Structural Basis for Autoinhibition and Mutational Activation of Eukaryotic Initiation Factor 2α Protein Kinase GCN2*[boxs]

The GCN2 protein kinase coordinates protein synthesis with levels of amino acid stores by phosphorylating eukaryotic translation initiation factor 2. The autoinhibited form of GCN2 is activated in cells starved of amino acids by binding of uncharged tRNA to a histidyl-tRNA synthetase-like domain. Replacement of Arg-794 with Gly in the PK domain (R794G) activates GCN2 independently of tRNA binding. Crystal structures of the GCN2 protein kinase domain have been determined for wild-type and R794G mutant forms in the apo state and bound to ATP/AMPPNP. These structures reveal that GCN2 autoinhibition results from stabilization of a closed conformation that restricts ATP binding. The R794G mutant shows increased flexibility in the hinge region connecting the N- and C-lobes, resulting from loss of multiple interactions involving Arg794. This conformational change is associated with intradomain movement that enhances ATP binding and hydrolysis. We propose that intramolecular interactions following tRNA binding remodel the hinge region in a manner similar to the mechanism of enzyme activation elicited by the R794G mutation.

Mammalian and yeast cells respond to starvation or stress by down-regulating overall protein synthesis, while increasing translation of specific mRNAs encoding transcription factors responsible for ameliorating starvation or stress conditions. These responses are induced by protein kinases that phosphorylate serine 51 in the ␣-subunit of translation initiation factor 2 (eIF2␣). 1 The eIF2 (␣␤␥) heterotrimer forms a ternary com-plex with GTP and methionyl-tRNA Met , which delivers Met-tRNA i Met to the small ribosomal subunit (1). Thereafter, this 43 S preinitiation complex binds to mRNA and assembles an 80 S initiation complex at the AUG start codon, with hydrolysis of eIF2-bound GTP and release of eIF2-GDP (2). eIF2-GDP is recycled to eIF2-GTP by the guanine nucleotide exchange factor eIF2B. Phosphorylation of eIF2␣ converts eIF2-GDP into an inhibitor of eIF2B, thereby decreasing ternary complex formation and protein synthesis (1).
GCN2 is the most widespread eIF2␣ kinase superfamily member, first identified as an inducer of GCN4, a transcriptional activator of amino acid biosynthetic genes in budding yeast. Four short open reading frames in the GCN4 mRNA leader support ribosomal reinitiation and increased GCN4 translation in response to moderate reductions in eIF2-GTP. Accordingly, eIF2␣ phosphorylation in amino acid-starved cells induces GCN4 with attendant increases in amino acid production (10). Similarly, activation of GCN2 in starved mouse cells, or of pancreatic endoplasmic reticulum kinase in cells experiencing endoplasmic reticulum stress, induces translation of the stress-responsive transcription factor ATF4, while reducing overall rates of translation initiation (5). Activation of GCN2 triggers transcriptional rescue via NF-B signaling in mammalian cells subjected to UV stress (11).
GCN2 is activated in amino acid-starved yeast cells via binding of uncharged tRNA to its histidyl-tRNA synthetase (HisRS)-like domain (12)(13)(14) (Fig. 1a). Starvation for any amino acid activates GCN2 (12,15). The HisRS domain binds various uncharged tRNAs with similar affinities, discriminating only against their aminoacylated forms (14). A ribosomebinding and dimerization domain (RB/DD) at the extreme C terminus is also required for GCN2 activation (16 -19). The RB/DD domain interacts with the HisRS and PK domains in vitro, and there is evidence that the RB/DD-PK interaction impedes GCN2 activation by decreasing the affinity of GCN2 for uncharged tRNAs (14,18). A degenerate kinase-like domain located immediately N-terminal to the PK domain (PK domain, Fig. 1a) and the RWD region at the extreme N terminus of GCN2 serve as additional regulatory domains. RWD supports binding of the GCN1⅐GCN20 complex (20,21), which interacts with translating ribosomes and may facilitate transfer of uncharged tRNA from the ribosomal decoding site to the HisRS domain (22,23).
The isolated PK domain of GCN2 is completely inert in vitro, but, remarkably, activity is rescued by single amino acid substitutions at Arg 794 (R794G) or Phe 842 (F842L). These constitutively activating (GCN2 c ) mutations bypass the tRNA binding requirement for kinase activation in vivo, and we proposed previously that they alter the PK active site in a way that mimics conformational changes induced by interactions with the HisRS or RB/DD domains on tRNA binding (24). The PK domain interacts with an N-terminal segment of the HisRS domain, and mutations in the latter abolished kinase function, without impairing tRNA binding or dimerization of the HisRS domain. Thus, the HisRS and PK domains probably participate in stable intramolecular interactions, with tRNA binding eliciting kinase activation via conformational changes that propagate from the HisRS domain to the PK domain (18). Other GCN2 c mutations have been described, e.g. E803V (25), whose activated phenotype remains dependent on tRNA binding to the HisRS domain (24). It was proposed that that these latter mutations lower the threshold concentration of uncharged tRNA required for kinase activation, allowing high-level GCN2 activity in nonstarved cells containing basal levels of uncharged tRNA.
To examine the molecular mechanisms responsible for GCN2 autoinhibition and activation by uncharged tRNAs, we determined a series of x-ray structures of the dimeric catalytic domain of Saccharomyces cerevisiae GCN2 in apo-and ATPbound forms at 3.0-and 2.75-Å resolution, respectively. These structures reveal partial closure of the active site cleft that restricts ATP binding by restraining the conformation of the hinge region between the N-and C-lobes of the PK domain. In addition, we determined x-ray structures of the R794G mutant PK domain in apo-and AMPPNP-bound forms at 1.95-and 2.0-Å resolution, respectively. These two structures demonstrate that this activating mutation increases the flexibility of the hinge segment and opens a "molecular flap" that increases the inter-lobate space and accessibility of the enzyme active site. Hence, our work reveals a novel nucleotide gating mechanism via conformational modulation of the hinge region that controls kinase activity. We propose a two-step activation mechanism in which tRNA binding to the HisRS domain leads to a comparable structural remodeling of the hinge region of wild-type GCN2 that facilitates ATP binding. Subsequent autophosphorylation of the activation loop is predicted to facilitate an additional realignment of active site residues necessary for substrate phosphorylation.

EXPERIMENTAL PROCEDURES
Protein Preparation and Crystallization-The GCN2 fragment (from S. cerevisiae) encompassing residues 594 -997 with deleted segment 665-767, and the same fragments bearing the D835N or R794G mutations, were expressed in Escherichia coli BL21(DE3*) cells as N-terminal His 6 -Smt3 fusion proteins (26) using a pET26b-derived expression vector (both S-and Se-methionine forms). The methionine prototrophic strain expressing GCN2PK D835N was used for producing selenomethionine-labeled protein via growth condition-dependent inhibition of the methionine biosynthesis pathway (27). A 1:1000 preculture in LB medium was used to inoculate 300 ml of M9 minimal medium supple-mented with 0.4% glucose, 2 mM MgSO 4 , and 30 g/ml kanamycin. After overnight growth (A 600 ϭ 0.6) the culture was diluted 1:20 into 6 liters of pre-warmed M9 minimal medium as above. A mixture of 100 mg/liter of L-selenomethionine, 100 mg/liter lysine, threonine, and phenylalanine, and 50 mg/liter of leucine, isoleucine, and valine were added after 8 h of culture growth (A 600 ϭ 0.6). The culture was induced with isopropyl 1-thio-␤-D-galactopyranoside (0.3 mM) after 30 min, and cells were grown for 12 h before harvesting.
All proteins were purified to homogeneity by nickel ion affinity chromatography, followed by Ulp1-mediated removal of the tag, subtractive nickel ion, heparin-Sepharose, and gel filtration chromatography. Matrix-assisted laser desorption ionization and electrospray ionization mass spectrometric analyses confirmed that purified GCN2PK WT , GCN2PK D835N , and GCN2PK R794G were neither degraded nor posttranslationally modified. Both wild-type and the D835N mutant showed observed mass ϭ 35,095 Ϯ 1 Da (calculated mass ϭ 35,094.3Da). The R794G mutant showed observed mass ϭ 34,995 Ϯ 1 Da (calculated mass ϭ 34,993.5 Da).
Se-GCN2PK D835N crystals were obtained via sitting drop vapor diffusion against a reservoir containing 11% polyethylene glycol 3350, 50 mM CAPSO, pH 9.4, at 21°C, using a protein concentration of 17.5 mg/ml (stored in 150 mM NaCl, 20 mM Tris-HCl, pH 7.5, and 0.2 mM Tri(2-carboxyethyl)phosphine hydrochloride), after cross-seeding with S-GCN2PK D835N microcrystals obtained under identical conditions. Rectangular bars, diffracting to 2.6-Å resolution, grow in the orthorhombic space group P2 1 2 1 2 1 (a ϭ 79.6 Å, b ϭ 154.1 Å, c ϭ 157.4 Å) with three homodimers/asymmetric unit. Crystals of GCN2PK WT were obtained under similar conditions, after cross-seeding with GCN2PK D835N microcrystals following a reservoir buffer substitution of 100 mM CHES, pH 9.0. Crystal cryoprotection was achieved by adding glycerol to a final concentration of 25% (v/v). Crystals of the GCN2PK WT ⅐Mg 2 ATP complex were obtained by co-crystallization under conditions similar to those used for GCN2PK WT , following addition of 2.5 mM ATP and 5 mM MgCl 2 to the crystallization buffer. GCN2PK WT ⅐Mg 2 ATP crystals were cryoprotected by addition of 2-methyl-2,4-pentanediol to a final concentration of 30% (v/v). Crystals of GCN2PK R794G grow spontaneously via vapor diffusion against a reservoir containing 9% polyethylene glycol 6000 and 100 mM CHES, pH 9.5. GCN2PK R794G ⅐Mg 2 AMPPNP co-crystals were obtained under conditions similar to those used for GCN2PK R794G , following addition of 2.5 mM AMPPNP and 5 mM MgCl 2 to the crystallization buffer. Apo-GCN2PK R794G and GCN2PK R794G ⅐Mg 2 AMPPNP crystals were cryoprotected by adding glycerol to a final concentration of 25% (v/v). All crystals grew within 7 days, with the exception of the GCN2PK WT ⅐Mg 2 ATP complex, which required 2 months.
Data Collection, Structure Determination, and Refinement-All diffraction data were collected at SGX-CAT (Advanced Photon Source) under standard cryogenic conditions. A single wavelength anomalous dispersion experiment was conducted at an x-ray wavelength near the selenium absorption peak with a single Se-GCN2PK D835N crystal. 48 of 54 possible selenium sites were located. Experimental phases were estimated at 2.6-Å resolution using MLPHARE (28), and improved by density modification and noncrystallographic symmetry averaging using DM (28). Automated electron density map interpretation with MAID (29) yielded an atomic model encompassing an average of 105 residues per molecule. Iterative rounds of model building/refinement, performed with REFMAC (28) and CNX (30), yielded a model each of three homodimers giving a working R factor ϭ 21.2% and free R value ϭ 25.9% (Table I). PROCHECK (28) showed no unfavorable (,) combinations with main chain and side chain stereochemical parameters better than average at this resolution limit (overall G value ϭ 0.2). Subsequent structures of GCN2PK WT , GCN2PK WT ⅐Mg 2 ATP, and GCN2PK R794G were determined via molecular replacement using MOL-REP (28) with a GCN2PK D835N monomer search model and were refined to convergence. The structure of GCN2PK R794G ⅐Mg 2 AMPPNP was determined by calculating difference Fourier syntheses, using the isomorphous structure of the GCN2PK R794G dimer. Residues not included in the final structural models frequently correspond to a few residues (773-779) within the kinase insert, a portion of activation loop (863-882), and the C-terminal ␣J-helix (986 -997). Coordinates of a single polypeptide chain (Chain A) from each structure were used for illustrations.
Kinase Assay-The purified proteins (GCN2PK WT and GCN2PK R794G ) were analyzed for kinase activity as described previously (24) by incubating with 25 M [␥-32 P]ATP (6000 Ci/mmol, Amersham) and 2.5 M recombinant eIF2␣-⌬C purified from E. coli in 20 l of kinase buffer for 20 min at 30°C. Reactions were stopped by adding 4ϫ SDS-PAGE sample buffer and boiling, resolved by SDS-PAGE, stained with Coo-massie Blue, and the dried gel was subjected to quantitative phosphorimaging analysis. By varying the concentration of eIF2␣-⌬C in the assays (data not shown), we verified that saturating amounts of substrate were employed in these assays.
Multiple Sequence Alignments and Homology Modeling-Protein sequences similar to GCN2PK WT (E-value Ͻ10 Ϫ4 ) were identified using -BLAST (31). Multiple sequence alignments were prepared using CLUSTALX (32), and amino acid conservation was calculated with BLOSUM62 (33). The homology model of GCN2 HisRS, residues 1034 -1490, was obtained with MODELLER (34) using the structure of histidyl-tRNA synthetase from Staphylococcus aureus (Protein Data Bank code 1QE0) residues 13-420 as a modeling template (E value ϭ 3 ϫ 10 Ϫ11 , sequence identity 19%, model score ϭ 1.0). Tests with known structures have shown that models with scores from 0.7 to 1.0 have the correct -fold at a 95% confidence level (35).

RESULTS
Crystallization and Structure Determination-Limited proteolysis/mass spectrometry studies of the GCN2 PK domain (residues 593-998) confirmed the presence of a large (102 residue) proteolytically sensitive insert that is poorly conserved among GCN2 homologs (Fig. 1b). An optimal E. coli expression construct, encoding the active kinase domain (residues 594 -997) lacking the insert loop (residues 665-767), was used to produce various proteins for crystallographic study. Three distinct proteins, including wild-type enzyme (GCN2PK WT ) and the two mutant forms containing the D835N inactivating mutation (GCN2PK D835N ) or the R794G activating mutation (GCN2PK R794G ), were purified to homogeneity and crystallized. GCN2PK WT and GCN2PK R794G were assayed for autophosphorylation and eIF2␣ kinase activity (see below). Unmodified native GCN2 PK domains containing the insert region failed to yield crystals under various crystallization conditions. The residue-numbering scheme used in this article corresponds to the native GCN2 PK domain (Fig. 1b).
The structure of Se-Met GCN2PK D835N was determined at 2.6-Å resolution via single wavelength anomalous dispersion and noncrystallographic averaging of three homodimers within the asymmetric unit (final refinement model gave R work ϭ 21.1% and R free ϭ 25.9% with excellent stereochemistry). Thereafter, the refined structure of a GCN2PK D835N monomer was used as a search model for additional structure determinations, including autoinhibited apoenzyme (GCN2PK WT ), a hyperactive mutant form of the apoenzyme GCN2PK R794G , and a GCN2PK WT ⅐Mg 2 ATP complex. Finally, the structure of the GCN2PK R794G ⅐Mg 2 AMPPNP complex was obtained via difference Fourier syntheses using the isomorphous structure of GCN2PK R794G (Table I and "Experimental Procedures").
The larger C-terminal lobe (C-lobe, residues 795-982) containing amino acids implicated in catalysis, activation, and substrate recognition is predominantly ␣-helical (␣D-␣I), and is connected to the N-lobe by the hinge region (residues 790 -794, Figs. 1b and 2a). The conformation of the Asp-Phe-Gly (DFG) motif of apo-GCN2PK WT closely resembles that of the cAPK in its active conformation (PDB code 1ATP). In both structures, the DFG motif projects into the active site, whereas the activation loop extends out into solvent (Fig. 2a, magenta and orange segments, respectively). The 42-residue activation loop in GCN2 (residues 853-894, Fig. 1b) is longer than is typically seen in protein kinases.
Approximately 22 residues at the center of the activation loop (residues 861-882) appear disordered and are invisible in experimental electron density maps (Figs. 1b and 2a). Autophosphorylation sites within the activation segment, Thr 887 and Thr 882 (42), are not phosphorylated in any of the purified proteins used for crystallization ("Experimental Procedures"). A segment between ␣F and ␣H (residues 921-951) exhibits high B-factors, consistent with disorder. The most C-terminal ␣-helix (␣J, residues 985-999), which serves as a linker to the HisRS domain, exhibits minimal non-covalent interactions with the PK domain and is disordered in various crystalline forms of GCN2. The remarkable level of pairwise sequence identity (Ͼ34% among all eukaryotes, Fig. 1b) and the pattern of amino acid differences among GCN2 PK domains from different organisms allow us to conclude that all known GCN2 proteins share essentially the same three-dimensional structure.
GCN2 PK Forms a Symmetric Homodimer-Consistent with previously published studies indicating that GCN2 functions as a homodimer (17)(18)(19), the GCN2 PK domain repeatedly crystallized as a symmetric homodimer (Fig. 2b,c) independent of lattice packing arrangements (Table I). On dimer formation, ϳ2,600 Å 2 of solvent-accessible surface area is buried, which is consistent with a stable homodimer (43). Size exclusion chromatographic studies demonstrated that GCN2 PK is dimeric in solution (data not shown). The dimer interface is composed equally of hydrophobic and polar side chains, and is stabilized by ϳ26 amino acids from each monomer ( Fig. 1b; within residues 594 -830) that participate in 22 hydrogen bonding interactions. All but four of these residues are located in the N-lobe. Given this dimerization interface, the mode of dimerization observed in our crystals almost certainly represents the PK domain dimer found within the dimer of full-length GCN2.
Inactive Conformation of GCN2 PK-The GCN2PK WT apoenzyme structure exhibits several features characteristic of autoinhibited kinase conformations with both a displaced helix ␣C and a closed bi-lobate conformation. Structural superposition of the C-lobes of the apo-and Mg 2 ATP-bound forms of GCN2PK WT documents that ATP binding requires opening of the cleft between the N-and C-lobes. Among known kinase structures, a murine cAPK ternary complex (with bound ATP and inhibitor peptide, PDB code 1ATP) exhibits one of the most closed bi-lobate conformations (44). Apo-GCN2PK WT adopts an even more compact, closed conformation, with the N-lobe ϳ4°S  Residues in additionally allowed region (%)  closer to the C-lobe than observed in murine cAPK (Fig. 3a). In contrast, the inter-lobe angle of Mg 2 ATP-bound GCN2PK WT (ϳ91°) closely resembles that of the ATP-bound cAPK (ϳ92°), with the two N-lobes rotated by ϳ11°with respect to each other. Examination of our structural superposition of the Clobes of apo-and Mg 2 ATP-bound GCN2PK WT shows that in the absence of a conformational change, ATP cannot bind to the apo-conformation of GCN2PK WT (Fig. 3b). We believe that GCN2 autoinhibition can be understood as described next.
Inter-lobe mobility in protein kinases reflects mechanical flexibility of the hinge region, which often contains conformationally less restricted glycines (as in the cAPK hinge). There are no glycines in the hinge of GCN2PK WT (Fig. 1b). The residue equivalent to Gly 126 of cAPK is Arg 794 , which makes a salt bridge with Glu 803 that binds to Arg 847 via a second salt bridge (Fig. 3c). Arg 794 also accepts a backbone hydrogen bond from the amide nitrogen of Ile 843 . Finally, the amide nitrogen of Arg 794 interacts with the carbonyl oxygen of Glu 792 , forming an unusual, strained i 3 iϩ2 CϭO . . . H-N hydrogen bond that clamps Asn 793 in a position that partially blocks ATP entry into the bi-lobal cleft. N-terminal to Glu 792 , C␤ of a conserved Cys 791 participates in a weakly polar C-H . . . interaction with the aromatic ring of conserved Phe 842 (45). Thus, an ensemble of non-covalent interactions rigidifies the hinge region, resulting in strong coupling of the two PK domain lobes. We suggest that these stabilizing interactions preclude ATP entry into the active site, thereby ensuring that the GCN2 PK domain is only active in the context of uncharged tRNA binding. The closed bi-lobate autoinhibited conformation of the GCN2 PK domain differs from those of all other closed, inactive kinase structures (reviewed in Ref. 46).
Structure of the Hyperactive GCN2 PK R794G Mutant Suggests a Model for Kinase Activation-In yeast, two mutations (R794G and F842L) bypass the requirement for tRNA binding to the HisRS domain, yielding constitutively active forms of GCN2 in vivo. Whereas the isolated wild-type PK domain is completely inactive, a double mutant form of the enzyme (R794G,F842L) is Ͼ500 times more active in phosphorylating eIF2␣ in vitro versus wild-type (24). Autophosphorylation experiments performed using GCN2PK R794G and GCN2PK WT prepared for crystallization show that the R794G mutation increases autophosphorylation activity ϳ75-fold (Fig. 4a). GCN2PK R794G is also significantly more active for eIF2␣ phosphorylation. We could not quantiate the extent of activation because GCN2PK WT is completely inert for substrate phosphorylation (Fig. 4a). Hence, GCN2PK R794G provides a useful model for the activated form of an eIF2␣ kinase.
Our structure of apo-GCN2PK R794G demonstrates relief of the closed autoinhibitory conformation when compared with that of GCN2PK WT (see supplemental materials Table I for r.m.s.d. values). As shown in Fig. 4b, both GCN2PK R794G structures (apo-and Mg 2 AMPPNP-bound) closely resemble the relatively open conformation of GCN2PK WT bound to ATP. Mutation of Arg-794 to glycine, the residue type found at that position in cAPK, disrupts multiple interactions that rigidify the hinge of GCN2 (Fig. 3c), giving rise to both short-and long-range structural perturbations of the PK domain. The conformational freedom of a glycine at position 794 eliminates the strained i 3 iϩ2 hydrogen bond described above, allowing Asn 793 to extend away from the ATP binding cleft in GCN2PK R794G (Fig. 4, b-d).
Substantial backbone conformational changes occur for hinge residues Glu 792 -Gly 794 in the GCN2PK R794G structure (Fig. 4, c and d, and supplemental materials Table II). The backbone torsion angles of G794 (, ϭ 105 o ,175 o ) adopt a conformation disallowed for non-glycine residues (47). Asn 793 , flanked by Glu 792 and Gly 794 , extends away from the active site toward the exterior of the PK domain, fully exposing the ATP binding cleft to solvent. A permissive Asn 793 conformation would provide an energetically favorable state for nucleotide binding, without the need for further opening of the interlobate space by domain movement (s). Comparison of the struc- tures of apo-GCN2PK R794G and its AMPPNP-bound form confirms this assertion (r.m.s.d. ϭ 0.3 Å for C␣ atomic pairs, Fig.  4b). The R794G mutation has a compound effect on ATP binding: it removes a restriction to ATP entry by opening the Asn 793 flap and it weakens the network of interactions that rigidify the hinge, thereby increasing nucleotide accessibility to the bilobate space. Alternatively, it is possible that hinge loosening in the R794G mutant form of the enzyme enables more dynamic inter-lobe movement, permitting ATP entry to the active site.
Importantly, the activating effects of the F842L (24) and E803V mutations (25) can also be explained by increased hinge flexibility. Replacement of the aromatic side chain of Phe 842 with that of leucine would weaken the interaction with Cys 791 (45). Combining this change with the R794G mutation would further reduce hinge rigidity, which could explain why a R794G,F842L double mutation produces greater activation of GCN2 than either single mutation (24). The E803V mutation would eliminate the salt bridge between Glu 803 and Arg 794 that also rigidifies the hinge in wild-type GCN2 (Fig. 3c).
The simplest explanation for the fact that the R794G and F842L mutations bypass the requirement for uncharged tRNA binding for kinase activation is that these mutations mimic structural alterations of the wild-type PK domain resulting from tRNA-binding to the HisRS domain. In contrast, the activated phenotype produced by the E803V mutation depends on tRNA binding to the HisRS domain (24), albeit at basal levels of uncharged tRNAs. Thus, it appears that loss of the Glu 803 -Arg 794 salt bridge does not activate kinase function directly, but does reduce the threshold for tRNA binding to the HisRS domain required to fully relieve the autoinhibitory conformation of the hinge, yielding activation under non-starvation conditions. Although residues Asn 793 , Arg 794 , and Glu 803 are not highly conserved in organisms other than yeast, we believe that similar but different constellations of polar/charged residues in these positions would secure the hinge via analogous interactions and provide the same overall mechanism of autoinhibition.
ATP Binding to GCN2 PK-ATP binding to GCN2 PK exhibits both similarities and differences in protein-ligand interatomic interactions as compared with other protein kinases. ATP binds in a deep pocket, making canonical contacts with conserved residues via the adenine base to main chain atoms of the hinge region between the two lobes (Glu 789 -Cys 791 ). As in other kinases, the adenine-ribose moiety of ATP resides in the conserved hydrophobic pocket (lined with residues Leu 605 , Val 613 , Ala 626 , and Phe 842 ) and makes water-mediated hydrogen bonds with the "gatekeeper" residue (48) Met 788 (Fig. 5a). Similar adenine-ribose interactions support AMPPNP recognition by GCN2PK R794G (Fig. 5b). However, considerable differences are seen in triphosphate recognition by GCN2PK WT and GCN2PK R794G versus other PKs. These differences could be attributed to a paucity of interactions with the P-loop and the displaced location of the ␣C-helix.
Orientation of ATP ␤and ␥-phosphates in other protein kinases is achieved by repositioning the P-loop within the Nlobe (49). Conversely, binding of ATP to GCN2PK WT and AMP-PNP to GCN2PK R794G yield similar P-loop conformations that preclude interactions between the P-loop and ␤and ␥-phosphates (Fig. 5). For GCN2PK WT , the C-lobe alone stabilizes these two phosphate groups, both directly and via co-ordination of two Mg ϩ2 ions (Fig. 5a). In comparison, Mg 2 AMPPNP bound to GCN2PK R794G exhibits fewer contacts between the phosphates and the C-lobe (Fig. 5b), possibly because of disorder (static or dynamic) of the bridging metal ion.
Anchoring of ␣and ␤-phosphate groups is essential for kinase activity. A conserved lysine residue (from strand ␤3) typically facilitates their proper orientation (49). In cAPK, conserved Lys 72 from strand ␤3, which is stabilized by Glu 91 from helix ␣C, positions the ␣and ␤-phosphates of ATP in the proper orientation for catalysis (49). In the structures of GCN2PK WT ⅐Mg 2 ATP and GCN2PK R794G ⅐Mg 2 AMPPNP, the equivalent basic residue (Lys 628 ) is separated from Glu 643 (corresponding to Glu 91 in cAPK) by ϳ11 Å. Instead of the classical ion-pair interaction, Glu 643 forms an inter-lobe salt bridge with Arg 834 (the Arg of the HRD motif), indirectly linking two invariant residues found in the N-and C-lobes, respectively (Fig.  5). The correct re-orientation of the ATP triphosphates requires significant conformational changes that involve both the P-loop and ␣C-helix. Next, we propose that these changes are triggered by autophosphorylation of the GCN2 PK activation loop in wild-type GCN2.

Implications for the "Second
Step" of GCN2 Activation-The activating mutations R794G and F842L lead to increased autophosphorylation of the GCN2 activation loop (Fig. 4a), and there is genetic evidence that phosphorylation of both Thr 887 and Thr 882 in this loop is required for full activity even in the presence of these GCN2 c mutations (24). We propose a two-step activation mechanism for GCN2 in which conformational alteration(s) of the hinge region induced by uncharged tRNA binding to the HisRS domain (step 1) would permit productive binding of ATP to the active site and autophosphorylation of the activation loop (step 2). Our structural analysis suggests that autophosphorylation of Thr 887 and Thr 882 leads to maximal GCN2 kinase activity by a mechanism observed throughout the "RD" subclass of kinases, wherein electrostatic neutralization of the phosphoresidue(s) drives conformational reorientation of a cluster of basic residues (49). One basic residue that would interact with phosphorylated Thr 887 or Thr 882 is the invariant RD arginine (Arg 834 ) that precedes the invariant Asp 835 (49). In GCN2, Arg 834 makes a salt bridge with Glu 643 , thereby stabilizing the orientation of the displaced ␣C-helix observed in our structures (Fig. 5, a and b).
By analogy with the structures of activated cAPK and CDK2, the conserved Arg 834 (Arg 165 in cAPK and Arg 126 in CDK2) of GCN2 is almost certainly involved in phosphoresidue recognition, which would release Glu 643 toward the catalytic site via axial rotation of the ␣C-helix to form an ion pair with Lys 628 . Cyclin activation of CDK2, causes a large translational (up to 8.5 Å) and axial rotation (90°) of the so-called PSTAIRE helix, which is required for formation of the canonical Lys-Glu ion pair (50). Examination of GCN2 suggests that a clockwise axial rotation of the ␣C-helix by ϳ90°with a minimal translation (ϳ3 Å) would position Glu 643 near Lys 628 , creating a salt bridge. An invariant Leu 856 residue, near the DFG motif, blocks such a rotation (Fig. 5). Interestingly, an Glu-Arg salt bridge similar to that of the Glu 643 -Arg 834 pair in GCN2 contributes to autoinhibition of c-Src and hematopoietic cell kinase (39 -41).
In CDK2, the ␣C-helix is amphipathic in character, whereas it is hydrophobic in the GCN2 PK. The GCN2 ␣C-helix contrib- Comparing GCN2PK WT at 1000 nM with GCN2PK R794G at 100 nM shows that the autokinase activity of GCN2PK R794G is about 75-fold higher than that of GCN2PK WT . No eIF2␣ kinase activity was detected for GCN2PK WT . Not surprisingly, given the dimeric nature of the GCN2 protein kinase domain, autophosphorylation is more efficient than eIF2␣-⌬C phosphorylation. The use of a truncated recombinant eIF2␣ subunit (eIF2␣-⌬C) instead of full-length multidomain eIF2 as substrate also may have reduced the efficiency of substrate phosphorylation. b, C␣ backbone views of superimposed wild-type and mutant GCN2 PKs in apo-and ATP/AMPPNP-bound forms. c, conformational changes within the hinge between apo-GCN2PK WT (magenta) and apo-GCN2PK R794G (cyan), with residues 792-794 shown as atomic stick figures. N-lobe and C-lobe surfaces are shown (gray). Asn 793 , the molecular flap, partially occludes the ATP binding cleft in wild-type GCN2. d, opening of the Asn 793 flap exposes the ATP binding cleft in apo-GCN2PK R794G , accommodating the superimposed AMPPNP with no steric clash. AMPPNP is shown as an atomic stick figure within the yellow translucent surface, Mg 2ϩ ion as a red sphere, apo-GCN2PK R794G as a solid gray surface with location of Asn 793 indicated in cyan.
utes extensively to PK domain dimerization (Fig. 1b) and is totally buried between the N-lobe and the dimer interface, where it participates in multiple hydrophobic interactions. We suggest that the hydrophobic character of the ␣C-helix permits the necessary axial rotation for repositioning Glu 643 near the catalytic Lys 628 . Hydrophobic interactions stabilizing the inactive conformation could be substituted by alternate hydrophobic interactions that stabilize the active conformation, whereas a rearrangement of the activation loop could relieve steric hindrance by Leu 856 . In summary, we propose that structural rearrangement induced by autophosphorylation, involving breakage of the Glu 643 -Arg 834 salt bridge via axial rotation of the ␣C-helix and formation of a Lys 628 -Glu 643 ion pair, constitutes the switching mechanism in the second step of GCN2 activation.
Hypothetical Structural Model of the GCN2PK 2 HisRS 2 Dimer-Both bacterial HisRS and the GCN2 HisRS occur as functional dimers, and dimerization is required for tRNA binding by GCN2 (18). Hence, we produced a homology model of the dimerized HisRS domain (HisRS 2 ) of GCN2 ("Experimental Procedures"). Two regions within the GCN2 HisRS domain, 1017-1127 and 1315-1383, suffice for self-association (18). The 1017-1127 region corresponds to the extensive dimer interface of E. coli HisRS (51) and constitutes a substantial portion of the dimer interface in our model (Fig. 6, magenta segment on HisRS 2 ). Point mutations that would abolish symmetrical hydrophobic interactions in this interface destroyed self-association of a GCN2 HisRS segment in vitro and inactivated GCN2 function in vivo (18). Although residues 1315-1383 also contribute to dimerization in our model, they are predicted to interact with other residues in the opposing protomer. We believe that the self-interaction of this isolated segment reported previously (18) was not physiologically relevant.
With an experimental structure of GCN2 PK 2 and a homology model of GCN2 HisRS 2 in hand, we built a hypothetical structural model of the PK 2 HisRS 2 dimer to gain insight into the biological regulation of GCN2. A previous publication (18) suggested that HisRS residues 1028 -1120 and PK residues 750 -810 mediate intermolecular interaction between HisRS and PK in vitro (18) (Fig. 6, magenta residues). Manual docking of the symmetric dimers was performed by linear alignment of the individual dimer axes with a common dimer axis and then minimizing the distance between HisRS 2 and PK 2 . HisRS 2 was rotated about the common dimer axis so as to optimize shapecomplementarity and minimize steric clashes in the putative dimer-dimer interface (Fig. 6). The predicted tRNA binding surface in each GCN2 HisRS monomer, as inferred from the bacterial HisRS-tRNA models (51,52), is freely accessible in our PK 2 HisRS 2 model. Thus, tRNA binding would not dissociate the PK 2 and HisRS 2 dimer (Fig. 6c). Importantly, multiple residues in the PK hinge occur in the predicted dimer-dimer interface of the PK 2 HisRS 2 model, suggesting that a conformational change in the HisRS domain elicited by tRNA binding could be transmitted directly to the PK hinge.
Regulation of GCN2 Catalytic Activity-Genetic and biochemical evidence that the R794G mutation bypasses the requirement for tRNA binding for kinase activation, combined with the structural data indicating that it increases hinge flexibility, provides a strong inference that tRNA binding overcomes hinge rigidity as a means of activating GCN2. Residues 750 -810, which encompasses the hinge, interact directly with HisRS in our model of the PK 2 HisRS 2 dimer. We suggest that under starvation conditions, binding of uncharged tRNA to the upper surface of PK 2 HisRS 2 (as depicted in Fig. 6c) causes a structural change in HisRS 2 that propagates to the hinge regions of one or both PK domains. Remodeling of the hinge would switch the PK domain from the closed, autoinhibited state to the relieved, active conformation, in a manner analogous to the effect of the R794G mutation (Fig. 4c). Although this modeling exercise is not a substitute for an experimental structure of PK 2 HisRS 2 , it is notable that residues 1370 -1392 of HisRS occur within or near van der Waals contact distance of the PK hinge.
Interestingly, the PK residues altered in the GCN2 c -E601K and GCN2 c -E821K activated alleles (25) (Fig. 1b) also map to the interface between the PK 2 and HisRS 2 dimers in our model. These two mutations resemble the E803V mutation, which requires tRNA binding for realization of an activated phenotype (24). Thus, they may directly alter the PK-HisRS interface in a way that lowers the tRNA occupancy of the HisRS domain required to fully rearrange the PK hinge. However, there is evidence that PK-HisRS interactions also impede tRNA binding (18), presumably to prevent kinase activation at low levels of uncharged tRNA. Thus, alterations of the dimer-dimer interface could increase the affinity of the HisRS domain for tRNA as an alternative means of achieving kinase activation in non-starved cells.
The E803V mutation disrupts interaction(s) between RB/DD and PK and increases the affinity of GCN2 for uncharged tRNAs (14), indicating a negative effect of PK-RB/DD interaction on tRNA binding. In our model of PK 2 HisRS 2 , the side chain of Glu 803 is exposed to solvent where it could interact with the RB/DD domain. As noted above, Glu 803 also interacts with a key hinge residue, Arg 794 , and could mediate an allos- teric effect of the RB/DD on PK activity. Mutation of the gatekeeper residue, M788V, is unlikely to affect interactions of the PK domain with either of the HisRS or RB/DD domains and probably increases access to the ATP binding cleft instead. In fact, this mutation partially overcomes the requirement for tRNA binding (24), as would be expected if it mimics the effect of the HisRS domain by facilitating ATP binding. Finally, it could be proposed that domain interactions within GCN2 would correct the displaced conformation of ␣C-helix as an alternative mechanism of kinase activation. However, we favor the idea that remodeling the hinge region through tRNA binding would still be required for kinase activation.
Conclusions-The structure of the wild-type GCN2 PK domain presented here provides an explanation for the mechanism of autoinhibition. The closed bi-lobate conformation imposed by a conformationally rigid hinge segment serves as an impediment to ATP binding and catalysis by the wild-type kinase domain. Comparing the structures of the activated R794G mutant and wild-type enzymes provides detailed molecular insights into the mechanism of GCN2 kinase activation via lobe opening and exposure of the ATP binding cleft, resulting from modest perturbations in the hinge. Elucidating the molecular mechanism of kinase activation by determining the structure of a mutant-activated enzyme is a unique aspect of our work. This mechanism also explains the strong activating effects of the F842L mutation, which should increase hinge flexibility, and of the gatekeeper residue mutation, M788V, that should increase ATP access to the active site. Our structure further provides support for a two-step activation process for GCN2 in which remodeling of the hinge, with attendant ATP binding, is followed by autophosphorylation of the activation loop to permit formation of the critical Glu 643 -Lys 628 salt bridge via re-orientation of the ␣C-helix. The PK 2 HisRS 2 dimer model provides a plausible explanation for tRNA-dependent activation of kinase activity, as the hinge region of the PK domain contributes to the predicted PK-HisRS interface. Hence, a conformational change in the HisRS region upon tRNA binding could be transmitted to the HisRS-PK interface and elicit remodeling of the hinge required to overcome autoinhibition. This process would serve as an efficient signaltransduction mechanism for coupling eIF2␣ phosphorylation to amino acid starvation in the cell. FIG. 6. Hypothetical structural model of the GCN2PK 2 HisRS 2 dimer. a, PK 2 (cyan and green monomers) and HisRS 2 (brown and blue) with their dimer axes perpendicular to the page are shown. Regions of interaction between the PK and HisRS domains determined experimentally (residues 780 -810 and 1034 -1120, respectively) (18) were mapped on the respective surfaces and shown in magenta. b, docking of HisRS 2 with the PK 2 along a common dimer axis (dashed line). c, PK 2 HisRS 2 with the predicted tRNA binding region indicated. d, rotated view of PK 2 HisRS 2 .