Note on De Novo Methods

In the absence of experimentally determined crystallographic coordinates, a number of de novo approaches to modeling GPCR structure have arisen over the past decade [8-11, 54]. Unlike comparative or homology modeling, these methods do not employ a known crystal structure as a template. Instead, de novo methods directly address the protein folding problem and attempt to construct a receptor model from the primary amino acid sequence using a combination of empirical and physical energy terms. These procedures generally begin with a coarse- grained description of the helical bundle and use empirical measures, such as hydrophobic moments or residue contact frequencies, to obtain a general orientation of the relative helix positions. In some cases, low-resolution experimental data arising from the 7.5-A frog rhodopsin electron density maps, site-directed spin labeling combined with EPR (SDSL-EPR), or fluorescence resonance energy transfer experiments (FRET) may be included as restraints [ 8, 10] . This is often followed by a higher resolution refinement step that may incorporate rotamer optimization of the receptor side chains, refinement of loop orientations, minimization of the transmembrane helix kinks, and molecular dynamics simulation in the presence or absence of a lipid bilayer. In several cases, de novo models have demonstrated qualitative agreement with experimental mutagenesis results [11, 55, 56], enabled the prediction of new ligand types [57], and allowed calculation of ligand binding affinities with correlation to experimental values [58].

Importantly, the accuracy of de novo methods should be considered in the global context of general protein folding methodologies [59]. In some cases, atomic resolution may be achieved for small, globular, soluble proteins, but in general, de novo model quality has remained poor relative to models built by template - based methods [60] . As few high - resolution seven transmembrane structures exist for de novo method validation, many procedures were evaluated by their accuracy in reconstructing bRho. The Membstruk method of Goddard et al. is reported to reproduce the TM helix backbone heavy atoms of bRho within 2.8Â, while PREDICT yields a RMSD of 2.9Â and Bundler predicts a RMSD of 3.2Â [8, 9, 61]. The most successful of these methodologies is TASSER, which uses a combination of template fragments obtained from the PDB with threading and refinement to obtain a final model 5 11] • TASSER reproduces bRho with an RMSD of 2.1 Â, even when templates with less than 30% sequence identity (SID) are removed from the PDB. However, examination of the TASSER threading results finds that approximately 72% of other modeled GPCRs use either the bRho, sensory rhodopsin, halorho-dopsin, or bacteriorhodopsin structures as a template, suggesting that the best models are obtained from a procedure more akin to homology modeling and refinement, rather that true de novo modeling [11] • Additionally, as in comparative modeling, the success of TASSER still depends strongly on the quality of the template [62]. While the results reported by the above de novo methods are indeed impressive, we note that the crystallographic p2AR and bRho structures exhibit a backbone heavy atom RMSD of 2.1 Â within the TM domains. Thus, even if no improvement was obtained from the starting template, a homology model of p2AR derived from bRho would be within the error reported by the de novo methods. The sequence identity between p2AR and bRho is only 22.8%; for templates with higher sequence identity to the modeled structure, the accuracy of comparative modeling methods is expected to be even greater (Table 15.3). Thus, while de novo methods can provide a useful description of receptor structure, and are critical in refining our understanding of the physical rules governing membrane protein assembly, in the present work, we focus on comparative modeling of other receptor types.

0 0

Post a comment