Decoy data set

Structural Data on Protein-Ligand Complexes and Decoy Data Sets... 

Two additional decoy data sets of misfolded proteins [17] and of predicted protein structures from the Critical Assessment of Techniques for Protein Stmcture Prediction (CASP) [67] are also used to illustrate the method and its utility. Individual components of the energy perform worse than the total energy for example, for the bulk of the well-packed decoys, the van der Waals energy provides very little information about structural similarity between a well-packed non-native structure and the native state. It is also shown that some aspects of the SGB model results can be mimicked by a screened electrostatic energy, although the SGB approximation provides a better discriminatory measure between non-native and native states. 

Examination of minimized energies for the seven extensive data sets of protein decoys (see Fig. 1) shows that using the OPLS-AA/SGB potential, no decoy scores better than the X-ray structure. The correlation between stractural similarity and score is strong only for structures with low RMSD. For RMSD > 4 A this correlation breaks down. Native-like states appear around 2 A at low energies, with the bulk of the decoys being in non-native-like conformations with RMSD above 4 A. 

Figure 4. The distribution of energy gaps from the native for the 3icb data set of the Park and Levitt decoys using various energy functions.

Physics-based all-atom molecular mechanics force fields have not been generally considered practical for fold detection because they are parameterized on small molecule data rather than on proteins directly the level of atomic detail contained in these models is considered poorly matched to the fold detection problem with respect to both accuracy and computational cost. Recent studies have shown, however, that a scoring function based on the potential energy from an all-atom molecular mechanics force field can recognize native protein conformations among a set of decoys as well as the best available knowledge-based scoring functions [6]. 

---