Transition from structure to mechanism is the next frontier. While time-averaged crystallographic snapshots frame biochemical and functional data in a structural context, achieving a mechanistic description of biological function requires an understanding of dynamics, the fourth dimension of protein structure. The function of channels, transporters, and receptors is intimately associated with their ability to execute movements that enable opening of a gate, alternate access of a substrate binding pocket to different sides of the membrane, or exposure of signaling sequences. Excursions between these conformers can be thermally activated; a view in stark contrast to the static picture communicated by crystal structures. In some cases, models of conformational changes can be inferred from a “patchwork of different homologs fortuitously crystallized in different states” (Krishnamurthy et al., 2009), but the caveat is that the observed distribution of structures may reflect “the idiosyncrasies of the different homologs” (Rees et al., 2009) rather than different intermediates in the functional cycle. Protein dynamics and conformational sampling can be altered by the crystallization process. Crystal contacts can act as a conformational selectivity filter distorting highly flexible but functionally critical segments and/or stabilizing conformations that may be sparsely populated in solution. Moreover, membrane proteins’ natural milieu is the lipid bilayer, which differs in its physico-chemical properties from detergent micelles, the preferred crystallography solvent. Accentuating this concern, detergent selection criteria often emphasize crystal and diffraction qualities at the expense of functional considerations thus dictating the use of harsh detergents. Together these factors may conspire to cloud the mechanistic interpretation of crystal structures (Cross et al., 2011). A detailed understanding of membrane protein functional cycles requires a description of the nature, amplitude and time scale of conformational equilibria and/or triggered conformational changes in a native-like environment.

The initial work consisted of extensive nitroxide scanning that targeted loop regions to simultaneously define the membrane boundary and map light-induced conformational changes (Reviewed in Hubbell et al., 2003). Changes in spin label mobility, and thus tertiary contacts, indicated the location, direction, and type of movements that underlie photoactivation—most significantly a rigid-body motion of TM6 away from the bulk of the protein. To better define the magnitude of this and other helical motions, Hubbell and colleagues measured distances by dipolar coupling between spin labels. In all of the short-range studies, spin labels had to be introduced on the inside of the protein in order to ensure sufficient proximity for dipolar coupling. Labeling buried positions likely biased the rotamers adopted by the label and possibly altered the local structure hindering a quantitative interpretation of the helix movement.

The EPR constraints were in reasonable agreement with the crystal structure of Rhodopsin in the inactivated state (Palczewski et al., 2000). Subsequently, Salom et al. (2006) reported the crystal structure of the putative photoactivated intermediate, achieved by the illumination of dark state crystals. Surprisingly, this structure was very close to the dark state with minimal movement of helices, initiating a controversy that continued until recently. However, the authors’ selection of crystallization conditions that preserved the crystalline lattice upon illumination likely suppressed the conformational flexibility of rhodopsin.

Although a benchmarked strategy for EPR-based modeling of structures is not yet available, simplifying approximations and assumptions have been applied to the systems reviewed here, and elsewhere in the literature, to successfully model structure and conformational changes with outcomes subsequently verified by other techniques. In most cases, the sampling of conformational space was restricted by prior data that suggested particular folds or motifs (Koteiche et al., 1998; Poirier et al., 1998), by structural simplicity such as small size and high symmetries (Cortes et al., 2001), or by restriction on the type of motion to rigid body or simple helix rotation (Altenbach et al., 2008; Perozo et al., 1999).

The more general question of whether EPR restraints restrict conformational space to a set of convergent models of acceptable resolution or enable detailed description of structural rearrangements starting from a high resolution structure has only been recently addressed. Alexander et al. (2008) carried out a systematic feasibility analysis of de novo protein structure determination from EPR restraints in T4L. This study also aimed to directly define the information content of EPR restraints and the impact of the sparse data on model quality. The distance between spin labels was converted into a distance range between β-carbons using a simple “motion on a cone” model, treating the spin label as an average vector relative to the β-carbon. Because of an assumed isotropic distribution of the label in this model, the function relating the distance between the two spin labels to that between the corresponding β-carbon was relatively broad, i.e. the derived restraint has large uncertainty. This study made two novel contributions. First, it heralded the use of the Rosetta folding algorithm (Das and Baker, 2008) as an alternative computation platform to MD simulations. Second, it demonstrated that a detailed model of the spin label conformations at each site may not be required. Even with a simple boundary function to interpret the restraints (Figure 10), 25 EPR restraints were sufficient to generate models with the correct fold. Subsequent high resolution refinement yielded structures that are within 1 Å RMSD from the crystal structure. This remarkable outcome was rationalized by the robust Rosetta knowledge-based energy function, which captures the principle of protein assembly encoded in known structures, compensating for the sparseness of EPR restraints. In turn, the EPR restraints efficiently restrict conformational space enabling Rosetta to find the global energy minimum.

This study set the stage for an analysis of the information content of EPR restraints. Alexander et al. demonstrated the importance of high information content (defined as the ratio between sequence separation and Euclidean distance) as a criterion for restraint quality. The improvement in model quality by Rosetta folding was attributed to 8 (of 25) restraints with the highest information content. Because throughput of EPR methods and the ensuing restraint sparseness is defined by the number of pairs to be constructed, a rational approach for the selection of spin labeled sites is required. Kazmier et al. (2010) developed an algorithm for selection of distance restraints with optimized information content. An optimal algorithm was generated, and its applicability was experimentally established through prediction of a set of pairs for spin label incorporation, experimental determination of the distances and then restrained Rosetta modeling of the T4L C-terminal domain. Improvement in model quality required a limited number of restraints determined by the pairwise connectivity of T4L α-helices (21 restraints for 7 helices). Finally this study introduced a practical criterion for identifying the “native-like” model out of the thousands generated by Rosetta. These models have simultaneously low Rosetta energy and restraint violation scores.