In summary, the proposed simulation protocol makes possible fast and reliable assessment of high resolution structures for relatively large proteins. Bradley et al. [2] have shown recently that a high resolution de novo structure prediction can be achieved for small proteins by using an all-atom refinement procedure in the last stage of prediction. The cost of application of the high resolution refinement for large proteins was estimated by authors to require orders of magnitude more computing power, than the 150 CPU days required for small proteins [2]. A single, 500-step, minimization sufficient for the ranking of the smallest proteins (of a comparable size to those from Bradley et al. work) took in present work approximately 1.5 minutes. The approach described here may be a good alternative for the refinement of small proteins and could be applied as a means of the best model selection in a large scale modeling, regardless of protein length.
Protein model filtering in the endgame of protein structure prediction protocols faces the following two challenges: fold identification (particularly in de novo modeling) and the selection of the best models from a set of good models (especially important in comparative modeling). The results of this work apply mostly to the model selection in comparative modeling. Recent CASPs test have shown that the best comparative models are built with a lot of human intervention using an assortment of well known modeling tools [1]. The challenge is to automate the protein prediction and produce even more accurate models with no need of the human assistance during the prediction protocol. Thus, consistent and accurate last stage of modeling is needed, i.e., producing and filtering the high-resolution predictions. Elaborate human intervention can be compensated by a high-throughput modeling, employing sampling of alternative alignments (which should lead to a sufficient sampling of the near-native regions of the conformational space) and an efficient scoring of the large number of obtained decoys.
The need for a reliable detection of the best native-like models from a set of different predictions produced in the recent CAFASP4 experiment (CAFASP assesses the performance of methods without the user intervention allowed in CASP) led to a new category of model assessments – Model Quality Assessment Programs (MQAPs). Although performance of currently available MQAPs indicate that some of these methods may be useful for new automated procedures, high false positive rates are observed, and the MQAP methods were suggested to be used as additional elements of the prediction protocols rather than as a simple post-filter [31].
Our method gives surprisingly consistent correlation of the all-atom energies with RMSD distance from the native structure. Decoys within 3 Å from native were examined and no false positive cases were noted. As mentioned before, the correlation of the CABS energy with RMSD in this range is often insignificant. The starting energy of the all-atom structures are uncorrelated with the CABS energy and are extremely high, mostly due to the overlaps of the side chains. The Molecular Mechanics energy decreases rapidly during the initial stage of minimization, mainly due to the improvement of side-chain rotamers position. For a fraction of decoys (see Table 1) the energy plateaus at a high level due to non-resolvable steric clashes. The subset of the low-energy structures, easily distinguishable from the high-energy ones, exhibits the above mentioned, nearly perfect correlation with the values of RMSD from the native structure. This provides a very strong support for the idea of multiscale high-resolution protein modeling. More extensive molecular dynamics simulation, than described here, might lead to even better model ranking and refinement. In the case of fixed Cα-traces, a longer than performing 1000 iterations minimization is not necessary – the results do not change anymore.
Finally, we would like to return to the two important assumptions of the present method: fixed positions of the alpha carbons during the minimization and the in vacuum Molecular Mechanics. Obviously, these assumptions significantly reduce the cost of computations for large sets of decoys. The frozen alpha carbon approximation works very well for the model ranking, although it eliminates possibility of a significant refinement of the entire structures. Model ranking exercises performed on the MOULDER set by others have shown clearly that there is very little added value with use of more rigorous Molecular Mechanics procedures [12]. The same conclusion could be drawn from our experiments with defrosted alpha carbons and with a continuous model of solvent (unpublished). The ability to generate sets of decoys containing a significant fraction of the near-native structures by coarse-grained modeling (followed by the all-atom refinement and model ranking) remains a key factor for the high-accuracy structure prediction.
Answers to all your Biology Questions
