Edited by Robert L. Baldwin, Stanford University Medical Center, Stanford, CA, and approved December 15, 2008
↵1A.V.G. and M.M.P. contributed equally to this work. (received for review August 20, 2008)
Here, we report the application of a computational approach that allows the rational design of enzymes with enhanced thermostability while retaining full enzymatic activity. The approach is based on the optimization of the energy of charge–charge interactions on the protein surface. We experimentally tested the validity of the approach on 2 human enzymes, acylphosphatase (AcPh) and Cdc42 GTPase, that differ in size (98 vs. 198-aa residues, respectively) and tertiary structure. We show that the designed proteins are significantly more stable than the corresponding WT proteins. The increase in stability is not accompanied by significant changes in structure, oligomerization state, or, most importantly, activity of the designed AcPh or Cdc42. This success of the design methodology suggests that it can be universally applied to other enzymes, on its own or in combination with the other strategies based on redesign of the interactions in the protein core.
Until man duplicates a blade of grass, nature will laugh at his so-called scientific knowledge.
Rational engineering of proteins to enhance stability and yet retain their enzymatic activity is well motivated (1). One motivation is the practical significance of expanding the use of enzymes in many areas of the modern world, including protein therapeutics, enzymes for food industry, diagnostics, and other areas of industrial biotechnology. Another motivation is validation of the existing scientific knowledge. In this case, predictions made by the existing models for protein stability are subjected to thorough experiments, testing their applicability to protein design. In this paper, we present the results of rational design of enzymes with enhanced stability and unchanged enzymatic activity. This approach has 2 major differences from previously described successful protein design methods (,2–,5): (i) it concentrates only on the residues on the protein surface, and (ii) it optimizes just one type of interactions, namely, charge–charge interactions on the protein surface (6–,15).
One of the most important aspects of engineering proteins with enhanced stability, retaining the enzymatic activity, is often forgotten. However, for all of these design efforts to be practically useful, it is important that the engineered proteins retain their biological and enzymatic activity. This issue is particularly important when enhanced protein stability is achieved by redesigning the charge–charge interactions on the protein surface. Such redesign can lead to several potentially detrimental effects on the activity: (i) it can affect the electrostatic potential in the active center, thus reducing or even abolishing the activity; (ii) it can affect substrate/product binding and again reduce or abolish the enzymatic activity; or (iii) it can have effects on the kinetics of substrate binding and, thus, lower the activity via reduced rates of electrostatic steering (3, ,16–,22).
To this end, it is important to test whether redesign of the surface charges that leads to the increase in stability affects enzyme activity. As the test model systems, we chose 2 human enzymes, acylphosphatase and Cdc42. These proteins were chosen for several reasons. First, they are both enzymes, and what is more important is that they both bind charged substrates. Second, they differ in size and topology, thus providing a test for universality of the design strategy. Third, the 3D structures of these proteins are known, which is crucial for setting up the computational energy-optimization procedure. The designed variants of these 2 proteins were characterized by a battery of biophysical methods showing that they are more stable than the respective WT proteins and, by using biochemical methods, showing that they do retain their enzymatic activities at levels comparable with the WT proteins.
Two model proteins were chosen for the test: human acylphosphatase, AcPh (EC-Number 3.6.1.7), and human cell-division cycle 42 factor, Cdc42 (EC-Number 3.6.5.2). Acylphosphatase is a 98-residue protein that catalyzes hydrolysis of acylphosphates to produce carboxylate and inorganic phosphate (23). This enzyme is proposed to be involved in several metabolic pathways and, in particular, in glycolysis/gluconeogenesis, pyruvate metabolism, and benzoate degradation via CoA ligation (,23). Acylphosphatase activity has been identified in both prokaryotes and eukaryotes (,23). According to SCOP classification, this protein belongs to the structural family of α+β sandwich proteins with antiparallel β-sheet (,Fig. 1). The actual location of the substrate-binding site is not known. It is believed to be located near the sulfate-binding site observed in the 2ACY structure of the bovine protein (,Fig. 1), with residues R23, K24, N41, and K98 located in close proximity. Biochemical and mutagenesis analyses have provided further support for this notion and identified R23 and N41 (,23) as residues that are critically important for the activity of AcPh. This observation is also supported by our results on redesign of AcPh that were not concerned with maintaining activity. It was shown that a variant with 4 substitutions (K24E/E63K/N81K/Q95K), although more stable than the WT (,7), is in fact inactive.
The diagram structures of 2 proteins studied in this paper. The locations of active sites are shown by van der Waals surfaces for the substrates. Positions of the substitutions are shown by gray circles corresponding to the backbone CA atoms. (A) Acylphosphatase (AcPh): comparison of the structures of bovine protein (PDB code 2ACY) (34) with the 3D structures of human protein AcPh-wt (blue) and its designed variant AcPh-des (green). The structures of AcPh-wt and AcPh-des were determined in this work by using multidimentional NMR spectroscopy (see Materials and Methods and SI Appendix for details). (B) Cdc42: structural model of human protein (PDB code 1AN0) as solved by x-ray crystallography in ref. 40.
Cdc42 is a eukaryotic protein and belongs to the family of small GTPases (24). It is a key enzyme that is involved in regulation of numerous vital cellular pathways such as gene expression, cell-cycle progression, and rearrangement of the actin cytoskeleton (,25). It has 198 amino acid residues arranged into a fold common for all Ras GTPases: a 6-stranded β-sheet surrounded by 5 α-helices (,Fig. 1). Eleven residues were identified to be in direct contact with the GTP molecule, of which three (K16, D57, and D118) are ionizable residues.
Distribution of surface charges in these 2 proteins was optimized by the Tanford–Kirkwood surface accessibility genetic algorithm (TKSA-GA), which is based on a GA for optimization of charge–charge interactions in proteins (26), in which the energies of charge–charge interactions are calculated by using TK formalism (,27) that includes SA correction as introduced in ref. ,28. The dependence of the energy of charge–charge interactions on the number of substitutions as identified by TKSA-GA is shown in ,Fig. 2 A and C. The favorable energies of charge–charge interactions increase with the number of substitutions until reaching saturation values. Such saturation behavior has been observed for other proteins and is consistent with 2 facts (7, ,8, ,29): (i) for a given protein topology, there is a finite number of surface sites where charged residues can be introduced, and (ii) addition of a charged residue at a new site has both favorable (with residues of the opposite charge) and unfavorable (with residues of the same charge) interactions.
The dependence of the energy of charge–charge interactions on the number of amino acid substitutions (or the percentage of substitutions relative to the total number of amino acid residues) in AcPh (A) and Cdc42 (C). Each small dot corresponds to a different sequence. The sequences selected for experimental verification are shown in large symbols: AcPh-wt or Cdc42-wt, black circles; AcPh-des, red squares; Cdc42-des1, blue squares; and Cdc42-des2, red triangles. (B and D) The corresponding per residue energies of charge–charge interactions as calculated by TKSA model are given in B (AcPh-wt, black bars; AcPh-des, red bars) and D (Cdc42-wt, black bars; Cdc42-des1, red bars; Cdc42-des2, blue bars). Positive energies represent overall unfavorable interactions of a given residue with all other ionizable residues in the proteins, whereas negative values of ΔGqq reflect overall favorable interactions.
For experimental tests, we selected sequences that have the largest increase in the energy of charge–charge interactions but have the number of substitutions at 5–6% of the total number of amino acid residues. These sequences are shown in Fig. 2 B and D. For AcPh, the substitutions were made at 5 positions, 3 of which led to charge reversal (H60E, E63K, and K72E) and 2 of which introduced new charges (Q50K and N81K). For Cdc42, we fully characterized 2 variants with 7 (Cdc42-des1) and 8 (Cdc42-des2) substitutions. Cdc42-des1 has 4 charge-reversal substitutions (E95K, K107E, D121K, and K131E) and 3 new charges (Q74E, N167K, and V189K). Cdc42-des2 has an additional charge-reversal substitution (E178K). Calculations show that the designed variants should have more favorable energies of charge–charge interactions, in particular because of a decrease in the number of amino acid residues with overall unfavorable energies of charge–charge interactions, ΔGqq (see Fig. 2 B and D for TKSA calculations and Fig. S1 for calculations using the MCCE software package) (30).
These designed proteins were cloned, expressed, and purified (see Material and Methods), and their physico–chemical properties were examined and compared with the properties of the corresponding WT proteins. First, we compared the oligomerization properties of the WT and designed proteins. Substitutions on the protein surface can influence the oligomerization state in solution by increasing the size of oligomer, by creating complementary surfaces, or by inducing domain swapping (31, ,32). Any of these mechanisms would increase the association constant, thereby making the oligomerization stronger. Such changes in the quaternary structure can lead to changes in stability. However, this increase/decrease in stability will not be directly relevant to the main hypothesis of this work, which is the demonstration that protein stability can be enhanced by optimization of charge–charge interactions on the protein surface. Both WT AcPh and Cdc42 are monomeric in solution as measured by equilibrium analytical ultracentrifugation. Importantly, all of the designed proteins were also found to be monomeric in solution (Fig. S2). This finding suggests that if the designed AcPh and Cdc42 are more stable, the increase in stability would not be due to changes in the oligomerization state of the proteins.
Changes in the structure of proteins upon amino acid substitutions can be another reason for changes in stability and thus create additional challenges for testing the hypothesis. Such substitutions can sometimes lead to dramatic structural changes (33). These changes in the structure of the native-state protein could, in turn, alter stability in a way that could compromise experimental proofs of enhancing protein stability solely by optimization of charge–charge interactions on the protein surface. To address this issue, structures of proteins were examined first at the low level of resolution by using circular-dichroism spectroscopy. Comparison showed no changes in the far-UV CD spectra between WT and designed proteins, suggesting that there are no large structural changes in these proteins with substitutions (Fig. S3). To further validate this conclusion, at higher structural resolution, we solved solution structures of AcPh-wt and AcPh-des by using multidimensional NMR spectroscopy (see SI Appendix for procedures on structure determination and detailed statistics of the final structures). Fig. 1A shows the overlay of the averaged structures of these 2 proteins, AcPh-wt and AcPh-des, together with the reported X-ray structure of bovine AcPh (34) that was used for the TKSA-GA optimization protocol. The rms deviation of the backbone atoms between bovine, ensemble-averaged AcPh-wt, and AcPh-des is <1.1 Å. This observation further supports the notion that, structurally, the designed protein is very similar to the WT AcPh.
Because there are no changes in the oligomerization state of the proteins and there are no dramatic changes in their 3D structure, all changes in stability are probably directly related to differences in charge–charge interactions in the WT and designed proteins.
Stability of the AcPh-wt and AcPh-des proteins in solution was estimated from temperature-induced reversible unfolding by using 2 different methods: differential scanning calorimetry (DSC) and monitoring the changes in ellipticity at 222 nm by using CD spectroscopy (Fig. 3 A and B). Temperature-induced unfolding was fully reversible, and the transitions were analyzed by using a 2-state unfolding model (see SI Appendix). Importantly, both methods (DSC and CD) produced similar thermodynamic parameters, suggesting that there are no changes in the unfolding mechanism upon substitutions in AcPh-des (Table 1). Notably, AcPh-des has an unfolding temperature 10 °C higher than AcPh-wt, which translates into an increase in Gibbs energy of 9 ± 1 kJ/mol. This increase in stability is rather substantial (,4), suggesting that indeed, the TKSA-GA protein-stabilization approach can identify protein sequences that will produce more stable proteins.
Biophysical characterization of the designed and WT acylphosphatase (AcPh) and Cdcd42. (A) Comparison of the experimental partial molar heat capacity profiles of AcPh-wt (white circles) and AcPh-des (red squares). Solid lines are the results of the fit according to a 2-state unfolding model. The results of the fit are given in Table 1. (B) Comparison of the experimental temperature-induced unfolding profiles of AcPh-wt (black thin line) and AcPh-des (red thin line) as monitored by the changes in ellipticity at 222 nm. Thick solid lines are the results of the fit according to a 2-state unfolding model. The results of the fit are given in Table 1. (C) Comparison of the experimental temperature induced unfolding profiles of Cdc42-wt (thin black line), Cdc42-des1 (thin blue line), and Cdc42-des2 (thin red line) as monitored by the changes in ellipticity at 222 nm. (D) Dependence of specific activity of Cdc42 variants (Cdc42-wt, black circles; Cdc42-des1, blue squares; Cdc42-des2, red triangles) on temperature. Lines are drawn to guide the eye. (E) Resistance of Cdc42 variants to aggregation after exposure to high temperature. The fraction of soluble monomer remained in solution after exposure to elevated temperatures was measured by using size-exclusion chromatography as described in Materials and Methods. Shown are Cdc42-wt (black circles), Cdc42-des1 (blue squares), and Cdc42-des2 (red triangles), and lines are drawn to guide the eye.
Comparison of properties of the wild-type and designed AcPh proteins
How does this increase in stability affect the enzymatic activity of AcPh? The activity of AcPh proteins was compared by using synthetic substrate benzoylphosphate (35, ,36). The results of activity measurements at different temperatures are given in ,Table 1 (actual kinetic traces are shown in Fig. S4). It is remarkable that not only does the AcPh-des protein remain active, it also has very similar kinetic parameters as the WT. At all temperatures, the catalytic constant Kcat is the same for both enzymes (AcPh-wt and AcPh-des). Although there is some decrease in Michaelis-Menten constant Km at room temperature, the difference becomes very small at 55 °C (Table 1). Thus, for AcPh, the amino acid substitutions that were introduced by design via optimization of charge–charge interactions on the protein surface produced an enzyme that is not only 10 °C more stable but also equally catalytically active.
Temperature-induced unfolding of Cdc42 is irreversible because of aggregation in the unfolded state of the protein (see below). To compare the transition profiles, temperature-induced unfolding was monitored by using changes in ellipticity at 222 nm at different protein concentrations and solvent conditions (2 M urea) (Fig. 3C and Table 2). Melting profiles for Cdc42-wt, Cdc42-des1, and Cdc42-des2, when compared at identical concentrations (,Fig. 3C and Table 2), show that the designed variants have ≈8 °C and ≈10 °C higher midpoints of transition, respectively, than the Cdc42-wt protein.
Comparison of properties of the wild-type and designed Cdc42 proteins
The enzymatic activity of all 3 proteins (Cdc42-wt, Cdc42-des1, and Cdc42-des2) at room temperature remains unchanged (Table 2). The temperature dependence of the specific activity for these 3 proteins is compared in ,Fig. 3D. It is clear that both designed Cdc42 proteins remain active at temperatures higher than the WT protein. Estimates show that the difference in midpoint of temperature inactivation for the Cdc42-des1 and Cdc42-des2 proteins is ≈10 ± 2 °C, which is consistent with the estimates based on the CD-melting profiles (Fig. 3D and Table 2). The changes in CD signal and loss-of-activity are due to the protein unfolding followed by irreversible protein aggregation. To test whether substitutions in the designed Cdc42 variants led to offset in the temperature of aggregation, the temperature dependence of aggregation was determined by using size-exclusion chromatography to quantify the fraction of protein that remains as monomer in solution. ,Fig. 3E shows the percentage of the monomeric Cdc42 remaining in solution after exposure to different temperatures. The midpoint temperatures of aggregation for both designed Cdc42 variants are ≈9 ± 2 °C higher than that of the WT protein, which is also consistent with the data from CD-monitored unfolding and temperature dependence of the specific activity (Fig. 3 C and D). These results suggest that the offset in the aggregation temperature in designed Cdc42 variants is due to the higher stability of their native states that start to unfold at higher temperatures, and thus the onset of aggregation is at higher temperatures as well.
Is it fortuitous that the selected sequences are more stable? Is it that the reference (i.e., WT) proteins chosen were not very stable to begin with? Both AcPh-wt and Cdc42-wt have midpoint-transition temperatures >55 °C, which suggests that they are rather stable proteins. Are the introduced substitutions more frequent in the corresponding sequence positions of the homologous series of AcPh and Cdc42? If so, can they be identified by using a consensus-design approach that uses evolutionary conservation as a means to engineer more stable proteins? To answer these questions, we analyzed the normalized probabilities of the residues at the substitution sites. The normalized probability of a residue at a given position in sequence (Pi) indicates how many times over what would be expected by random chance does a given residue type occur at that position (see Materials and Methods for details). Interestingly, of 13 substitutions in the 2 proteins, 12 are substitutions from more frequent (Pi > 1) to less frequent (Pi < 1) amino acid residues. The ratio Pi,wt>/Pi,des, ranges from 2 to 75 with an average value of 13 (Table S1). The only exception is the H60E substitution in AcPh, for which WT histidine is 1.6 times less frequent than glutamic acid. Overall, for all 12 substitution positions, the WT residues have normalized Pi, > 1. Moreover, in many cases, WT-residues have the maximal or second maximal Pi, and yet the substitutions in these residues lead to increase of protein stability without changes in enzymatic activity. One possible explanation of these observations is that the interactions with other proteins and ligands in the cell dictate the preference for the residues at these positions, i.e., activity-for-stability tradeoff (37, ,38). Although cellular partners interacting with AcPh are not known, proteins interacting with Cdc42 are well characterized, and disruption of these interactions frequently leads to lethal phenotype. For example, a set of alanine-scanning mutations has been generated in the yeast homolog of Cdc42, and the effect of these mutations on the growth of yeast has been examined (,39). A number of mutations were found to be detrimental. However, for the mutations in positions E95, D121, K131, and E178 (i.e., site of substitutions in our Cdc42-des1 and Cdc42-des2), the growth of haploid cells on YPD plates had a WT phenotype (,39). Further support that the site of the substitutions in Cdc42-des1 and Cdc42-des2 does not affect the function of this protein was obtained from the analysis of the Cdc42 interactions with Cdc42 GTPase-activating protein Cdc42GAP. Cdc42GAP binds Cdc42 and activates GTPase activity of Cdc42 (,40). We analyzed the enzymatic activity (e.g., GTPase activity) enhancement of Cdc42 in the presence of Cdc42GAP protein. For all 3 proteins (Cdc42-wt, Cdc42-des1, and Cdc42-des2), it was observed that the addition of Cdc42GAP leads to an enhancement of GTPase activity in a concentration-dependent manner (Fig. S5). Moreover, the enhancement in activity among the WT and designed Cdc42 variants is practically indistinguishable.
To further characterize the interactions of the Cdc42 variants with Cdc42GAP, we performed 2 types of experiments. First, we measured the GTPase activity of the Cdc42 variants at a fixed concentration of GTP (100 μM) and different concentrations of Cdc42GAP. The results of these measurements are shown in Fig. 4 and suggest that all 3 Cdc42 proteins have similar enhancement of activity by Cdc42GAP. Analysis shows that the Kd,app is similar for Cdc42-wt, Cdc42-des1, and Cdc42-des2 and is on the order of 0.3 μM (Table 2). Interestingly, the same dissociation constant of Cdc42-wt and Cdc42GAP was measured by Cerione et al. (41), who used fluorescence polarization spectroscopy to monitor Cdc42 to Cdc42GAP binding in the presence of GTP. Second, we used isothermal titration calorimetry to measure Cdc42 binding to Cdc42GAP in the presence of GDP (Fig. S6). These experiments were performed in the presence of excess GDP, and thermodynamic analysis of the binding isotherms shows that under these conditions, the apparent dissociation constant for Cdc42 binding to Cdc42GAP is on the order of 10 ± 3 μM and is the same for all 3 Cdc42 variants. This estimate for Cdc42 binding to Cdc42GAP in the presence of GDP agrees well with the dissociation constant of 2.8–2.9 μM (24, ,42) and is in reasonable agreement with the 50 μM value reported by Hoffman et al. (41). These experiments establish the feasibility of engineering a protein with enhanced stability without perturbing its enzymatic or signaling functions.
Stimulation of the GTPase activity of the Cdc42 proteins by the increasing concentrations of Cdc42GAP. Experiments were performed at the initial GTP concentration of 100 μM. Shown are Cdc42-wt (black circles), Cdc42-des1 (blue squares), and Cdc42-des2 (red triangles). Concentration of all 3 Cdc42 proteins was kept constant at 0.5 μM. Data were fitted to a binding equation
where y is the fractional saturation, LT is the total Cdc42GAP concentration, MT is the total Cdc42 concentration, and Kd,app is the apparent dissociation constant of Cdc42 to Cdc42GAP. Dashed lines are the results of the fit for each individual protein, and the solid line is the fit of all datapoints to a single Kd,app. See the results of the fit in Table 2.
The results presented here clearly support the idea that the stability of many proteins and enzymes is not fully optimized from the evolutional point-of-view and, thus, the stability of the enzymes can be increased via optimization of surface charge–charge interactions without perturbing the enzymatic activity. As such, this approach is a viable strategy that can be used on its own but also in combination with other design strategies that rationally optimize other types of interactions, predominantly in the protein core (2–,5).
Optimization of energies of charge–charge interactions on the protein surface, as calculated by the TKSA model (7, ,14, ,43) by using GA, was identical to the approach used by us (,7, ,26) (see SI Appendix for details). X-ray structures of human Cdc42 and bovine acylphosphatase (PDB entries 1AN0 and 2ACY, respectively) were used as starting templates in the design (34, ,40). To assess errors in the calculated energies of charge–charge interactions and to at least partially account for the dynamics of the side chains in solution, ensembles of 11 structures were generated from each starting template by using Modeler version 7.7 (,44). Average surface accessibilities, distances between the side chains, and interaction potentials were calculated from these structural ensembles and used as input parameters during GA optimizations. New sites for introducing ionizable residues were picked from those that are >50% solvent exposed. Side chains with <50% solvent accessibility, side chains forming multiple hydrogen bonds, and side chains located within 10 Å of the active sites and/or ligand-binding sites were excluded from optimization. For AcPh, the sites included in the optimization were E2, D4, K31, K32, D43, Q44, Q50, H60, E63, E66, K68, K72, H74, D76, R77, S79, H81, N82, K84, K88, D90, D93, and Q95. For Cdc42, the optimizations sites were Q2, N26, K27, N39, M45, E49, E62, D63, K66, Q74, E91, K94, E95, H103, K107, D121, E127, K128, K131, N132, Q134, K135, E140, K144, R147, D148, K150, K166, N167, E178, E181, K183, K184, and V189. During the TKSA-GA run, each and every site that was included in optimization was allowed to a have positive, negative, or neutral charge, which is independent of the charges on the other residues. The energies of these different combinations of charges were evaluated, and those that have energies higher than a certain predetermined value were discarded.
All proteins used in this work were cloned into pGia expression vectors, overexpressed in BL21(DE3) pLys E. coli strain, and purified to homogeneity by using column chromatography (7). Detailed description of the cloning, expression, purification, and characterization is provided in SI Appendix.
The sequences homologous to the human Cdc42 were identified by using BLASTP 2.2.17 and nonredundant protein sequences from all major databases (45); 995 sequences were identified this way. For AcPh, sequence alignment was taken from the Pfam database (,46) and contained 682 sequences. The normalized probability of finding a residue type (i) at a substitution site (j) was calculated as:
where ni(j) is the number of residues of type i at site j, Σi ni(j) is the total number of sequences in the alignment, Ni is the number of residues of type i in the sequence alignment, and Σi Ni is the total number of residues in the alignment.
All NMR spectra were obtained at 27 °C on 600- and 800-MHz Bruker spectrometers equipped with cryoprobes. Protein concentrations of 0.4 to 1 mM were used for all experiments, in 20 mM phosphate buffer, 50 mM NaCl, at pH 5.7. Polypeptide backbone assignments were obtained with HNCO, HNCACB, and HN(CO)CACB (Fig. S7). Side-chain assignments were obtained through the analysis of 3D 13C-separated HCCH-TOCSY and COSY spectra. Distance constraints were obtained from a 3D 15N-NOESY spectrum recorded at 1H frequency of 600 MHz and 2 3D 13C-NOESY spectra recorded with a 150-ms mixing time at a 1H frequency of 800 MHz with the 13C-carrier frequency in the aliphatic and the aromatic region, respectively. The 3 datasets were peak-picked and automatically assigned by using the CANDID macro for CYANA (47). Dihedral-angle constraints were obtained by using the program Talos (,48). The structural statistics are listed in Table S2.
We thank Dr. Cerione (Cornell University, Ithaca, NY) for the Cdc42 clone and Drs. Thomas Spratt and Susan Gilbert for advice on enzyme kinetic data acquisition and analysis. This work was supported by National Science Foundation Grant MCB-0110396 (to G.I.M.).
Author contributions: A.V.G. and G.I.M. designed research; A.V.G., M.M.P., J.L., S.A.M., and C.W. performed research; A.V.G. and M.M.P. analyzed data; and A.V.G., M.M.P., and G.I.M. wrote the paper.
↵2Present address: Department of Biochemistry and Molecular Biology, University of Texas Medical Branch, Galveston, TX 77555.
The authors declare no conflict of interest.
This article is a PNAS Direct Submission.
Data deposition: The atomic coordinates and structure factors have been deposited in the Protein Data Bank, www.pdb.org (PDB ID codes 2K7J and 2K7K).
This article contains supporting information online at www.pnas.org/cgi/content/full/0808220106/DCSupplemental.
