Reply to Nikiforovich et al.: Restoration of the loop regions of G-protein–coupled receptors

We agree with Nikiforovich et al. that the protocol we used in our G-protein–coupled receptor (GPCR) loop calculations (1) differed in some of its details from the protocol they used in their earlier work (2). Our objective in demonstrating that we could restore loops in GPCR crystal structures in good agreement with X-ray crystallographic data was to show that our energy model and sampling algorithms were able to take on this specific task with reasonably good accuracy. We view these capabilities as a necessary, but not sufficient, condition for actually predicting GPCR loops in a realistic application, such as homology modeling. By taking an incremental approach, our intention is to separate the problem into different components: energy model, localized sampling algorithm, and fully delocalized sampling algorithm (i.e., capable of accurate structural prediction when starting with a model that potentially has errors in all regions of the structure, as opposed to starting with a perfect structure and restoring the loop in this context). Our article has provided encouraging results for two of these three component—the energy model and localized sampling algorithm; we are presently working on the third. We reported the results of Nikiforovich et al. (3) in our study for one main reason: to show how difficult the full GPCR loop prediction problem is. This should have been made clearer in our article.

With regard to the remaining points brought up by Nikiforovich et al. (2), we have the following responses:

First, we do not enforce experimental disulfide bonds; the possibility of disulfide bonding is present in the model, but formation of such bonds occurs during predictions when the cysteine sulfurs get sufficiently close in distance.

Second, it is true that Nikiforovich et al. (3) defined their “loop” regions as, in general, being slightly larger than ours, typically one to three residues. However, based on previous experience with our methods, we do not believe these differences would cause significant changes in our prediction accuracy, as the loop length differentials are not large.

Third, we do not agree with the assertion that “the lowest of the rmsd values calculated for the set of predicted low energy conformations may serve as a more adequate measure for success of the predictions made by modeling” (2). In our view, using the lowest energy structure as the final prediction for the crystal structure is a highly significant test of the energy and solvation models used, as well as the capabilities of the sampling algorithm. This is a standard view in the field, and we do not see any argument made by Nikiforovich et al. (2) as motivating us to abandon it.