approximate protein structures obtained by homology modeling to near-experimental accuracy. We are developing an iterative protein structure refinement protocol that involves the combination of efficient sampling methods with the application of scoring functions that are able to identify the most native-like conformations from an ensemble of structures.


A key part to most protein structure prediction and refinement protocols is an efficient conformational sampling method that is capable of generating native-like or near-native conformations. In the context of an iterative refinement protocol, the sampling method has to be able to generate better structures at every cycle in order to be able to make progress towards the native structure. Sampling methods may range from molecular dynamics simulations to lattice-based Monte Carlo sampling of low-resolution models.


The picture above shows the incremental improvement from initial homology models towards the native state. A lattice-based sampling protocol was used, where the structure that is closest to the native at a given refinement cycle is used as the starting structure for the next cycle. With this protocol refinement is possible, but only up to about 2-3 Angstroms after a large number of cycles. We are now developing more efficient sampling methods that converge faster and reach structures within 1-2 Angstroms from the native conformation.


The other key element of a successful protein structure prediction and refinement protocol is a scoring function that can discriminate native-like conformations from non-native, unfolded or misfolded, structures. We are most interested in scoring functions that are based on all-atom force fields and include implicit solvation terms. Such scoring functions are expected to provide good approximations of relative conformational free energies between different conformations with the native structure at the global free energy minimum.


The diagram above shows the application of a physical scoring function to a series of conformation between 1.5 and 5.5 Angstroms. In this case, there is a downward slope towards the native structure with the most native-like structure at the minimum of the scoring function.


In practice, the combination of sampling and scoring methods in an iterative refinement protocol is not as straightforward as it may seem from the above examples. The picture below on the left shows an ensemble of structures generated with a lattice sampling protocol from a single initial structure. Although about 15% of the structures are better than the initial structure at 5.5 Angstroms, the scoring function is too noisy to easily identify the most native-like conformations. We have devised a new correlation-based method that reduces some of the noise and enhances the underlying trends. The result shown below on the right is still noisy, but the a large percentage of the best-scoring structures are now actually better than the initial structure.

Relevant Publications:

Andrew Stumpff-Kane, Michael Feig: A Correlation-Based Method for the Enhancement of Scoring Functions on Funnel-Shaped Energy Landscapes. Proteins (2006) in press

Michael Feig, Charles L. Brooks III: Evaluating CASP4 Predictions with Physical Energy Functions. Proteins (2002) 49, 232-245 PDF