De Novo High-Resolution Protein Structure Determination from Sparse Spin labeling EPR Data

Summary

As many key proteins evade crystallization and remain too large for nuclear magnetic resonance spectroscopy, electron paramagnetic resonance (EPR) spectroscopy combined with site-directed spin labeling offers an alternative approach for obtaining structural information. Such information must be translated into geometric restraints to be used in computer simulations. Here, distances between spin labels are converted into distance ranges between β-carbons using a “motion-on-a-cone” model, and a linear-correlation model links spin label accessibility to the number of neighboring residues. This approach was tested on T4-lysozyme and αA-crystallin with the de novo structure prediction algorithm Rosetta. The results demonstrate the feasibility of obtaining highly-accurate, atomic-detail models from EPR data by yielding 1.0Å and 2.6Å full-atom models, respectively. Distance restraints between amino acids far apart in sequence but close in space are most valuable for structure determination. The approach can be extended to other experimental techniques such as fluorescence spectroscopy, substituted cysteine accessibility method, or mutational studies.

Introduction

The accelerated pace of genome sequencing has sparked the development of rapid structure determination methods and ambitious proposals for genome-scale structure determination utilizing primarily X-ray crystallography and nuclear magnetic resonance (NMR) spectroscopy (Stevens, Yokoyama et al. 2001; Berman, Battistuz et al. 2002; Lesley, Kuhn et al. 2002; Westbrook, Feng et al. 2003). However, it has become clear that static and dynamic structural information for a significant subspace of the protein universe continues to evade these tools. Important examples include the static structure of membrane proteins (Tusnady, Dosztanyi et al. 2004), conformationally heterogeneous water-soluble proteins (Haley, Bova et al. 2000), and large protein complexes involved in major cellular processes (Harrison 2004). Insight into conformational motions that mediate function is restricted to proteins amenable to NMR spectroscopy or to crystallization in multiple intermediate states. Furthermore, the absence of representative structures of entire protein families, whose members often share difficulties in structure determination, reduces the efficiency and accuracy of comparative modeling (Sali 1998).

A complement of methods with intrinsically lower resolution can provide insight into these problems. Among them are probe-based approaches such as electron paramagnetic resonance (EPR) spectroscopy in combination with site-directed spin labeling (SDSL) (Hubbell, Mchaourab et al. 1996; Mchaourab, Berengian et al. 1997; Koteiche, Berengian et al. 1998; Perozo, Cortes et al. 1999; Liu, Sompornpisut et al. 2001; Brown, Sale et al. 2002; Dong, Yang et al. 2005). EPR analysis of spin labeled proteins results in a set of structural restraints that describes, in a native-like setting, local environments as well as aspects of the global fold of the protein. Spin label accessibility and mobility can be used to determine secondary structure location and topology (Farahbakhsh, Altenbach et al. 1992; Altenbach, Froncisz et al. 2005). Distance measurements between pairs of spin labels in the range from 5−60Å (Rabenstein and Shin 1995; Borbat, Mchaourab et al. 2002) reflect the relative packing of domains and secondary structures. In cases where these parameters are obtained in various conformational intermediates of the protein, they allow for a detailed mapping of structural changes involved in function (Dong, Yang et al. 2005). There are relatively few limits on the size and environment of the protein, particularly when compared to X-ray crystallography and NMR spectroscopy.

Despite the widespread application of SDSL (Fanucci and Cafiso 2006), the use of EPR restraints for structure determination has not been systematically explored. A central question is the number and nature of EPR restraints necessary to obtain a structural model at a biologically relevant resolution. The most extensive use of spectroscopic data along with computational methods for structure determination is in NMR spectroscopy (Wüthrich 1986). Typically consisting of distances not greater than 5−6Å with upper and lower bounds, the geometric information is derived from NOE-based experiments. The number of such restraints required for the determination of a structure depends on the range and quality of such restraints, but is generally assumed to be above 15 restraints per residue (Nederveen, Doreleijers et al. 2005).

Although EPR distance restraints have a longer range than their NMR counterparts, they are fundamentally less accurate since they report distances between probes introduced into the protein sequence. The significant length of the spin label linking arm implies that the EPR distances will have a rather large uncertainty when translated into distances between α- or β-carbons unless the conformation of the spin label is known at every site. Therefore, previous efforts have focused on either using molecular dynamics simulations to define their trajectories (Sale, Song et al. 2005) or on determining a library of rotamers from crystal structures of spin labeled T4-lysozyme (Langen, Oh et al. 2000). These studies are critical since spin label conformations are likely stabilized by weak specific interactions with neighboring amino acid side chain or backbone atoms. However, such calculations are time and resource intensive and not practical without a high-resolution structural model of the protein.

Sparse experimental data, such as EPR restraints, aid computational protein structure prediction algorithms by restricting the conformational space that must be considered in order to obtain the correct structure. For instance, the Rosetta de novo protein structure prediction algorithm (Simons, Kooperberg et al. 1997; Bonneau, Strauss et al. 2001; Bonneau, Tsai et al. 2001; Bradley, Chivian et al. 2003; Rohl, Strauss et al. 2004; Bradley, Malmstrom et al. 2005) predicts high-resolution (better than 1.5Å) structures of proteins with less than 80 amino acids in the absence of experimental restraints (Bradley, Misura et al. 2005). In combination with sparse (less than one restraint per amino acid) NMR NOE distances and/or residual dipolar couplings (Bowers, Strauss et al. 2000; Rohl and Baker 2002), the structure of proteins with up to 200 amino acids can be determined to medium-high-resolution (1.5−3.0Å).

In the present work, Rosetta is combined with sparse EPR distance restraints and solvent accessibility measures for high-resolution structure determination of the mostly helical T4-lysozyme (Weaver and Matthews 1987) and the all β-sheet protein αA-crystallin (Horwitz 1992; Horwitz 1993). We address the question of whether the EPR restraints can restrict the protein conformational space without assuming an atomic-detail model for the spin label's dynamics and without accounting for its context-dependent specific interactions. Also addressed are the questions of how many restraints are needed to obtain a high-resolution structure and what type of EPR restraint is most efficient.

The results demonstrate that sparse EPR restraints derived from a non-atomic model of the spin label lead Rosetta to high-resolution structures for both proteins. Also, distance restraints are more efficient in restricting conformational space than spin label accessibilities. Further analysis reveals that those between two amino acids far apart in sequence but close in Euclidian space are the most valuable.

Results

EPR distance and accessibility data were transformed into structural restraints as described in Experimental Procedures. Briefly, distances between spin labels were translated into distances between β-carbons using a motion-on-a-cone model of the spin label location relative to the a-carbon. The accessibilities of spin labels were computationally interpreted in terms of the exposed surface area. The effectiveness of the restraints to aid Rosetta in the folding process was then evaluated. Because distance restraints proved vastly more efficient in preliminary experiments, accessibility data was not used during modeling. De novo models were compared to the crystal structure of T4-lysozyme and a comparative model of aA-crystallin.

Evaluation of the “Motion-on-a-cone” model for interpretation of distance restraints

The ‘‘motion-on-a-cone’’ model (see Figure 1) yields a predicted distribution for the difference between the distance separating the spin labels (d_SL) and that separating the two corresponding Cβs carbons (Cβs; d_Cβ). Comparison of the predicted d_SL - d_Cβ distribution with the d_SL - d_Cβ obtained from the T4-lysozyme and αA-crystallin structures (Figure 1D) demonstrate that they essentially encompass the same range of d_SL - d_Cβ and reveals a common bias in experiment and model for d_SL > d_Cβ. The comparison also reveals that the model over-predicts the frequency with which large (> 4Å) d_SL - d_Cβ values occur and underestimates the frequency with which low (< 4Å) d_SL - d_Cβ values occur. However, the present application depends only on the ability of the model to predict the appropriate range of values for d_SL - d_Cβ; the frequency with which these values occur is not a part of the utility of this simple model.

Figure 1.

Rational for translating d_SL into d_Cβ for use as a restraint. A) Chemical structure of a nitroxide spin label side chain with the distance from the Cβ atom to the spin label atom indicated (Borbat, Mchaourab et al. 2002). B) Illustration of how the maximum distance from Cβ to spin label, SL, is reduced to an effective distance, SL^eff (depicted by a double line). C) d_SL is a starting point for the upper estimate of d_Cβ, and subtracting the effective distance of 6Å twice from d_SL gives a starting point for the lower estimate of d_Cβ. D) A histogram compares T4-lysozyme crystal structure (black bars, left y-axis, bottom x-axis) and αA-crystallin comparative model (white bars, left y-axis, bottom x-axis) d_SL–d_Cβ values with those obtained from the simple cone model (circles and line, right y-axis, top x-axis).

Given a d_SL, the “motion-on-a-cone” model provides a restraint in the form of a predicted range for d_Cβ. The accuracy of the range can be evaluated by comparing it to the d_Cβ calculated from the T4-lysozyme structure and the αA-crystallin comparative model. Practically all calculated d_Cβ lie within the range predicted by the model (Figure 2C, G).

Figure 2.

Map of the EPR restraints on the T4-lysozyme crystal structure (A-D) and on the αA-crystallin comparative model (E-H). A and E) Red dotted lines show d_Cβ distances, which are restrained by respective d_SL. B and F) Residues for which accessibilities e_SL were measured are depicted as space-filling models. C and G) Diagram shows d_SL (blue circle), the range of the derived distance restraints (blue), and the corresponding crystal/comparative model d_Cβ (red bar). D and H) Diagram illustrating the correlation of e_SL with e_Cβ. The lines indicate the consensus model fit ±3•σ_Cβ, where σ_Cβ, was recalculated based on the consensus fit to be 1.70Å. In B, F, D, and H the residues are color-coded with decreasing e_SL from blue – cyan – yellow – orange – red; black indicates amino acids in αA-crystallin that show reduced experimental accessibility due to intermolecular contacts with other αA-crystallin units in the oligomeric protein.

A Linear Regression Relation for Interpretation of Accessibility Restraints

Analogous to an experimental distance measurement, the experimental accessibility of a spin label (e_SL) needs to be translated into accessible surface areas of the protein structure for use as a restraint. For this purpose, a consensus linear regression relation between e_SL and the number of Cβ atoms within 8Å of the Cβ of the corresponding amino acid (e_Cβ) was determined from T4-lysozyme and αA-crystallin structures. The linear relation is given by e_Cβ = (0.76–e_SL)•20.87 and has a correlation coefficient of −0.83 to experimental T4-lysozyme and αA-crystallin data (Figure 2D, H). The strong correlation suggests the simple method of linearly relating e_SL to e_Cβ is a sufficient means for obtaining a structural restraint from EPR accessibility data.

Influence of EPR data on de novo fold determination

To avoid the introduction of noise through unconstrained regions and focus on evaluating the contribution of EPR restraints in structure prediction, regions in both proteins that were not probed with spin labels were excluded from the calculations. For T4-lysozyme, the C-terminal 107 residue helical domain (amino acids 58−164) was modeled (Figure 2A). For αA-crystallin, the C-terminal 88-residue β-sandwich domain (amino acids 60−147) was modeled (Figure 2E).

The influence of the experimental EPR restraints on de novo protein folding with Rosetta was evaluated by building 10,000 models for each protein (a) without the use of experimental data, (b) with only the use of distance restraints, (c) with only the use of solvent accessibility restraints, and (d) using both sets of restraints. The average model quality was monitored by the root-mean-square-deviation (RMSD). The results for both T4-lysozyme and αA-crystallin follow the same trends: there is an improvement in the quality of models created using distance restraints compared to models created without distance restraints (Figure 3A and B, T4-lysozyme; Figure 3I and J, αA-crystallin); there is very little to no improvement when accessibility restraints are included (Figure 3A and C, T4-lysozyme; Figure 3I and K, αA-crystallin); using accessibility restraints in conjunction with distance restraints provides little to no improvement over using distance restraints alone (Figure 3B and D, T4-lysozyme; Figure 3J and L, αA-crystallin). It is clear from these analyses that the distance restraints are critical for improving the RMSD distribution of models, while solvent accessibility data only marginally improve the quality of water soluble protein models.

Figure 3.

Illustration of the value of the experimental restraints in de novo protein folding for T4-lysozyme (A - H) and αA-crystallin (I - P). The backbone RMSD distribution of 10,000 T4-lysozyme de novo models created A) without the use of EPR restraints, B) with only the use of EPR distance restraints, C) with only the use of EPR accessibility restraints, D) with the use of EPR distance and accessibility restraints. E) The backbone RMSD distribution of 10,000 T4-lysozyme de novo models created with the use of 1/3 of the EPR distance restraints: top bar) those with the largest information content; second bar) those between amino acids furthest apart in sequence. The third and fourth black and white bars denote the sum percent of information content of the restraints used for the top and second bars, respectively. The width of the blocks comprising the black and white bars denotes the information content of individual restraints. F) Same as for E) but using the distance restraints with the lowest information content (top bar) and nearest in sequence (second bar). G) Same as for E) but using 2/3 of the total distance restraints. H) Same as for F) but using 2/3 of the total distance restraints. I – P) Same as A - H but for αA-crystallin.

Influence of spin label placement on de novo fold determination

Spatial contacts of amino acids that are distant in sequence define the protein fold best (Baker 2000; Bonneau, Ruczinski et al. 2002). However, whereas the sequence separation s_Cβ can be chosen when designing an EPR experiment, the Euclidean distance d_Cβ is generally unknown. EPR experiments do not provide contact data, but, instead, distances of up to 50Ǻ. Thus, the information content (I_Cβ) of an EPR distance restraint can be defined as directly proportional to (s_Cβ) but inversely proportional to I_Cβ ∼ s_Cβ/d_Cβ.

To investigate the influence of spin label location and resulting I_Cβ on de novo structure determination, two experiments were designed with subgroups of all available restraints. First, all restraints are ranked by I_Cβ to assess their power in an idealized experiment. Second, all restraints are ranked by s_Cβ in order to simulate the choices the experimentalist can make when selecting sites for labeling. 10,000 models for T4-lysozyme and αA-crystallin were built which used a) the one-third restraints with highest I_Cβ, b) the one-third restraints with lowest I_Cβ, c) the two-third restraints with highest I_Cβ, and d) the two-third restraints with lowest I_Cβ. The experiment was then repeated using s_Cβ instead of I_Cβ,.

Once again the trends of the results are the same for both proteins and for both I_Cβ and s_Cβ experiments. a) Using the one-third restraints with highest I_Cβ shifts the RMSD distribution into the same range that is obtained when using all of the available distance restraints (Figure 3E and B, T4-lysozyme; Figure 3M and J, αA-crystallin). b) Using the one-third restraints with lowest I_Cβ only slightly shifts the RMSD distribution towards lower RMSDs and is similar to that in the absence of distance restraints (Figure 3F and A, T4-lysozyme; Figure 3N and I, αA-crystallin). c) There is little shift in the RMSD distribution when the two-third restraints with highest I_Cβ are used compared to only one-third (Figure 3G and E, T4-lysozyme; Figure 3O and M, αA-crystallin); the extra restraints increase the number of lower RMSD models that are created. d) When the two-third restraints with lowest I_Cβ are used, there is a small shift in the RMSD distribution compared to when no distance restraints are used (Figure 3H and A, T4-lysozyme; Figure 3P and I, αA-crystallin). However, this shift is not nearly as drastic as the shift obtained when the one-third most informative distant restraints are used.

Using s_Cβ to select restraints instead of I_Cβ results in only a slight reduction in model quality and slight variation in total I_Cβ of the selected restraints compared to that of the restraints employed in the first experiment (Figure 3E – H, T4-lysozyme, Figure 3M – P, αA-crystallin). This indicates maximal sequence separation can be used to effectively define spin label placement and select for restraints with large I_Cβ. However, it should be noted that some of the sites for spin labeling T4-lysozyme were chosen with the crystal structures at hand which might bias the restraint sets for increased information content. Furthermore, this experiment does not test how spin labels should be distributed along the sequence. Additional experiments indicate that, besides maximizing s_Cβ, a uniform distribution of spin labels over the sequence is optimal (data not shown).

Rosetta folding of T4-lysozyme and αA-crystallin

No accessibility restraints were used in the large scale folding simulations, due to their minimal influence on structure determination. Of the 500,000 models built for T4-lysozyme, the lowest RMSD obtained was 2.39Å with a total of 117 models having an RMSD value smaller than 3.5Å. Of the 500,000 models built for αA-crystallin, the lowest RMSD obtained was 3.36Å with a total of 46 models having an RMSD value smaller than 4.0Å.

Filtering the 500,000 models of T4-lysozyme and αA-crystallin reduces the number considered for high-resolution refinement to a manageable number and enriches the high-resolution refinement pool for low RMSD models. Enrichment is measured as the fraction of low RMSD models in the filtered ensemble divided by the fraction of low RMSD models in the original ensemble. For T4-lysozyme, requiring full agreement with all distance restraints and an overall Rosetta score better than −35 points prunes down the number of candidate structures to 10,906, keeping 27 models with RMSD values smaller than 3.5Å. This enriches low RMSD (≤ 3.5Å) models in the dataset by a factor of 27 / 10,906 ÷ 117 / 500,000 = 10.6.

For αA-crystallin, in order to keep approximately 10,000 structures for high-resolution refinement, models were required to have an overall Rosetta score better than or equal to −75, β-strand pairing score better than −31, and total sum of all distance violations smaller than 3.0Å. These criteria limit the number of structures to 9,796. Of the 9,796 models, 26 models have an RMSD of less than 4.0Å, which is an enrichment of low RMSD (≤4.0Å) models of 26 / 9,796 ÷ 46 / 500,000 = 28.8.

Structure Determination of T4-lysozyme and αA-crystallin

After high-resolution refinement, models were filtered by agreement with distance restraints. T4-lysozyme models were again required to be in full agreement with all distance restraints. αA-crystallin models were required to have sum total distance restraint violations of less than 1 Å. The remaining models were sorted by Rosetta full-atom energy, and the lowest energy model for each protein was compared to the crystal structure of T4-lysozyme and the comparative model of αA-crystallin. For RMSD analysis, loop regions of the α-helical and β-sandwich domains were disregarded.

In addition to RMSD analysis, the agreement of side chain conformations can be captured by comparing the dihedral angles χ_1...4. A specific set of such angles χ_1...4 is called a “rotamer” (Dunbrack and Karplus 1993; Dunbrack 2002). If all angles χ_1...4 of an amino acid side chain deviate less than 60°, the rotamer is the same and the conformation is closely recovered. The number reported for side chain conformation comparison is the percentage of non-agreeing rotamers (Figure 4B and Figure 4D).

Figure 4.

Correlation of de novo models' accuracy with the energy of the de novo models. A) and C) The non-loop RMSD versus Rosetta energy for T4-lysozyme and αA-crystallin models, respectively. B) and D) The percentage of incorrectly built side chain conformations versus Rosetta energy for T4-lysozyme and αA-crystallin models, respectively. In all diagrams the minimized crystal structure or comparative model is depicted as a circle; the lowest energy model is shown as a square.

The Rosetta energy of T4-lysozyme de novo models decreases as their RMSD and side chain rotamer disagreement to the native structure diminish (Figure 4A, Figure 4B). This allows the selection of high-resolution models based on energy alone. The lowest energy de novo model achieves an RMSD to the crystal structure of 1.0Å in the α-helical domain and 2.0Å over all modeled residues (Figure 5A). 80% of all rotamers are in agreement with the crystal structure.

Figure 5.

Overlay of lowest energy de novo models on crystal structure or comparative model. A) and B) For T4-lysozyme and αA-crystallin, respectively, superimposition of the lowest energy model (rainbow colored) with the crystal structure or comparative model (gray). The backbone is given as a ribbon diagram. Side chains of T4-lysozyme and of the β-sandwich of αA-crystallin are shown as stick models without hydrogen atoms.

The Rosetta energy of αA-crystallin de novo models decreases as their structure approaches that of the comparative model (Figure 4C). However, side chain rotamer agreement does not correlate with Rosetta energy (Figure 4D). The lowest energy de novo model achieves an RMSD of 2.6Å for the β-sandwich and 4.0Å over the whole protein (Figure 5B). The rotamer agreement in the β-sandwich is 54.5%. Note that similar rmsds and side chain agreements are also found between the comparative model and two additional comparative models based on alternative templates (data not shown).

Discussion

Structure determination from sparse EPR restraints

The major conclusion of this paper is that structural restraints obtained from EPR analysis of spin labeled proteins can be used in combination with de novo prediction methods to determine atomic-detail structures of proteins. Furthermore, the structural interpretation of the EPR data for the purpose of de novo modeling does not require a detailed understanding of the position or conformation of the spin label relative to the backbone. A “motion-on-a-cone” model can provide inter-Cβ distance restraints and the spin label accessibility can restrain the number of close neighbors when building a structural model.

Only a few specific distance restraints are needed in order to add substantial information and restrict the accessible fold space significantly. Using all twenty-five distance restraints, 50% T4-lysozyme models have an RMSD smaller than 8Å compared to less than 5% in their absence. This means that less than one distance restraint per four amino acids was sufficient to focus the de novo structure determination method on sampling the correct fold in more than half of all runs of T4-lysozyme. Only 0.074 distance restraints per residue are needed in order to obtain an equivalent RMSD distribution if the eight restraints between amino acids farthest apart in sequence are used. A significant increase in the quality of models created with distance restraints was also noted for αA-crystallin. By using the distance restraints between amino acids farthest apart in sequence, only 0.136 distance restraints per residue are needed in order to obtain an RMSD distribution similar to that seen when using all restraints. The reduced frequency at which the correct fold is obtained compared with T4-lysozyme can be attributed to the more challenging folding pathway for a β-sandwich.

Relative importance of accessibility and spatial restraints

Distance restraints – even with the comparably large uncertainties resulting from EPR measurements – are more valuable for de novo protein structure determination than solvent accessibility data. Whereas distances reflect specific geometric relationships, the solvent accessibility reflects a convolution of local interactions and rather unspecific interactions with the solvent. The Rosetta energy function already contains knowledge-based terms for these two types of interactions: the amino acid pair potential and the environment potential. The pair potential describes the likelihood of two amino acid types to be spatially close. The environment potential describes the likelihood of an amino acid to be exposed to the solvent or buried in the core. Thus, the EPR accessibility measurements, which reflect the expected buried/exposed distribution, add little information beyond the empirical potentials used in Rosetta.

It should be noted, however, that the importance of the spin label accessibility restraints may be understated by the use of water-soluble proteins as a test case. In general, de novo prediction methods for secondary structure and for sequence–specific residue environment are quite accurate for water soluble proteins (Rost and Sander 1993; Jones 1999; Rost 2001; Meiler 2003). In contrast, the apolar character of the membrane core makes a computational distinction between membrane and protein core more difficult. Thus, experimental accessibilities are expected to be critical in defining the secondary structure and topology of initial models for membrane proteins.

Structural interpretation of EPR parameters

The fundamental assumption in the “motion-on-a-cone” model is that the spin label cannot point towards the interior of the protein. Thus, possible specific interactions of spin label and protein are disregarded. This is manifest as a bias towards over-predicting the frequency with which large (> 4Å) d_SL - d_Cβ values occur. Additionally, the model underestimates the frequency with which low (< 4Å) d_SL - d_Cβ values occur. This is because the spin labels cannot adopt conformations that closely mimic non-spherical arrangements on the surface of proteins such as β-strands and α-helices. A more precise estimation of the distribution of d_SL - d_Cβ values might be possible as a comprehensive understanding of spin label rotamers emerges (Langen, Oh et al. 2000).

While there is a robust linear relation between the experimental spin label accessibility, e_SL, and the predicted accessibility, e_Cβ, the applied consensus fit procedure used to obtain the relation has two disadvantages. First, because a comparative model was used in the development of the consensus linear regression model, using it for de novo folding simulations is somewhat circular. Second, part of the accessibility data for αA-crystallin are influenced by oligomerization and had to be excluded because only a model of monomeric state of αA-crystallin is computationally feasible. Nevertheless, using this relation in simulations is arguably an acceptable test of the usefulness of such data for de novo protein folding.

Conclusion

De novo protein structure prediction samples as much of the conformational space as possible in order to find the native structure; the number of models reflects the extent of sampling. The increase in the number of high quality models indicates that the conformational search space has been reduced by the restraints, which allows the remaining space to be sampled more densely. It is remarkable that sparse distance restraints with as large an uncertainty as those obtained from the “motion-on-a-cone” model provide such a drastic improvement in the RMSD distribution of models. The most efficient reduction in conformational search space results from restraints between residues far apart in the primary sequence but close in Euclidean space. A uniform distribution of restraints throughout the protein sequence should be taken into consideration in order to maximize efficiency. Experimental accessibility data add little for structure elucidation of soluble proteins. Less than 0.25 restraints per amino acid are sufficient for protein structure elucidation at atomic detail.

The tendency for side chains to achieve their native rotamer as the protein model backbone approaches its native conformation has been termed “backbone memory” in protein design (Kuhlman and Baker 2000) and was also observed in very accurate high-resolution de novo protein structure prediction (Bradley, Misura et al. 2005). As a protein model backbone approaches its native conformation, backbone memory allows Rosetta to accurately place side chains into their native rotamer (Kuhlman and Baker 2000). This is demonstrated with T4-lysozyme; 80% of all rotomers are correct in the lowest energy de novo model although no side chain conformational restraints are used. For αA-crystallin definite placement of side chain atoms cannot be conclusively analyzed, since no high-resolution crystal structure is available. The energies between comparative and de novo models are similar; however, no convergence in side chain conformation was achieved. Therefore, it remains unclear whether the rotamers predicted by the comparative or the de novo model are more accurate.

Overall, this benchmark study sets the stage for application of EPR restraints to protein targets where no structural model is yet available. Advancements in protein folding algorithms and incorporation of other experimental techniques will further improve the efficiency and accuracy of de novo protein structure determination from sparse experimental data.

Experimental Procedures

Introduction of site-directed spin labels and EPR conditions

For the introduction of spin labels, cysteine residues were systematically introduced into the (cysteine-free) T4-lysozyme and αA-crystallin amino acid sequences through single or double point mutations (Koteiche, Berengian et al. 1998; Borbat, Mchaourab et al. 2002; Altenbach, Froncisz et al. 2005). After recombinant protein expression and purification, the mutant was reacted with methanethiosulfonate nitroxide reagent. A total of 25 double mutants and 57 single mutants of T4-lysozyme (Table 1, and Table 2) and 36 double mutants and 87 single mutants of αA-crystallin (Table 3, and Table 4) resulted in the restraints used for the current analysis. Sample preparation and EPR measurement have been described elsewhere (Mchaourab, Lietzow et al. 1996; Koteiche and Mchaourab 1999).

Table 1.

T4-lysozyme EPR distance restraints in comparison with crystal structure distances.

AA1-AA2[a]	dCβ (Å)[b]	dSL (Å)[c]	σSL (Å)[d]	dSL+σSL+ 2.5 (Å)[e]	dSL-σSL−12.5 (Å)[f]	Reference
061−135	37.7	47.2	2.2	51.9	32.5	(Borbat, Mchaourab et al. 2002)
065−135	34.3	46.3	2.2	51.0	31.6	(Borbat, Mchaourab et al. 2002)
061−086	34.5	37.5	2.0	42.0	23.0	(Borbat, Mchaourab et al. 2002)
065−086	28.9	37.4	2.7	42.6	22.2	(Borbat, Mchaourab et al. 2002)
080−135	26.7	36.8	1.0	40.3	23.3	(Borbat, Mchaourab et al. 2002)
061−080	28.7	34.0	2.2	38.7	19.3	(Borbat, Mchaourab et al. 2002)
065−080	22.6	26.5	3.8	32.8	10.2	(Borbat, Mchaourab et al. 2002)
119−131	13.2	25.0	5.0	32.5	7.5	new data
123−131	14.6	23.0	5.0	30.5	5.5	new data
065−076	16.8	21.4	2.8	26.7	6.1	(Borbat, Mchaourab et al. 2002)
116−131	11.1	19.0	10.0	31.5	0.0	new data
119−128	10.4	19.0	4.0	25.5	2.5	new data
140−151	15.5	18.0	9.0	29.5	0.0	new data
089−093	9.8	16.0	3.0	21.5	0.5	new data
086−119	10.0	15.0	3.0	20.5	0.0	new data
120−131	10.5	14.0	3.0	19.5	0.0	new data
127−151	9.6	14.0	2.4	18.9	0.0	new data
140−147	10.1	13.0	7.0	22.5	0.0	new data
131−150	8.7	5.7	0.4	8.6	0.0	new data
127−154	5.9	7.0	3.0	12.5	0.0	new data
131−154	9.5	6.5	4.0	13.0	0.0	new data
134−151	10.7	7.0	0.8	10.3	0.0	new data
131−151	10.4	9.0	8.0	19.5	0.0	new data
088−100	8.9	<6.0	3.0	11.5	0.0	new data
089−096	8.4	<6.0	3.0	11.5	0.0	new data

^[a]indices of spin labeled amino acids with respect to the crystal structure.

^[b]Cβ distance as reported in the crystal structure.

^[c]spin label distance as observed by EPR.

^[d]standard deviation as observed by EPR.

^[e]maximum Cβ atom distance predicted by cone model.

^[f]minimum Cβ atom distance predicted by cone model.

Table 2.

T4-lysozyme EPR solvent accessibility in comparison with crystal structure.

AA[a]	eCβ[b]	eSL[c]	eCβ[d]
086	9	0.36	7.7
093	6	0.41	5.8
094	10	0.28	10.8
096	10	0.23	12.7
097	12	0.17	15.0
100	13	0.15	15.8
101	14	0.13	16.5
102	13	0.14	16.2
103	14	0.16	15.4
104	10	0.28	10.8
105	10	0.33	8.9
106	7	0.31	9.7
108	10	0.32	9.3
109	6	0.45	4.3
111	12	0.23	12.7
113	8	0.40	6.2
114	11	0.30	10.0
115	6	0.36	7.7
116	7	0.52	1.6
117	10	0.29	10.4
118	13	0.21	13.5
119	8	0.35	8.3
120	11	0.23	12.7
121	16	0.10	17.7
122	9	0.36	7.7
123	6	0.39	6.6
124	6	0.42	5.4
125	8	0.40	6.2
126	11	0.32	9.3
127	7	0.44	4.7
128	8	0.42	5.4
129	13	0.14	16.2
130	12	0.30	10.0
131	8	0.37	7.4
132	10	0.35	8.1
133	15	0.23	12.7
134	10	0.30	10.0
135	6	0.38	7.0
136	9	0.31	9.7
137	5	0.41	5.8
138	11	0.29	10.4
139	13	0.33	8.9
140	6	0.37	7.4
141	8	0.25	11.9
142	9	0.27	11.2
143	8	0.39	6.6
144	6	0.36	7.7
145	9	0.24	12.3
146	13	0.27	11.2
147	11	0.31	9.7
148	12	0.28	10.8
149	14	0.21	13.5
150	12	0.29	10.4
151	11	0.37	7.4
153	14	0.24	12.3
154	10	0.34	8.5
155	8	0.25	11.9

^[a]indices of spin labeled amino acids with respect to the crystal structure.

^[b]number of Cβ atom neighbors in the crystal structure.

^[c]spin label accessibility as observed by EPR (Sompornpisut, Mchaourab et al. 2002).

^[d]number of Cβ atom neighbors predicted by the consensus linear regression relation.

Table 3.

αA-crystallin EPR distance restraints in comparison with comparative model.

AA1-AA2[a]	dCβ (Å)[b]	dSL (Å)[c]	σSL (Å)[d]	dSL +2.5+ σSL (Å)[e]	dSL—σSL—12.5 (Å)[f]	Reference
065−072	4.5	6.0	3.0	11.5	0.0	(Koteiche and Mchaourab 1999)
068−122	13.0	19.5	5.0	27.0	2.0	new data
070−072	6.3	6.6	1.5	10.6	0.0	(Koteiche and Mchaourab 1999)
070−074	12.8	12.0	4.0	18.5	0.0	(Koteiche and Mchaourab 1999)
070−142	5.2	6.0	2.0	10.5	0.0	(Koteiche and Mchaourab 1999)
072−142	5.7	5.5	2.0	10.0	0.0	(Koteiche and Mchaourab 1999)
073−118	5.6	6.0	3.0	11.5	0.0	new data
074−140	6.0	6.0	2.0	10.5	0.0	(Koteiche and Mchaourab 1999)
084−099	4.5	7.6	0.6	10.7	0.0	(Koteiche, Berengian et al. 1998)
084−101	8.6	7.5	8.0	18.0	0.0	(Koteiche, Berengian et al. 1998)
088−095	4.6	6.0	0.7	9.2	0.0	(Koteiche, Berengian et al. 1998)
088−097	7.7	7.0	0.8	10.3	0.0	(Koteiche, Berengian et al. 1998)
090−093	6.8	6.9	1.1	10.5	0.0	(Koteiche, Berengian et al. 1998)
090−099	18.7	18.5	4.0	25.0	2.0	(Koteiche, Berengian et al. 1998)
090−126	11.1	12.0	4.0	18.5	0.0	(Koteiche and Mchaourab 1999)
092−123	10.4	12.0	6.0	20.5	0.0	(Koteiche and Mchaourab 1999)
095−117	5.0	7.6	1.0	11.1	0.0	(Koteiche, Berengian et al. 1998)
101−103	5.4	6.3	0.8	9.6	0.0	(Koteiche and Mchaourab 1999)
101−111	11.9	7.6	1.0	11.1	0.0	(Koteiche, Berengian et al. 1998)
103−105	7.5	9.5	8.0	20.0	0.0	(Koteiche and Mchaourab 1999)
105−111	11.6	18.0	9.0	29.5	0.0	(Koteiche and Mchaourab 1999)
119−123	10.5	13.0	6.0	21.5	0.0	(Koteiche and Mchaourab 1999)
119−125	17.1	23.0	3.0	28.5	7.5	(Koteiche and Mchaourab 1999)
119−127	21.2	24.0	6.0	32.5	5.5	(Koteiche and Mchaourab 1999)
119−130	19.3	23.0	5.0	30.5	5.5	(Koteiche and Mchaourab 1999)
119−132	21.9	23.0	4.0	29.5	6.5	(Koteiche and Mchaourab 1999)
125−144	4.0	7.0	4.0	13.5	0.0	(Koteiche and Mchaourab 1999)
127−142	13.1	20.0	4.0	26.5	3.5	new data
128−144	5.0	4.0	4.0	10.5	0.0	(Koteiche and Mchaourab 1999)
128−146	10.4	13.0	4.0	19.5	0.0	(Koteiche and Mchaourab 1999)
130−144	9.7	6.0	1.0	9.5	0.0	(Koteiche and Mchaourab 1999)
131−133	8.0	7.0	1.0	10.5	0.0	(Koteiche and Mchaourab 1999)
131−134	12.3	17.0	4.0	23.5	0.5	(Koteiche and Mchaourab 1999)
132−135	11.5	16.0	5.0	23.5	0.0	(Koteiche and Mchaourab 1999)
132−142	8.8	6.0	1.0	9.5	0.0	(Koteiche and Mchaourab 1999)
132−144	16.1	9.0	5.0	16.5	0.0	(Koteiche and Mchaourab 1999)

^[a]indices of spin labeled amino acids with respect to the protein sequence.

^[b]Cβ atom distance in comparative model.

^[c]spin label distance as observed by EPR.

^[d]standard deviation as observed by EPR.

^[e]maximum Cβ atom distance predicted by cone model.

^[f]minimum Cβ atom distance predicted by cone model.

Table 4.

αA-crystallin EPR solvent accessibility in comparison comparative model.

AA[a]	eCβ[b]	eSL[c]	eCβ[d]
060	5	0.01	10.9
061	5	0.04	10.5
062	11	0.00	11.0
063	6	0.02	10.8
064	10	0.05	10.5
065	9	0.07	10.2
066	10	0.05	10.4
067	9	0.04	10.6
068	5	0.19	8.9
069	6	0.21	8.7
070	10	0.16	9.3
071	16	0.05	10.5
072	12	0.06	10.3
073	15	0.07	10.2
074	13	0.05	10.5
075	12	0.02	10.8
076	11	0.02	10.8
077	12	0.01	10.9
078	9	0.02	10.8
079	11	0.03	10.6
080	14	0.01	10.9
081	10	0.09	10.0
082	8	0.27	8.0
083	5	0.29	7.8
084	10	0.25	8.2
085	16	0.05	10.4
086	7	0.22	8.5
087	12	0.14	9.5
088	8	0.35	7.2
089	11	0.18	9.0
090	9	0.26	8.2
091	4	0.41	6.5
092	7	0.35	7.1
093	12	0.27	8.1
094	17	0.06	10.4
095	13	0.23	8.4
096	15	0.14	9.5
097	12	0.21	8.6
098	12	0.01	10.9
099	9	0.13	9.6
100	16	0.01	10.9
101	9	0.19	9.0
102	10	0.14	9.5
103	6	0.17	9.2
104	6	0.15	9.3
105	12	0.19	8.9
106	5	0.10	9.9
107	7	0.08	10.1
108	4	0.07	10.2
109	10	0.17	9.1
110	3	0.04	10.6
111	7	0.15	9.3
112	9	0.00	11.0
113	9	0.08	10.1
114	11	0.01	10.9
115	9	0.15	9.4
116	11	0.00	11.0
117	10	0.16	9.3
118	12	0.01	10.9
119	7	0.26	8.1
120	11	0.01	10.9
121	5	0.19	8.9
122	8	0.43	6.2
123	7	0.30	7.7
124	16	0.08	10.1
125	9	0.34	7.3
126	7	0.20	8.8
127	5	0.38	6.8
128	9	0.32	7.5
129	14	0.21	8.7
130	9	0.18	9.0
131	9	0.16	9.3
132	8	0.22	8.6
133	11	0.12	9.7
134	7	0.07	10.3
135	4	0.19	8.9
136	5	0.08	10.1
137	12	0.08	10.1
138	10	0.02	10.8
140	13	0.05	10.5
141	15	0.01	10.8
142	11	0.11	9.7
143	14	0.03	10.7
144	11	0.20	8.8
145	12	0.10	9.8
146	7	0.41	6.5
147	2	0.39	6.7

^[a]indices of spin labeled amino acids with respect to the protein sequence.

^[b]number of Cβ atom neighbors in the crystal structure.

^[c]spin label accessibility as observed by EPR (Koteiche, Berengian et al. 1998; Koteiche and Mchaourab 1999).

^[d]number of Cβ atom neighbors predicted from the consensus linear regression relation.

EPR distance measurements

For T4-lysozyme, 25 distances were measured (Table 1, Figure 2A). Distances derived from Double Electron-Electron Resonance (DEER) or DQC experiments (Borbat, Mchaourab et al. 2002; Jeschke 2002; Borbat and Freed 2007) were distributed in different areas of the molecule with predicted distances larger than 25Å. They provide geometric restraints on the global fold. CW-EPR was used to measure distances between neighboring helices. For each pair of interacting helices, doubly labeled mutant sets were created by designating a reference spin label in one helix and moving another spin label along the exposed surface of the second helix.

The αA-crystallin EPR data (CW-EPR, Table 3) consists of 36 distances, including β-strand to β-strand and β-strand to loop distances covering most of the overall topology of the molecule (Koteiche, Berengian et al. 1998; Koteiche and Mchaourab 1999) (Figure 2E). For both T4-lysozyme and αA-crystallin, when measurements provided multiple distances, the most contributing distance was used.

For the CW-EPR experiments, dipolar coupling between spin labels was analyzed both in the liquid state and in frozen solutions using a modification of the deconvolution method (Rabenstein and Shin 1995). This approach requires two EPR spectra of the double mutant: one in the absence and one in the presence of the dipolar interaction (Figure 6A, B, respectively). The former is obtained from the digital sum of the spectra of each single mutant. A Levenberg-Marquardt algorithm was used to minimize the difference between the experimental EPR spectrum of the double mutant and the spectrum obtained from the convolution of a broadening function with the EPR spectrum of the corresponding sum of single mutants. The broadening function consisted of either one or two Gaussian distributions for the distance between spin labels (Figure 6C). The relatively wide distance distributions obtained is consistent with a highly dynamic motional state of the spin label obtained at the predominantly exposed sites. The results obtained in the solid and liquid states are in agreement both in terms of the average distance and the overall distribution as previously reported (Altenbach, Oh et al. 2001).

Figure 6.

Distance measurements at room temperature and in the solid state between spin labels using EPR. A) Representative reference EPR in the absence of dipolar coupling obtained from the digital sum of the corresponding single mutant spectra. B) Spectra of double mutants along with the non-linear least squares fit obtained by the convolution method as described in the experimental methods section. C) Distance distributions obtained from CW-EPR spectra. D) Distance measurements by DEER for representative double mutants. E) Raw DEER signals were background corrected and then fit using Tikhonov regularization to obtain F) average distances and distance distributions.

DEER measurements were performed on a Bruker 580 pulsed EPR spectrometer, using a standard four pulse protocol (Jeschke 2002). Experiments were performed at 80 K using Ficoll as cryoprotectant. Sample concentration was 200 μM and sample volume 20 μl. DEER signals were analyzed by the Tikhonov regularization (Chiang, Borbat et al. 2005) to determine average distances and distributions in distance, P(r), as illustrated in Figure 6 D, E, and F.

EPR accessibility measurements

For T4-lysozyme, e_SL of 57 spin labels was measured (Sompornpisut, Mchaourab et al. 2002) (Table 2). For αA-crystallin, e_SL of 87 spin labels was measured (Table 4). Accessibility is assessed by measuring the Heisenberg exchange rate between the nitroxide spin label and either molecular oxygen, in the case of T4-lysozyme, or NiEDDA, in the case of αA-crystallin. For the latter, power saturation measurements were carried out under nitrogen and in the presence of 3 mM NiEDDA (Farahbakhsh, Altenbach et al. 1992). e_SL was calculated as previously described (Farahbakhsh, Altenbach et al. 1992; Altenbach, Froncisz et al. 2005).

αA-crystallin comparative model preparation

The 88 amino acid C-terminal domain of αA-crystallin was submitted to the BioInfo metaserver (Fischer 2000; Kelley, MacCallum et al. 2000; Shi, Blundell et al. 2001; Ginalski, Elofsson et al. 2003; Karplus, Karchin et al. 2003; McGuffin and Jones 2003; Ginalski, von Grotthuss et al. 2004; Rost, Yachdav et al. 2004; Bryson, McGuffin et al. 2005; Soding, Biegert et al. 2005; Finn, Mistry et al. 2006). The server identified three heat shock proteins (PDB identifiers 1gmeA (van Montfort, Basha et al. 2001), 1shsA (Kim, Kim et al. 1998), and 2bolA (Stamler, Kappe et al. 2005)) as possible templates with a 3D-Jury score of over 60, where a score over 40 indicates a ∼90% chance that the identified proteins have the same fold as the submitted amino acid sequence (Ginalski, Elofsson et al. 2003). Obtaining the correct fold is the most important and difficult aspect of de novo protein folding, so having such a large likelihood that the identified proteins have the same fold as αA-crystallin is essential to ensuring the comparative model provides an adequate benchmark with which the fold of de novo models can be compared. There was approximately 20% sequence homology between the αA-crystallin amino acid sequence and the three template sequences.

A multiple sequence alignment was performed for the αA-crystallin amino acids with the template proteins. The aligned αA-crystallin amino acids were then mapped onto the template proteins' atomic coordinates and Rosetta was used to reconstruct the loop regions of αA-crystallin while holding the β-sandwich region fixed. Afterwards, Rosetta was used to perform a high-resolution refinement of the αA-crystallin comparative models.

The lowest energy comparative model was used to compare against the de novo Rosetta models. This model was based on the PDB structure 1gmeA (van Montfort, Basha et al. 2001). The model is 2.3Å RMSD to a previously published (Koteiche and Mchaourab 1999) comparative model based on a different template protein (Hsp16.5) but achieves a lower Rosetta energy after both models are refined at high-resolution.

The spin label “motion-on-a-cone” model

A simple cone model for the relative position of the spin label with respect to the Cβ of an amino acid was developed using three assumptions. First, the spin label's motion follows the perimeter of the base of a right circular cone with an opening angle of 90° whose vertex is the Cβ. The average position of this motion is √2/2 of length of the extended chain (8.5Å (Borbat, Mchaourab et al. 2002)) which gives a maximal effective distance between spin label and Cβ of 6Å (Figure 1A, B). Second, the protein is globular. Third, the angle defined by the center of the protein, the Cβ, and the spin label is between 120° and 180°, and, therefore, spin labels point away from the interior of the protein (Figure 1B, C). The “motion-on-a-cone” model can also be adapted for spin labels of different linking arms or ring substituents that restrict the amplitude of its motion.

Using the “motion-on-a-cone” model to translate EPR spin label distances into structural restraints

The difference between d_SL and d_Cβ given by the model described above was analyzed using the software package Mathematica (2005):

1. An ellipsoid with the main radii 10Å ≤ r_x ≤ r_y ≤ r_z ≤ 20Å was created with otherwise randomly chosen r_x, r_y, and r_z. Its center is C.

2. Two points, Cβ_i and Cβ_j, on the surface of this ellipsoid were selected by randomly choosing the polar coordinates φ_i,j, and Ψ_i,j. From these points, d_Cβ is computed as the Euclidean distance.

3. Two numbers between 120° and 180° for the angle C–Cβ_i–SL_i and C–Cβ_j–SL_j are chosen randomly and the position of the spin labels, SL_i and SL_j, is computed. From these points, d_SL is computed as the Euclidean distance.

4. The difference d_SL–d_Cβ is computed.

5. Steps 1−4 are repeated 10,000 times and the values d_SL–d_Cβ are plotted as a histogram (Figure 1D).

This analysis of the difference between d_SL and d_Cβ showed that (d_SL+2.5Å) ≥ d_Cβ ≥ (d_SL−12.5Å) (Figure 1D).

σ_SL, which is the experimentally determined standard deviation in d_SL, is a measure of the magnitude of the spin label's motion, or, its static distribution relative to the Cβ. Since increased magnitude an of motion increases the ambiguity of the derived d_Cβ, σ_SL is added as an additional allowance to the restraint which gives (d_SL+σ_SL+2.5Å) ≥ d_Cβ ≥ (d_SL–σ_SL–12.5Å).

Development of a model to translate EPR spin label solvent accessibility into structural restraints

Obtaining a structural restraint from EPR spin label solvent accessibility is accomplished by building a consensus linear regression relation of e_SL to e_Cβ in a three step procedure:

1. Using the crystal structure of T4-lysozyme, e_Cβ was computed for all residues with known e_SL (Table 2). A linear regression was fit to a plot of e_Cβ of a residue versus the corresponding e_SL, yielding the relation e_Cβ = (0.71–e_SL)•24.23 with a correlation coefficient of −0.80. Using this relation to calculate the number of neighbors for a residue gives e_Cβ^fit for that residue. The standard deviation (σ_Cβ) of e_Cβ from e_Cβ^fit was calculated to be 1.65.

2. For those residues in αA-crystallin that have an experimentally determined accessibility (Table 4), e_Cβ was determined using the comparative model of αA-crystallin. In addition, the equation from 1) above was used to calculate the number of neighbors, e_Cβ^fit, for each residue. Amino acids were excluded from a linear fitting of e_Cβ versus e_SL when |e_Cβ–e_Cβ^fit| > 2•σ_Cβ. This procedure was necessary in order to exclude amino acids in αA-crystallin that show reduced experimental accessibility due to intermolecular contacts with other αA-crystallin units in the oligomeric protein (Figure 2F, H). Fitting a linear regression to the remaining data gives a relation similar to that seen for T4-lysozyme: e_Cβ = (0.72–e_SL)•21.63 with a correlation coefficient of −0.87.

3. Combining the data for both proteins in a single consensus linear regression model yields e_Cβ = (0.76–e_SL)•20.87 with a correlation coefficient of −0.83 (Figure 2D, and Figure 2H).

Implementation of structural restraints for de novo structure determination

The distance restraints are used as an additional penalty in the energy function of Rosetta. This penalty is zero if d_Cβ lay within the range predicted from d_SL. As d_Cβ ventures outside this range a quadratic penalty function is applied. The detailed implementation of this penalty function and its use to guide the folding simulation is described in detail in the respective RosettaNMR publications (Bowers, Strauss et al. 2000; Rohl and Baker 2002).

Similarly, ${\left({\mathrm{e}}_{\mathrm{C}\beta }-e_{C\beta }^{\text{consensus mode}l}\right)}^2$ is used as a quadratic penalty function for the accessibility data. The relative weight of this penalty function was optimized by a series of experiments varying its weight in a wide range of two orders of magnitude.

Rosetta folding simulations

De novo model generation using Rosetta was performed in four steps:

1. The protein is folded using RosettaNMR with EPR distance restraints to guide the simulation (Bowers, Strauss et al. 2000; Rohl and Baker 2002). In this step, amino acid side chains are embraced in a single super-atom – a “centroid” (Simons, Kooperberg et al. 1997).

2. Choosing the models with lowest energy and best agreement with experimental restraints prunes the large number of ∼500,000 models from Step 1 to ∼10,000.

3. Models obtained from Step 2 are refined to high-resolution. High-resolution refinement is described as follows:

   • After replacing side chain centroids with full-atom side chain representations from a backbone dependent rotamer library (Dunbrack and Karplus 1993), an iterative protocol of all-atom gradient minimization and side chain repacking is repeated eight times. The details of the protocol are published elsewhere (Bradley, Misura et al. 2005; Misura and Baker 2005). No restraints were used during these refinement simulations in order to fully leverage the discriminative power of the Rosetta energy function (Kuhlman and Baker 2000; Bradley, Malmstrom et al. 2005; Bradley, Misura et al. 2005; Misura and Baker 2005; Misura, Chivian et al. 2006).

4. Models from Step 3 are again filtered for good agreement with the experimental restraints.

Specific standard Rosetta procedures were used which are described in detail elsewhere (Simons, Kooperberg et al. 1997; Simons, Ruczinski et al. 1999; Bowers, Strauss et al. 2000; Bonneau, Strauss et al. 2001; Rohl and Baker 2002; Meiler, Bradley et al. 2003; Rohl, Strauss et al. 2004; Bradley, Malmstrom et al. 2005). Secondary structure predictions were obtained from the primary sequence of the C-terminal 107 amino acids of T4-lysozyme and the C-terminal 88 amino acid primary sequence of αA-crystallin using Jufo (Meiler, Müller et al. 2001; Meiler and Baker 2003), PsiPred (Jones 1999), and Sam (Karplus, Sjolander et al. 1997). All T4-lysozyme and αA-crystallin homologs were excluded from the protein database prior to modeling in order to simulate structure elucidation of an unknown protein fold as closely as possible. A library of 200 fragments for each position was built.

Models were obtained in 500,000 independent simulations on a cluster in Vanderbilt University's Advanced Computing Center for Research & Education (ACCRE) using up to 300 parallel 2.2 GHz JS20 IBM PowerPC processors. The average time to complete a model was approximately 100s for T4-lysozyme and 180s for αA-crystallin. The high-resolution refinement protocol requires about 500s of computation time per model.