Multistart simulated annealing refinement of the crystal structure of the 70S ribosome

Abstract

A macromolecular X-ray crystal structure is usually represented as a single static model with a single set of temperature factors representing a simple approximation of motion and disorder of the structure. Multiconformer representations of small proteins have been shown to better describe anisotropic motion and disorder and improve the quality of their electron density maps. Here, we apply multistart simulated annealing crystallographic refinement to a 70S ribosome-RF1 translation termination complex that was recently solved at 3.2 Å resolution. The analysis improves the interpretability of the electron density map of this 2.5-MDa ribonucleoprotein complex and provides insights into its structural dynamics. We also used multistart refinement and conventional Fourier difference maps to address a recent study in which cross-crystal averaging between two crystal forms of the 70S ribosome was used to evaluate reported differences between two ribosome crystal structures solved at 2.8 and 3.7 Å resolution. Our analysis suggests that results obtained from cross-crystal averaging are inherently biased toward the higher-resolution dataset.

Crystallographic techniques and methods of structure determination have progressed rapidly in recent years, accompanied by an explosion in the amount of structural information about biological molecules and macromolecular assemblies. X-ray crystal structures have traditionally been represented by a single set of atomic coordinates accompanied by a set of temperature factors (B factors) to characterize atomic thermal motion. At most experimentally attainable resolutions, a spherical approximation for thermal motion is used, accounting only for the isotropic component of atomic displacement. It has been shown that the use of a static model with isotropic thermal parameters underestimates the total disorder and conformational variability of macromolecules (1–5). The limitation of the single-model approach is exemplified by several cases where high-resolution data are available. Nearly half the amino acid side chains and even some backbone atoms were found to have alternative conformations in the 0.54 Å structure of crambin and in the 1.0 Å structure of calmodulin (1, 6). To account for anharmonic contributions and conformational variability in X-ray structures, the use of an ensemble of models, as has long been standard practice in structure determination by NMR, has been suggested as a replacement for the traditional single-model approach (7). Multiconformer refinement has been implemented using both real- and reciprocal-space refinement approaches (2, 3, 8–13) and was demonstrated to yield improved agreement with diffraction data in most cases. In the case of molecular replacement phasing, electron density maps are often biased toward the search model; when a structure ensemble is used, however, model bias can be reduced significantly (14).

In X-ray structures of the ribosome, the largest asymmetric structures so far solved, modeling at medium resolution has often been compromised by the quality of the electron density map, especially for protein side chains. Furthermore, the use of molecular replacement phasing leads to inheritance of model error in subsequent ribosome structures. One of the most popular ways of reducing model bias is calculation of simulated-annealing composite omit maps (15, 16). Application of this method to the ribosome, however, is computationally impractical due to the large number of atoms and structure factors. As an alternative, local omit maps can be used (17); however, they, and composite omit maps, tend to be noisier than those calculated from the complete model, such as maximum-likelihood-weighted maps (18–20), and contain model bias if the resolution of the data and the accuracy of the starting model do not allow aggressive simulated annealing atomic refinement (17). Electron density maps calculated from an ensemble of structures contain phase information from multiple structures. Due to averaging of the errors present in individual models, these maps can be superior not only to single-model density maps (14, 21) but also to omit maps, where phase information for an omitted region is obtained from the rest of a single atomic model and experimental structure factor amplitudes.

In this work, we apply a multistart simulated annealing (MSSA) approach in crystallographic refinement of the 70S ribosome. It involves obtaining a set of individual, independently refined structures rather than simultaneous refinement of an ensemble, which would require a high data-to-parameter ratio and thus is not applicable at moderate resolution. We find that variability in the atomic coordinates of the multiconformer model of the RF1 termination complex (22) correlates with isotropic temperature factors and with anisotropic temperature factors derived from TLS refinement of this same complex.

We also use MSSA refinement and conventional Fourier difference maps to address a recent study (23) in which cross-crystal-form averaging was used in an attempt to evaluate structural differences between two ribosomal complexes solved in two different crystal forms at 3.7 Å (24) and 2.8 Å resolution (25). That study concluded that several features that were found in different conformations in the two structures or were absent from the 2.8 Å structure were the same as the 2.8 Å structure in both cases. Our findings suggest that cross-crystal averaging, as applied, leads to bias toward the stronger 2.8 Å dataset, yielding electron density maps that reflect the features of the higher-resolution structure. We conclude that multistart simulated annealing refinement is a useful method for model building, dynamics analysis and validation of large macromolecular structures.

Results

Dynamics of a 70S Ribosome Termination Complex.

We refined the structure of an RF1 ribosomal termination complex at 3.2 Å (22) using different simulated-annealing protocols (see Methods), providing several independently refined structures with similar R_free values that were used to create an ensemble of structure models. To eliminate potential B factor-driven bias during refinement, all atomic temperature factors were reset to a single (global) value before refinement. Calculated atomic positional covariance tensors and root-mean-square differences (RMSD) between equivalent atomic positions in the resulting structures (Fig. 1 A, D, and G) were found to correlate with the translation-libration-screw (TLS)-derived anisotropic (Fig. 1 B and E and Figs. S1–S4) and isotropic (Fig. 1H) temperature factors of the refined model (22). This correlation improved as more structures were used for comparison; 4–5 independently refined structures appear to be sufficient for the analysis of dynamics (Fig. 1I). Although RMSDs were reported to correlate with area exposed to solvent in some previous studies (9), there is no such correlation for the ribosomal termination complex (compare Fig. 1 A and D with C and F). The correlation coefficient between atomic RMSDs for four 23S rRNA structures and atomic solvent exposure of 23S rRNA is 0.07, whereas that between the RMSDs and B factors is 0.7 (Fig. 1I). The possibility that the mobility of atoms during refinement is dominated simply by atomic packing can therefore be excluded in this case.

Fig. 1.

Correlation between the structural dynamics of the ribosome inferred from multistart simulated annealing (MSSA) refinements and from atomic B factors for (A–C) the 50S subunit, (D–F) the 30S subunit and (G and H) RF1 from the RF1–70S termination complex. (A, D, and G) Structures are colored according to RMSD values (pseudo anisotropic B factors) from the refined MSSA model. (B, E, and H) Structures are colored according to anisotropic (B and E) or isotropic (H) atomic B factors from the TLS-refined model. (C and F) Structures are colored according to solvent accessibility in the ligand-bound 70S termination complex. In all panels, parameters are ramped from blue (lowest values) to red (highest values) using bins containing equal number of atoms. (I) Correlation of isotropic atomic B factors with RMS differences between the structures of a MSSA model for the RF1-bound 70S termination complex (black squares) and 23S rRNA (blue triangles) as a function of the number of independently refined structures used for the multiconformer model.

TLS refinement is intrinsically biased toward the choice of TLS groups, whereas MSSA refinement is devoid of this kind of bias. Multistart simulated annealing refinement, therefore, can be used to define TLS groups less subjectively. One way to approach this would include interpretation of multiconformer coordinates as pseudoanisotropic B factors followed by conversion to and optimization of TLS groups and TLS tensors. Our preliminary optimization of TLS tensors (Figs. S1b, S2b, S3b, and S4b), calculated from atomic positional tensors (pseudo anisotropic B factors) of the multiconformer ensemble (Fig. 1A and Figs. S1a, S2a, S3a, and S4a) yields results that are generally similar to those obtained from a straightforward TLS refinement (Fig. 1B and Figs. S1c, S2c, S3c and S4c), if the same subjective choice of TLS groups is used. For the 70S ribosome, the all-atom correlation between the atomic positional tensors converted into anisotropic B factors using 11 TLS groups (pseudoTLS representation; see Methods) and the anisotropic B factors derived from TLS refinement using the same groups is 0.56 (0.74 for 23S ribosomal RNA, the largest constituent of the ribosome); importantly, its positive value reflects similar directionality of anisotropic B factor tensors between those resulting from MSSA refinement and those obtained from TLS refinement. At the periphery of the ribosome, however, the TLS B-factor ellipsoids derived from the MSSA model appear to be more anisotropic than those derived from the TLS-refined structure (Figs. S1–S4). This may reflect the limitations of TLS refinement using the current conservative choice of TLS groups; an automated approach to the choice of TLS groups based on multiconformer anisotropy would likely yield a more accurate description of the anisotropic modes for the dynamics of a macromolecule.

Analysis of RMSDs between the structures comprising our multiconformer model suggests that conformational motion or disorder tends to be higher around the periphery of the ribosomal subunits (Fig. 1), consistent with the ability of the ribosome to undergo spontaneous intersubunit rotational movement (26). Interestingly, RMSDs between separately refined structures of the 70S ribosome-RF1 complex also correlate with anisotropic temperature factors derived from a lower-resolution structure of a 70S ribosome-tRNA complex by TLS refinement and obtained in a different crystal form (27), suggesting that the large-scale dynamic properties of the ribosome are similar for these complexes. Two of the most highly mobile elements of the ribosome revealed by MSSA refinement are the L1 (Fig. 1 and Fig. S2) and L11 stalks (Fig. 1 and Fig. S3), in agreement with conclusions based on the results of numerous other approaches (27–37). In some cases, multiconformer analysis may provide information about molecular motions that is not evident from TLS-derived B factor values alone. For example, helix 38 (h38) of 23S rRNA, which forms an intersubunit bridge with protein S13, and helix 84 (h84), which, bound to protein L5, also forms an intersubunit bridge with S13, and the L11 stalk, are similarly flexible according to TLS analysis (Fig. 1B), whereas consideration of the multiconformer model suggests that the mobilities of h38 and h84 are much lower than that of the L11 stalk (Fig. 1A).

The most conformationally variable parts of the small subunit are at the beak of the head and the spur, in agreement with results from other approaches (Fig. 1 D and E and Fig. S1). Anisotropic dynamics of the head obtained by multistart refinement (Fig. S4 a and b) and TLS refinement (Fig. S4c) are consistent with structural rearrangements in the 30S subunit implicated in translocation, where the head appears to rotate along the direction of translocating tRNAs (33, 38). The beak of the head and spur appear to be similarly flexible according to TLS analysis (Fig. 1E), whereas multiconformer analysis suggests that the spur is much more dynamic than the beak.

Interestingly, the conformational variability of domain 1 of release factor RF1 appears to be correlated with that of its ribosomal contact sites in the beak of the small subunit and in the L11 stalk of the large subunit (Fig. 1 G and H). In contrast, domain 2, which bears the codon-recognition elements and domain 3, which contains the GGQ motif involved in hydrolysis of the peptidyl-tRNA ester linkage, are the least mobile regions of RF1 in its ribosome-bound state, suggesting that low conformational noise may be critical to their roles in termination of protein synthesis.

As an alternative to the proposal to deposite multiple structures for a given X-ray dataset in the Protein Data Bank (7), which can be impractical for very large structures such as the ribosome, we suggest deposition of a structure corresponding to the average of the ensemble, along with pseudoanisotropic B factors, in addition to a conventionally refined structure obtained without the use of multistart refinement. In this case, the latter structure would be available for detailed structure analysis, whereas the averaged multistart refined structure would be useful for evaluation of anisotropic dynamics and disorder.

Improvement of Electron Density Maps.

Multistart simulated annealing refinement and other multiconformer approaches have been shown to be able to improve the quality of electron density maps of small proteins (3, 8, 14, 39). Here, we tested the impact of the multiconformer approach on the quality of the moderate 3.2 Å resolution 70S ribosome maps, where the number of atoms is two orders of magnitude greater than in previous multiconformer studies (3, 8, 14, 39). Because the overall diffraction intensity (I/σI) of a ribosome crystal is significantly lower than that observed for low-molecular-weight macromolecules, and the property that ribosomal RNA is the main contributor to X-ray scattering from ribosome crystals, the electron density for ribosomal proteins and factors bound to the ribosome is generally more difficult to interpret than that seen for small, isolated proteins, even at similar resolutions. In the case of the ribosome, the global indicators of the fit of a single model to the data, R and R_free, are relatively insensitive to protein side-chain conformation variability, or even errors in register, making model-building a more challenging task.

We undertook two approaches to assess the benefits of multistart simulated annealing refinement. In the first case, an ensemble of structures for the 70S ribosome-RF1 termination complex was obtained from multiple, independent simulated-annealing refinements at 3.2 Å resolution, starting from a single refined model and creating different refinement trajectories by the use of different starting temperatures. We then examined the effect of MSSA refinement on the quality of the electron density map for domain 1 of RF1, a region of the map where interpretation was particularly difficult (22). The map calculated from the ensemble showed a dramatic improvement over that calculated from the single model, allowing unambiguous interpretation of the density for several previously unobserved side chains (Fig. 2 A–C). In a second test, the 3 Å structure of the RF2 termination complex was used (Fig. 2 D–F) (40). We asked whether a simulated ensemble of 70S ribosome-RF2 models containing a GGQ loop that was intentionally built incorrectly is able to reconstruct the correct electron density for the loop. The structures were obtained by simulated-annealing refinements of a model in which an error in sequence register was introduced at the GGQ motif, which was also manually distorted to create an incorrect fold (Fig. 2E). The RMSD of the resulting refined structures from the correct solution was at least 4 Å (all-atom; 3 Å for C-α atoms) for residues 240–265, comprising the GGQ motif of RF2. A σ_A-weighted 2F_obs − F_calc electron density map, calculated at 3 Å resolution from this set of structures (Fig. 2 E and F), was at least as interpretable as a simulated-annealing omit map for this region (Fig. 2D) and better than maps calculated using any individual model of the ensemble; the multiconformer-averaged electron density corresponding to the polypeptide backbone was slightly more continuous, and the helical fold of the GGQ region was more readily interpretable than in the omit map, yet none of the models of the ensemble provides a generally better fit to the densities than any of the others. This example demonstrates that in a real-life situation, when the quality of an omit map is insufficient to unambiguously fit part of a structure, building an approximate model, followed by multistart simulated annealing refinement can provide a clearer map, allowing for more accurate modeling. On the basis of these tests, we conclude that multiconformer-averaged maps for proteins in the context of the ribosome are less biased and can help substantially in model building, even where the starting model includes errors in fold or in sequence register.

Fig. 2.

Stereo images showing the effect of multistart simulated annealing refinement on weak regions of electron density maps. (A–C) Improvement of the electron density map for residues 23–28 of domain I of RF1. (A) Conventional σ_A-weighted 2F_obs − F_calc map. (B and C) MSSA σ_A-weighted 2F_obs − F_calc electron density map, showing all of the models (B) and only the final model (C). (D–F) Reconstructing the electron density for the GGQ loop of domain III of RF2. An error in sequence register was intentionally introduced into the starting model in the region of the GGQ motif, which was also manually distorted to create an incorrect fold. (D) Simulated-annealing 2F_obs − F_calc omit map. (E and F) MSSA σ_A-weighted 2F_obs − F_calc electron density map, showing the distorted starting (E) and the final refined (F) models. The maps are contoured at 1.2σ.

Evaluating Differences Between 2.8 Å and 3.7 Å Structures of 70S Ribosome Complexes.

In the study in ref. 23, cross-crystal averaging was used to evaluate differences between the structures of two different ribosome complexes containing P and E site tRNAs at 3.7 Å resolution (24) and A, P and E site tRNAs in the presence of paromomycin at 2.8 Å resolution (25). Although the detailed structures of the ribosomes and tRNAs were generally very similar in the two complexes, a number of detailed differences were observed, including the conformations of certain features of the peptidyl-transferase center (PTC), the L1 stalk, the sarcin-ricin loop (helix 95) of 23S rRNA, the A site finger (helix 38 of 23S rRNA) and the occupancies of certain ribosomal proteins. In this section, we reexamine these issues, using the multiconformer refinement approach and conventional crystallographic methods. Cross-crystal form averaging, or multicrystal averaging, is a method that was developed for improving electron density for a macromolecular structure crystallized in two or more different crystal forms (41). Cross-crystal averaging involves local alignment of electron density maps of equivalent regions in different crystal forms, followed by density averaging. The method has been particularly successful in cases when only very weak initial experimental phasing is available (42).

In the study in ref. 23, cross-crystal averaging was applied to a 2.8 Å resolution structure from a complex crystallized in the orthorhombic space group P2₁2₁2₁ (25) and a 3.7 Å resolution structure from a tetragonal I422 crystal form (24). Although most of the structural details reported for 23S rRNA were found to be essentially identical for the two published structures, a number of differences, including a shift in the position of the noncanonical A2450-C2063 base pair and the adjacent A2451 in the PTC of the 3.7 Å structure, were observed (24). When phases modified by cross-crystal averaging were used to calculate an electron density map from the 3.7 Å diffraction data, the map in the peptidyl transferase center was found to be identical to that calculated from the 2.8 Å data (23). Further differences were found in the L1 stalk and protein L36, where electron density for helix 78 (residues 2130–2160) of 23S rRNA and L36 was present in the 3.7 Å map (24) but absent in the 2.8 Å map due to low occupancy and/or disorder (25). In the map calculated using phases from cross-crystal averaging and structure factor amplitudes from the 3.7 Å data, electron density was reported to be “of poor quality” or “missing completely” for h78 and L36, respectively (23). By using different starting models and arriving at similar final averaged electron density maps, the authors concluded that the cross-crystal averaging approach is unbiased toward the starting model. However, the much stronger 2.8 Å data (I/σI is estimated to be >14 when truncated at 3.7 Å, as described in Methods, compared with I/σI = 3.0 for the 3.7 Å dataset) and the presence of noncrystallographic symmetry in the orthorhombic crystal form are potential sources of bias toward the 2.8 Å structure. The authors (23) attempted to rule out possible bias toward the 2.8 Å experimental data by calculating cross-crystal-form averaged maps around the ribosomal A site. As in the published structures, the anticodon stem-loop of A site tRNA could be visualized in the map calculated using the 2.8 Å amplitudes but not in the map using the 3.7 Å data. The absence of A site anticodon density in the 3.7 Å map was taken as evidence for the absence of bias in cross-crystal averaged maps. However, this conclusion is called into question by our observation that calculation of such a map using the 3.7 Å data and an artificially biased refined model where the anticodon stem-loop is present, also shows no density in this region (Fig. S5).

In a further effort to resolve the discrepancies between the 2.8 Å and 3.7 Å structures, we used two alternative approaches to evaluate the occupancies and/or order of protein L36 and helix 78 of 23S rRNA and the conformational details of the peptidyl-transferase center. In agreement with our original conclusions, and contrary to the results of cross-crystal averaging, electron density for L36 and h78 is clearly visible in σ_A-weighted F_obs − F_calc Fourier difference maps (Fig. 3A and B) calculated using the 3.7 Å data and 2.8 Å model, where these features are absent, therefore excluding model bias. L36 is also present in other structures of the Thermus thermophilus 70S ribosome determined from the I422 crystal form (43–46), suggesting that the presence of L36 may be related to the details of ribosome preparation leading to formation of the I422 crystals.

Fig. 3.

Conventional Fourier difference maps and MSSA maps support the original conclusions for the 3.7 Å structure (24). (A and B) σ_A-weighted F_obs − F_calc difference maps calculated using phases from the 2.8 Å model (26) and structure factors from the 3.7 Å data (24) clearly show the presence of helix 78 of 23S rRNA (A) and protein L36 (B), which were reported as disordered or absent in the 2.8 Å structure and in the cross-crystal averaged maps (23). The absence of any potential bias from the 3.7 Å model suggests that the cross-crystal averaged maps were biased toward the stronger 2.8 Å dataset. The maps were contoured at 1.1σ (A) and 1.4σ (B). (C–F) Twelve-model multiconformer averaged 2F_obs − F_calc density map in the peptidyl transferase center for nucleotides A2450 and C2063 (C and D) and A2451 (E and F). The map, calculated using phases from the 2.8 Å model refined against the 3.7 Å diffraction data (as described in Results), shows that electron density corresponding to nucleotides A2450, A2451 and C2063 is closer to the ribose of A76 of the P site tRNA in the 3.7 Å structure (24) than in the 2.8 Å structure (25). (C and E) The model refined against the 3.7 Å data are in blue (23S rRNA) and orange (P-tRNA). (D and F) The 2.8 Å model is shown in gray (23S rRNA) and orange (P-tRNA). The map is contoured at 1.4 σ.

To evaluate the conformational differences reported for certain nucleotides in the peptidyl transferase center between the 2.8 Å and 3.7 Å structures, two types of maps were calculated. To exclude any possible bias toward the published 3.7 Å model, all maps were calculated using phases from the 70S ribosome and P and E site tRNA models from the 2.8 Å structure and measured structure factor amplitudes from the 3.7 Å dataset. First, a σ_A-weighted 3F_obs − 2F_calc map was calculated using phases from the 2.8 Å model after subjecting it to one round of rigid-body refinement followed by a conservative simulated-annealing torsion angle refinement employing protein and RNA secondary structure restraints against the 3.7 Å data. Second, a multiconformer σ_A-weighted 2F_obs − F_calc map was calculated averaging F_calc from 12 multistart structures refined against the 3.7 Å data using starting temperatures ranging from 500 K to 2,000 K. Six of the structures were refined with the conformation of the PTC (i.e., all nucleotides within a 10 Å radius of A2450) fixed and the other six were refined conventionally; in all of the structures obtained by the latter procedure, nucleotides A2450, A2451 and C2063 were shifted toward the ribose of A76 of the P site tRNA. We then averaged the 12 structures to calculate an electron density map from the 3.7 Å data and obtained real-space R factors and correlation coefficients for each of these structures with respect to this averaged map. The real-space statistics for the 6 structures where the three nucleotides are shifted (CC = 0.679 ± 0.003; R = 0.249 ± 0.001), compared with those where the PTC was preserved as in the 2.8 Å structure (CC = 0.659 ± 0.002, R = 0.2567 ± 0.0007), show that the fit of the shifted nucleotides to the averaged map is significantly better than that of the 2.8 Å structure.

Both the conventional 3F_obs − 2F_calc (Figs. S6 a and b and S7 a and b) and multiconformer maps (Fig. 3 C–F and Figs. S6 c and d and S7 c and d) support our original findings that nucleotides A2450, A2451 and C2063 are closer to the ribose of A76 of the P site tRNA in the 3.7 Å structure (24) than in the 2.8 Å structure (25) (Fig. 3 C–F). The B factors of the starting models used in the multiconformer procedure were reset to a single global value to avoid possible bias toward any particular model; yet, the features of the multiconformer map in the PTC support the displaced positions of A2450, A2451, and C2063.

Together, these results support our original findings (24) and strongly suggest that the conclusions derived from the cross-crystal averaging study were biased toward the stronger 2.8 Å experimental data. Our control experiments, in the form of straightforward conventional difference map calculations for h78 and L36, clearly show that the electron density map derived from multicrystal averaging calculations in this case (23) does not correctly represent the electron density for these regions and further point to the existence of bias toward the stronger dataset. In the absence of a detailed description of the protocols applied by Simonovic et al. (23), it is not possible to draw strong conclusions concerning the precise source of this bias. For example, it is not clear whether Simonovic et al. (23) modified the 2.8 Å dataset (e.g., by introducing random error into F_obs) to scale down its signal/noise to values compatible with those of the weaker 3.7 Å data. In addition, the details of calculation of figure of merit (FOM), that are used for weighting in multicrystal averaging, are not clear from the description of their work. FOMs calculated from refined models are likely to be systematically overestimated (47) and thus can adversely affect multicrystal averaging, which is most effective when weak experimental phases rather than model phases are used. The cross-crystal averaged maps appeared most “improved” (23) when 3-fold averaging was used, i.e., when modified phases were obtained from averaging two noncrystallographic-symmetry related maps calculated from the 2.8 Å dataset and one map calculated from the weaker 3.7 Å dataset. We suggest that the apparent improvement in the interpretability of the maps in this case was the result of phase bias toward the 2.8 Å dataset.

We conclude that the reported differences regarding the occupancies of L36, h78, and the conformation of the PTC (24) are consistent with the 3.7 Å dataset, and may be the result of different ionic conditions, crystal packing, crystallization conditions and/or ribosome isolation and purification approaches. The resolution of the data, however, is not sufficient to draw strong conclusions concerning details of the mechanism of peptide bond formation. The structure of the 70S ribosome-tRNA complex crystallized in a form similar to that reported for the 3.7 Å structure and solved at a higher resolution will be required to further address any potential functional implications of the observed differences.

Methods

Multiconformer structures for the 70S-RF1 and 70S-RF2 complexes were obtained using two approaches. Atomic temperature factors of the starting models (PDB ID codes 3D5A–3D5D and 3F1E–3F1H for RF1 and RF2 complexes, respectively) were set to 60.00. The structure was then subjected to different simulated-annealing torsion-angle refinement protocols in CNS (16) in parallel; in one, multitemperature, approach, starting temperatures ranging from 1,000 to 10,000 K were used, whereas in a second, multiseed, approach, the structures were refined all starting at 5,000 K, but using different random seeds leading to different initial velocities (multistart). The resulting structures were characterized by similar R/R_free values (R ranged from 0.3235 to 0.3248; R_free ranged from 0.3685 to 0.3715; these values are higher than the published ones because atomic B factors were set to a global constant value). A multiconformer model consisting of the multiseed-refined (MSSA) set of structures was used for analysis of overall structural variability and for calculation of electron density maps. The results obtained from the analyses of the multitemperature-refined set of structures were similar to those of the multiseed set.

Detailed information concerning calculation of MSSA maps, calculation of pseudoanisotropic B factors from a set of MSSA-refined structures, calculation of pseudoTLS tensors and correlation coefficients, and re-refinement of the 3.7 Å structure can be found in SI Methods and SI Appendix.