eIF2B is a decameric guanine nucleotide exchange factor with a γ2ε2 tetrameric core

eIF2B facilitates and controls protein synthesis in eukaryotes by mediating guanine nucleotide exchange on its partner eIF2. We combined mass spectrometry (MS) with chemical cross-linking, surface accessibility measurements and homology modelling to define subunit stoichiometry and interactions within eIF2B and eIF2. Although it is generally accepted that eIF2B is a pentamer of five non-identical subunits (α–ε), here we show that eIF2B is a decamer. MS and cross-linking of eIF2B complexes allows us to propose a model for the subunit arrangements within eIF2B where the subunit assembly occurs through catalytic γ- and ε-subunits, with regulatory subunits arranged in asymmetric trimers associated with the core. Cross-links between eIF2 and eIF2B allow modelling of interactions that contribute to nucleotide exchange and its control by eIF2 phosphorylation. Finally, we identify that GTP binds to eIF2Bγ, prompting us to propose a multi-step mechanism for nucleotide exchange.

P rotein synthesis in eukaryotes is a highly regulated process, particularly at the initiation phase, which requires the coordinated interplay of at least ten initiation factors (eIFs). Translation initiation begins with GTP-bound translation initiation factor 2 (eIF2) forming a ternary complex with methionyl initiator tRNA (Met-tRNA i ) and delivering it to the P-site on the small 40S ribosomal subunit 1 . In association with other eIFs, this pre-initiation complex binds mRNA 5 0 -untranslated regions near the mRNA cap and scans downstream until it encounters an initiation AUG codon. Hydrolysis of eIF2-bound GTP and Pi release is triggered by correct basepairing of the Met-tRNA i anticodon with the AUG codon of mRNA. This results in the dissociation of eIF2-GDP and other associated eIFs to permit 60S joining and transition to the elongation phase 2 . eIF2-GDP is released from the ribosome in complex with eIF5 (ref. 3), preventing the release of eIF2-bound GDP 4 until eIF2B displaces eIF5 (ref. 5). eIF2B then acts as a guanine nucleotide exchange factor (GEF) catalyzing exchange of GDP for GTP, reactivating eIF2 and allowing subsequent rounds of translation initiation.
eIF2 is a well-conserved heterotrimer comprising a-, band g-subunits, encoded by EIFS1-S3 in humans and SUI2, SUI3 and GCD11 in yeast. eIF2g contains the GTP-binding site and together with band a-subunits binds Met-tRNA i 6,7 , whereas eIF2a is a major point of protein synthesis control. Alterations in protein synthesis, in response to diverse stresses such as viral infection or starvation, are mediated by phosphorylation of the asubunit at Ser-51 (or Ser52 according to SUI2 gene sequence; however, historically phosphorylation of eIF2a is referred to as Ser51, not taking into account the first methionine, which is posttranslationally removed) by eIF2a kinases 8 . Phosphorylation converts eIF2 from a substrate for its GEF eIF2B to a competitive inhibitor 9,10 and reduces the rate of translation initiation. It also activates translation of stress-response genes and is called the integrated stress response, as it provides a common node for multiple stresses. Diminished eIF2 phosphorylation response causes diseases in mammals ranging from metabolic disorders, to altered sensitivity to viral infection or altered brain functions, including enhanced long-term memory 1 . Clinical interest in these proteins has been raised because missense mutations within eIF2 and eIF2B subunits have been linked to distinct inherited neurodegenerative disorders 11,12 .
Detailed structural information for eIF2 is provided by X-ray crystallography of the archaeal homologue (aIF2) from Sulfolobus solfataricus 13 . In this structure, there is no contact between aand b-subunits, which both bind the g-subunit independently 13 . Less is known however about eukaryotic eIF2, the a-subunit being the best characterized [14][15][16] . An NMR solution structure of human eIF2a revealed three domains 14 , while the highly conserved loop containing Ser51 (the phosphorylation site) is well-resolved in the X-ray structure of the amino terminal region (2-175, domains 1 and 2) of yeast eIF2a 16 . In contrast to eIF2, the only atomic resolution structures of eIF2B subunits are of human eIF2Ba 17 , comprising an N-terminal a-helical domain (NaH) and the carboxy terminal domain with a Rossmann-like fold (RLF) and the C-terminal catalytic domain of the e-subunit (e-cat) 18 . No high-resolution structures are available for the remaining eIF2B subunits, for interactions between eIF2B subunits or between eIF2B and its substrate eIF2.
Considerable insight into the structure and function of eIF2B has been provided by genetic and biochemical methods, mostly applied to the yeast factor. Studies on the molecular basis for control of eIF2B by eIF2 phosphorylation have shown that three of the five eIF2B subunits (a, b and d) comprise a regulatory subcomplex that can sense and respond to this important regulatory event by binding to eIF2a 9 . This binding is enhanced by eIF2a phosphorylation 19 and abrogates the catalytic GEF activity of eIF2B. Single mutations in these subunits overcome regulation of eIF2B by eIF2 phosphorylation 20 , suggesting a broad interface between eIF2aP and eIF2B. Genetic co-depletion of b/Gcd7 and d/Gcd2 allowed the formation of a subcomplex containing the remaining three subunits but weakened its interaction with eIF2 in vivo 21 . GEF function is accomplished by the gand e-subunits (encoded by GCD1 and GCD6 in yeast) that form a subcomplex when cooverexpressed 19,22 . As indicated above, the C terminus of eIF2Be bears the catalytic domain 23 but it remains unclear how the gsubunit contributes to GEF activity. Direct interactions between e/GCD6 and eIF2b 24 , and e/GCD6 and eIF2g were determined by mutagenesis and pull-down assays 25,26 . Valuable insight into the interactions between catalytic eIF2Be and g-subunits has recently been reported by combining homology modelling, protein deletions and mutations. Combination of these techniques allowed identification of domain boundaries in catalytic subunits and their individual contributions to eIF2B complex formation in yeast 27 and humans 28 . In yeast 27 , the catalytic subunits of eIF2B (g and e) share sequence similarity and domain structure with each other and with a family of phospho-hexose sugar-nucleotide pyrophosphorylases for which X-ray structures are available (ADP-glucose pyrophosphorylase (AGP) tetramer-(PDB 1YP2) and GlmU trimer (PDB 2OI7)). These structures exhibit different oligomeric states and relative arrangements of the pyrophosphorylase-like (PL) and left-handed b-helix (LbH) domains. In yeast, PL and LbH domains were shown to contribute independently to the interactions between g and e of eIF2B, favouring half of the AGP-tetramer arrangement over GlmU-like interactions 27 .
The dynamic nature of the eIF2:eIF2B interactions have so far resisted X-ray crystallography and require alternative methods. Mass spectrometry (MS) allows the study of protein complexes from solutions in which the native state is preserved, proving informative for determining subunit architectures and protein interactions 29,30 . Here we apply nano-electrospray MS and proteomics, coupled with chemical cross-linking, surface accessibility measurements and homology modelling to define subunit interactions within eIF2B, eIF2, as well as eIF2B and eIF2. We show that eIF2B exists as a decamer with its catalytic core g 2 e 2 responsible for GTP binding and exchange. Our findings prompt us to propose a model for the subunit arrangement within the eIF2B decamer and provide new insights into its function as a GEF for eIF2.

Results
MS reveals an unexpected stoichiometry of eIF2B subunits. Mass spectra of eIF2 purified from a yeast strain co-overexpressing all three subunits shows that the dominant species has a molecular mass of 125 kDa, corresponding closely to the calculated mass of the a/b/g eIF2 trimer (Fig. 1a). Peak splitting, resulting from a mass difference of 450 Da, is assigned to GDP binding (443 Da), present during purification and binding to the g-subunit ( Fig. 1a and Supplementary Table 1). The a/g dimer (93.8 and 94.2 kDa) and the His-tagged g-subunit (58.9 and 59.3 kDa), both with and without GDP, were also observed at low intensity.
The mass spectrum of FLAG-tagged eIF2B (similarly obtained by purification from a yeast strain engineered to co-overexpress all five subunits) shows a species of higher mass than expected for the eIF2B pentamer (298.0 kDa) (Fig. 1b). The measured mass (B600 kDa) corresponds to twice the anticipated mass of an eIF2B pentamer (596.0 kDa). No charge states were observed for a pentameric form, indicating that eIF2B exists exclusively as a decamer. Under high-collision energy conditions, one or two a-subunits were expelled from the eIF2B decamer ( Fig. 1b and Supplementary Fig. 1). eIF2Ba is the only non-essential subunit of eIF2B 22 and its loss during purifications of native protein complexes, requiring sequential column chromatography steps, has been observed previously 31 . Taken together with previous data, the MS analysis is consistent with the idea that a-subunits are located on the periphery of the eIF2B complex rather than within its core.
Complex assembly and GTP binding involve eIF2Bc and e. To address the question of which eIF2B subunits are involved in complex assembly, we purified FLAG-tagged eIF2B lacking the asubunits. Mass spectra showed that this complex is an octamer with a mass of 532.2 kDa (Fig. 2a). Tandem MS (MS/MS; 57 þ charge state) of the complex showed that b-subunits dissociate readily ( Supplementary Fig. 2), indicating their probable peripheral location. In-solution disruption experiments were carried out with organic solvents to disrupt hydrophobic interfaces 29,32 . Addition of 10% acetonitrile (ACN) to the complex-containing solution resulted in the formation of a tetramer of gand e-subunits (305.2 kDa) (Fig. 2b), suggesting that gand e-subunits form a tetrameric core of the complex, with band d-subunits released from the octamer following disruption of hydrophobic interfaces.
To confirm the proposed tetrameric core, we purified the eIF2Bge complex applying the same FLAG-tag strategy (using FLAG-tagged g-subunit) 33 . We found that the ge-core is present as dimeric and tetrameric complexes (Fig. 2c), with the dimeric species being predominant. This implies that band d-subunits stabilize the g 2 e 2 -core in the intact eIF2B complex. Interestingly, the masses of both complexes (ge and g 2 e 2 ) were significantly higher than predicted by summing the masses recorded for eand g-subunits present at lower m/z (Fig. 2d). The mass differences between theoretical and observed masses (832 and 2,830 Da) suggest binding of one or more metabolites to either the eor g-subunits.
Prior analysis has indicated that GTP can bind to eIF2B 34 , and as pyrophosphorylase enzymes including AGP bind specific sugars and nucleotides 27 , we hypothesized that either the eor g-subunit PL domains might bind GTP. As GTP binding by eIF2B could contribute to its nucleotide exchange activity, we therefore investigated GTP binding to the g 2 e 2 -core complex or to FLAG-tag purified e-subunit alone by addition of 6-Thio-GTP, a photo-reactive base analogue of GTP. The low m/z region of the spectrum shows the e-subunit, FLAG-tagged e-(purified separately) and g-subunits. Importantly, two populations of gsubunit were observed, one with an additional mass of 637 Da, corresponding to binding of one GTP. After incubation with 6-Thio-GTP and ultraviolet irradiation, the intensity of the GTP-bound population of the g-subunit increased, while for both e-subunits (with and without FLAG-tag) no nucleotide binding could be observed (Fig. 2d). These data provide strong evidence that eIF2Bg, but not eIF2Be, can bind GTP.
Transient interactions between eIF2 and eIF2B. We applied two different strategies to form eIF2-eIF2B complexes and transfer them intact to the gas phase. First, we used a purification procedure of eIF2B with a FLAG-tagged g-subunit 33 under low ionic strength conditions to retain interactions with eIF2. All subunits were observed on SDS-polyacrylamide gel electrophoresis (SDS-PAGE) gels ( Supplementary Fig. 3a, insert) with eIF2 subunits being sub-stoichiometric to eIF2B subunits as judged by the intensity of stained gel bands. The identity of all proteins was confirmed by nanoflow liquid chromatography-coupled tandem mass spectrometry (nano LC-MS/MS) of tryptic peptides. Mass spectra of the eIF2-eIF2B complex purified at low ionic strength showed charge states for the eIF2B decamer (11,500-13,000 m/z) and the eIF2 trimer (5,000-6,000 m/z) ( Supplementary Fig. 3a). Again, no charge states were observed for a pentameric form of eIF2B. Other species corresponding to the eIF2 a/g dimer, eIF2g, eIF2Bg and eIF2Bge were also present in the spectra. The eIF2-eIF2B complex was retained during purification but the absence of an intact eIF2:eIF2B complex in mass spectra confirms that interactions are transient.
Attempting to stabilize these transient interactions in eIF2:eIF2B, we purified g-His tagged eIF2 phosphorylated in vitro by human PKR, as described previously 35 Fig. 3b insert), and is consistent with enhanced binding of eIF2 to eIF2B induced by eIF2a phosphorylation. Mass spectra of this phosphorylated eIF2-eIF2B complex however revealed the eIF2 and the eIF2B decamer as separate species, very similar to the low ionic strength preparation ( Supplementary Fig. 3). The fact that we are able to purify an eIF2-eIF2B complex, via a FLAG-tagged g-subunit of eIF2B, and to enhance eIF2 binding after eIF2a phosphorylation confirms their interaction during purification. The fact that the complex also readily dissociates before MS, suggests that the interactions between these two protein complexes are of a transient nature.
Subunit stoichiometry of eIF2-eIF2B complexes. As we have established individual eIF2 and eIF2B stoichiometries but could not detect the intact eIF2-eIF2B complex we used label-free quantitative proteomics to determine absolute quantities of protein subunits and thus stoichiometries of interacting eIF2 and eIF2B proteins. The eIF2-eIF2B complex was purified after phosphorylation of eIF2 as described above and proteins were digested with trypsin and subjected to nanoLC-MS/MS. Abundances of the various subunits in the complex were determined by intensity-based absolute quantification (iBAQ) 36 , which takes into account the sum of peak intensities of all the peptides ascribed to a specific protein divided by the number of theoretical peptides. Results of the quantification (Supplementary Fig. 4) show that aand b-subunits of eIF2 are close to stoichiometric with g showing higher abundance likely to be due to the presence of His-tag used for eIF2 purification (a:b:g corresponding to 1:1.08:1.67). For eIF2B subunits quantified by iBAQ, the abundance of a, e and d (1.12:0.76:0.95) correlates well with the FLAG-tagged subunit g (set to 1), whereas the b-subunit ratio was slightly higher than anticipated (1.4-fold over g).
On average, eIF2B:eIF2 subunits were 4:1 in line with the proposal that eIF2 is sub-stoichiometric, due to its transient interactions with eIF2B.
Interactions between protein subunits within the eIF2 and eIF2B complexes, as well as in the eIF2:eIF2B complex, were assessed by chemical cross-linking using deuterated (d4) and non-deuterated (d0) (BS3) cross-linker (Methods). Different complexes were cross-linked: the eIF2:eIF2B complex, the four-subunit eIF2B complex (lacking the a-subunit), the eIF2Bge core complex, and eIF2:eIF2Be and eIF2:eIF2e-cat subcomplexes. In total, we identified 167 unique cross-links (Supplementary Tables 2-5) within all (sub-) complexes. These included 36 inter subunit cross-links, of which 11 were assigned between eIF2B and eIF2. To place our cross-links in a structural context, we used homology models (Supplementary Table 6) generated by MODELLER 37 and SWISS-MODEL 38 web servers for yeast eIF2 and eIF2B subunits (Figs 3a and 4a-e). Using intraprotein cross-links (Supplementary Tables 2 and 4) assigned to individual subunits, we validated our cross-linking strategy, finding a good agreement with the folds of the homologous structures generated and the distances between intra-cross-linked residues measured 39 (Figs 3a and 4a-e).
The available crystal structure of the yeast eIF2a subunit is incomplete (PDB 1Q46, Fig. 3d) 16 and we therefore generated a new homology model for this subunit. Both eIF2 aand bsubunits were modelled using the crystal structure of archaeal aIF2 from S. solfataricus (PDB 3CW2) as a template. eIF2g was modelled using the crystal structure of eIF2g from Pyrococcus abyssi (PDB 1KK1A). Homology models for all three yeast eIF2 subunits were then aligned with the crystal structure of S. solfataricus (PDB 3CW2) (Fig. 3a). Modelled structures of a-(residues 7-267) and g-(residues 96-527) subunits were almost complete, while the model of the b-subunit is missing the aminoterminal domain (NTD) residues 1-126 in which three crosslinks with the g-subunit were found (bK19 or bK86 to gK184 and bK89 to gK401; Supplementary Table 4). Cross-linked peptides identified within eIF2 are in good agreement with both the crystal structure of the yeast a-subunit (Fig. 3c) and the fold of the homologous structures (Fig. 3a). One exception is the bK260-gK449 cross-link and implies that the C-terminal end of the b-subunit (bK260) is flexible enough to come into the vicinity of gK449. Another exception is domain I of eIF2a, where K61 cross- Table 4) and the g-subunit G domain (Fig. 3a).
The first two domains of eIF2a and the C-terminal region of eIF2b are thought to be highly mobile 40 , making many transient dynamic interactions, which probably explains these crosslinks.
eIF2B subunit domain arrangements based on inter-crosslinks. Similar to eIF2, we generated homology models for eIF2B subunits (Supplementary Table 6). The majority of these, however, are based on homologous structures of proteins with different or unknown functions and lower sequence identity than was observed for eIF2. For some proteins, we could only generate homology models for separate domains rather then full-length proteins. We therefore mapped intra-subunit cross-links at a domain level (Fig. 4a-e). Based on our findings that gand e-subunits probably form a tetrameric hydrophobic core, we propose that gand e-subunits interact to form a g 2 e 2 hetero-tetramer, resembling the arrangement of the structural homologue-homotetramer ADP-glucose  Table 2).
pyrophosphorylase (AGP). Similar to AGP, these eIF2B subunits both have an N-terminal PL domain and a second domain that is predicted to adopt an LbH (Fig. 4d,e). This arrangement agrees well with those proposed in recent studies mapping domain interactions between these two subunits by complementary methods 27,28 and with identified cross-links between gK249 and eK176 (Fig. 5a). e-cat is shown adjacent to the eLbH as it is not required for interactions with other eIF2B subunits 23,27 . Two cross-links between adjacent residues in the eIF2Bd RLF domain (K421 and K422) to 2BeK356 (LbH) and 2BgK410 (LbH), respectively, indicate that residues eK356 and gK410 are proximal, as expected when g and e interact via their LbH domains, and locate eIF2Bd relative to the g 2 e 2 core (Fig. 5b). Position of the d-subunit is further constrained by residues K579 and K580 in its RLF domain, proximal to K422 in the homologous structure, cross-linked to gK410. Two further cross-links in the d carboxy-terminal domain (CTD) RLF domain dK556 and dK564 to eK4, which in turn is cross-linked to the eK520 at the end of e-LbH, suggest that the d-subunit is inclined towards the e-subunit (Fig. 5b).
Arrangement of eIF2Ba is constrained by the cross-link between aK107 (NaH) and gK412 (LbH), two residues apart from gK410, which cross-links to dK422, restricting the position of eIF2Ba, not only relative to gbut also to the d-subunit   Tables 2 and 3).
NATURE COMMUNICATIONS | DOI: 10.1038/ncomms4902 ARTICLE (Fig. 5b). The cross-linked residues bK380 and eK9 of eIF2B are not present in the homology models; however, approximate location of the eIF2Bb can be inferred based on available residues in close proximity (bQ371 and eK4, Fig. 4b). The location of the b-subunit is further confirmed by cross-links K136 and K380 to the d-subunit residues K648 and K95, and the cross-link between bK329 and gK494 (Fig. 5c). These constraints position the b-subunit relative to aand d-subunits, which we propose form a trimer by extensive interactions between their RLF domains in agreement with previous studies, which demonstrate interactions between a, b and d to form an eIF2 phosphorylation-sensing regulatory subcomplex 9,41 (Fig. 5d).
We completed our structural model of the eIF2B decamer by symmetry, reflecting the organization of subunits on a protein domain level (Fig. 5d). We assembled the regulatory subunits a, b and d to form asymmetric trimers interacting through RLF domains with NaH domains oriented outwards (Fig. 5c). Binding of the regulatory subcomplex to the g 2 e 2 core occurs through both e and g LbH domains. band d-subunits are inclined towards the e-subunit, providing more extensive interactions with e LbH and extending to form interactions with the e PL domain.

Solvent accessibility measurements.
To provide complementary evidence for the eIF2B subunit arrangement, we probed the exposed surfaces of intact eIF2B by chemically labelling. We subjected the eIF2B complex to the labelling reagent (diethylpyrocarbonate (DEPC)) for 1 min. Following enzymatic digestion and assignment of the labelled peptides using nano LC-MS/MS, we located the solvent accessible residues modified by the labelling reagent, onto the proposed subunit structure models (Fig. 5e,f). Within the catalytic domain, for which a structure is available, the surface residues identified reassuringly map close to the identified catalytic residue E569 (ref. 18), showing this region is accessible for eIF2 interaction. The results also confirm that our proposed interaction interfaces between gand e-subunits are not modified by the labelling reagent and so support our proposed model of the g 2 e 2 core complex (Fig. 5e). The crystal structure for eIF2Ba revealed a strong dimer interface along one hydrophobic edge of the RLF domain (pdb 3ECS). The chemical labelling approach showed that although this subunit is generally solvent accessible, this hydrophobic RLF edge is not, suggesting a general peripheral location for eIF2Ba within the complex, but allowing interaction via the RLF hydrophobic edge. Similarly, the eIF2Bb and d-subunit solvent accessibility measurements support a model placing them at the exterior of the protein complex (Fig. 5f).
Model for eIF2-eIF2B interactions based on cross-links. As eIF2B is a dimer with two catalytic domains, it has the potential to interact with two molecules of eIF2. Similarly, a single eIF2 trimer has the potential to interact with residues from both copies of each eIF2B subunit. Therefore, our proposed 'working model' represents one possible solution to this complex problem derived from the cross-linking data we obtained, rather than a definitive interaction model. To position eIF2 relative to eIF2B in the eIF2:eIF2B complex we used five critical cross-links: eIF2a K175:eIF2Bd K365 and eIF2a K104:eIF2Bd K519 (red dashed lines, Fig. 6a) together with eIF2b K183, b K240 and g K113 to eIF2Bg K249 (green dashed lines, Fig. 6a). The cross-links between eIF2a and eIF2Bd subunits bring the N-terminal domain of eIF2a containing Ser51 close to the interface between band d-subunits of eIF2B. Although the other three eIF2 residues are close to the GDP/GTP nucleotidebinding site of the g-subunit of eIF2, eIF2Bg K249 is located in the PL domain and the observed cross-links place it in close proximity to the nucleotide-binding pocket of eIF2g.
Other cross-links between eIF2 and eIF2B confirm this arrangement. For example, eIF2g K518 cross-links to K89 of eIF2Bd located on the same side as eIF2Bg, involved in interactions with eIF2g. Taking into account that eIF2b is flexible and that it is not possible to identify the structural arrangement of subunits at the atomic level, we cannot determine which one of the two eIF2Ba K145 residues are cross-linked to eIF2b K170 (yellow dashed lines, Fig. 6a). Similarly, the cross-link between eIF2b K247 and eIF2Bd K422 could be formed to either of the two eIF2Bd subunits. In agreement with our proposed model, we have identified cross-links between the catalytic domain of eIF2Be (K691) and eIF2a (K67) (Fig. 6a). We have further cross-linked eIF2-eIF2Be subcomplex comprising (i) full-length e-subunit and (ii) the catalytic domain of e ( Supplementary Fig. 5). Cross-links observed in these complexes confirm interactions between the catalytic e-subunit and eIF2. Additional intra-cross-links (Supplementary Table 4) in eIF2 subunits that are not present when eIF2 is in complex with decameric eIF2B suggest that these residues may be shielded by the interaction with eIF2B subunits (Fig. 3). A similar explanation may account for our inability to identify any cross-links between the catalytic domain of e-subunit and eIF2g.

Discussion
In this study we assessed the composition of eIF2 and eIF2B complexes and their interactions by MS coupled with chemical cross-linking and quantitative proteomics. Although it is generally accepted that eIF2B is composed of five non-identical subunits in unit stoichiometry in yeast 19 and human 28 , we show that this complex exists as a decamer in 2:2:2:2:2 stoichiometry in line with very recent results for human eIF2B 42 . We used a BS3d 0 /d 4 cross-linker to determine interaction sites between eIF2B subunits, as well as interactions between eIF2B and eIF2. Based on these results, together with homology modelling, we present a model of subunit interactions within the eIF2B decamer and also its interactions with eIF2. We propose that the assembly of the eIF2B decamer occurs through the formation of a g 2 e 2 tetrameric core, resembling the subunit arrangement in AGP homotetramer. The a-, band d-subunits are arranged in asymmetric trimers associated to the core through e and g LbH domains. This arrangement is in agreement with prior mutagenesis and pulldown studies demonstrating that both e and g LbH domains mediate the interactions between the catalytic and regulatory subcomplexes in both yeast and human eIF2B 27,28 . We found that the catalytic g 2 e 2 -core has an additional mass bound, that may include GTP and showed directly that GTP can bind to the g-subunit, but not to the e-subunit. As pyrophosphorylase enyzymes bind specific nucleotides (for example, AGP binds ATP), it is reasonable to propose that the eIF2Bg PL domain within its N-terminal half binds GTP. This idea is supported by site-directed mutagenesis around the nucleotidebinding site in yeast e PL domain, which has only modest effect on eIF2B activity 27 . By contrast, mutation of Tyr-38, close to the putative nucleotide-binding site in human eIF2Bg, severely impairs catalytic function 28 , implying possibly that GTP binding to eIF2Bg is important for its function. Two models for GTP binding to eIF2B have been proposed previously. First, an allosteric role has been suggested in which GTP binding is not directly involved in the catalytic mechanism, but instead acts as a signal that energy in the form of GTP is available to support continued protein synthesis. The second idea is that GTP binding to eIF2Bg acts as a direct source of GTP for nucleotide exchange. This latter idea is supported by prior enzyme kinetic analysis with the yeast factor 43 and by our cross-linking result that eIF2Bg K249 (within the PL domain) crosslinks to eIF2g K113 (within the G domain), thereby placing the GTP-binding sites for each factor in close proximity to each other (Fig. 6a). We favour this latter idea.
Based on our findings and taking into account other data, we propose that eIF2B assembly provides not only the catalytic function for guanine nucleotide exchange but also acts as a scaffold to hold eIF2 in place during catalysis. The eIF2B decamer of B600 kDa may be capable of interacting with and catalysing exchange on two eIF2 complexes, simultaneously, but for simplicity we show only a single eIF2 here. From our model of eIF2-eIF2B interactions (Fig. 6a) and our identification of a GTPbinding site on eIF2Bg, it seems likely that guanine nucleotide exchange involves several steps and these ideas are shown schematically in Fig. 6b. As proposed previously, initial binding of eIF2-GDP to eIF2B allows the catalytic domain of e to reach the interface between band g-subunits of eIF2 (Fig. 6b-1) and to promote GDP release from eIF2g ( Fig. 6b-2) 25,26 . By analogy with other G proteins, GDP release probably involves conformational changes in eIF2 subunits (Fig. 6b-3) followed by subsequent binding of GTP (Fig. 6b-4). The novel aspect of our model is that GTP is transferred from eIF2Bg to the vacant site on eIF2g (Fig. 6b-4). As the affinity of eIF2 for GDP is much higher than for GTP 43 an immediate transfer of GTP to the vacated site on eIF2g might be necessary. It seems reasonable to speculate that GTP binding to eIF2g may then induce further conformational changes and dissociation of the two factors (Fig. 6b-5). Such a mechanism fits with the 'substituted' rather than 'sequential' mechanism for guanine nucleotide exchange, which is supported by prior enzyme kinetic analyses 34 . A sequential enzyme mechanism assumes that GDP is released from the eIF2-GDP:GEF complex before binding of GTP and GEF dissociation, whereas the substituted mechanism proceeds by the formation of a ternary complex containing both substrates eIF2:GDP:GTP:GEF before release of any the products 34,44 .
If we consider eIF2B being a platform for eIF2, accommodating conformational changes during catalysis, this may explain the regulatory mechanism of eIF2a phosphorylation at residue Ser51. Tight binding of the phosphorylated Ser51 to regulatory subunits is likely to interfere with conformational changes in eIF2 necessary for catalysis and, therefore, would abrogate eIF2B function. According to our model, Ser51 binds at the interface between band d-subunits of eIF2B, in line with extensive prior genetic and biochemical studies 19,20,25,41,45 . In our model, eIF2Ba stabilizes interactions within the regulatory sub-complex making the structure more rigid. Therefore, mutations within the eIF2Ba subunit, or its elimination from the complex, may increase spatial flexibility of b and d to allow conformational changes in eIF2 and permit nucleotide exchange.
Overall, our study advances understanding of the structural arrangement of the eIF2B complex and defines its interactions with eIF2. We have shown that eIF2B exists as a decamer formed around a tetrametic catalytic core. This hydrophobic g 2 e 2 core suggests that in an evolutionary context the catalytic function may have preceded its regulatory role. The regulatory function, enabled by the later addition of the regulatory subunits resulting in the structural intricacy described here, making eIF2B one of the most complex GEFs known.

Methods
Protein expression and purification. Yeast Saccharomyces cerevisiae strains engineered to overexpress eIF2, eIF2B, eIF2Be, eIF2Be-cat and four-subunit eIF2B complex (lacking a-subunit) employed in this study were-GP3511 (MATa leu2-3 leu2-112 ura3-52::HIS4-lacZ ino1 gcn2D pep4::LEU2 sui2D pAV1089[SUI2 SUI3 GCD11-His6 2 mm URA3]) 19  His-tagged yeast eIF2 was purified using Ni-chelate chromatography followed by chromatography on heparin-Sepharose 16 . FLAG-tagged eIF2B complexes and subunits were purified in a single affinity step in a high-salt buffer (500 mM KCl) according to ref. 33, with minor modifications for the five-subunit eIF2B: Briefly, after immobilizing the protein on anti-FLAG resin, we used 500 mM ammonium acetate as a wash and elution (containing FLAG peptide) buffer to reduce manipulation steps and time before native MS, which requires transfer of a protein complex to a volatile buffer (such as ammonium acetate) for efficient ionization in the electrospray source of a mass spectrometer. The eluted fraction (B400 ml) was concentrated in an Amicon Ultra (100 K MWCO) 0.5 ml centrifugal filter (Millipore) and washed ten times with 400 ml of 500 mM ammonium acetate to remove FLAG peptide, and then concentrated to a final volume of 50 ml. Three microlitres of the final sample were withdrawn for SDS-PAGE analysis.
MS of intact eIF2 and eIF2B protein complexes. For intact MS protein containing solutions were exchanged into 500 mM ammonium acetate using Amicon centrifugal devices. MS and MS/MS spectra were acquired on a high mass quadrupole time-of-flight-type instrument 46 adapted for a QSTAR XL platform (MDS Sciex) 47 using in-house prepared gold-coated glass capillaries 26 . Optimized instrument parameters were as follows: ion spray voltage 1.3 kV, declustering potential 100 V, focusing potential 200 V and collision energy up to 200 V, MCP 2,350 V. In MS/MS, the relevant m/z range was selected in the second quadrupole and subjected to acceleration in the collision cell.
Digestion with trypsin. The proteins were separated by gel electrophoresis using the NuPAGE gel system (Invitrogen) according to the manufacturer's protocol. Protein bands of interest were cut and the proteins therein were digested in-gel as previously described 48 . Alternatively, the proteins were precipitated with ethanol and the protein pellet was dissolved in 10 ml 1% (m/v) RapiGest SF Surfactant (Waters). The proteins were reduced by addition of 10 ml 25 mM dithiothreitol and incubation at 37°C for 1 h. Free cysteine residues were alkylated with 10 ml 100 mM iodoacetamide at 37°C for 1 h. The RapiGest concentration was reduced to 0.1% (m/v) by addition of 25 mM ammonium bicarbonate. The proteins were digested with 0.5 mg Trypsin (Promega) at 37°C overnight. RapiGest was decomposed by addition of 5% (v/v) trifluoroacetic acid and removed by centrifugation. The peptides were dried in a vacuum centrifuge.
The nanoLC system was coupled to an LTQ-Orbitrap XL hybrid mass spectrometer (Thermo Scientific) and peptides were directly eluted into the mass spectrometer. Typical mass spectrometric conditions were as follows: spray voltage of 1.8 kV; capillary temperature of 180°C; normalized collision energy of 35% at an activation of q ¼ 0.25 and an activation time of 30 ms. The LTQ-Orbitrap XL was operated in data-dependent mode. Survey full scan MS spectra were acquired in the orbitrap (m/z 300-2,000) with a resolution of 30,000 at m/z 400 and an automatic gain control target at 10 6 . The five most intense ions were selected for CID MS/MS fragmentation in the linear ion trap at an automatic gain control target of 30,000. Detection in the linear ion trap of previously selected ions was dynamically excluded for 30 s. Singly charged ions as well as ions with unrecognized charge state were also excluded. Internal calibration of the Orbitrap was performed using the lock mass option (lock mass: m/z 445.120025 (ref. 49)).
Protein-protein cross-linking. Ten microlitres of a 1:1 mixture of 2.5 mM deuterated (d4) and non-deuterated (d0) BS3 (Thermo Scientific) were added to 240 ml of purified eIF2-eIF2B complex. The reaction mixture was incubated for 30 min at 26°C and 450 r.p.m. in a thermomixer. One hundred and sixty microlitres of 100 mM ammonium bicarbonate were added to quench the remaining cross-linking reagent. The proteins were precipitated with ethanol and digested in solution with RapiGest SF Surfactant (see above) or separated by SDS-PAGE and digested in gel (see above). The peptides obtained from in-solution digestion were re-dissolved in 20% (v/v) ACN, 4% (v/v) FA and separated by cation exchange chromatography using SCX stage tips (Thermo Scientific) according to the manufacturer's protocol. Peptides were eluted with different concentrations of ammonium acetate (25 mM-500 mM) and dried in a vacuum centrifuge. The mixture of cross-linked and non-cross-linked peptides was analysed by nanoLC-MS/MS. Potential crosslinks were identified by using the MassMatrix Database Search Engine [50][51][52] . Search parameters were as follows: peptides were defined as tryptic with a maximum of two missed cleavage sites. Carbamidomethylation of cysteines and oxidation of methionine residues were allowed as variable modifications. The mass accuracy filter was 10 p.p.m. for precursor ions and 0.8 Da for fragment ions. Minimum pp and pp2 values were 5.0 and minimum pp tag was 1.3. Maximum number of crosslinks per peptide was 1. All searches were performed twice, including the deuterated and the non-deuterated cross-linker, respectively. Potential crosslinks were validated manually by (i) checking the presence of the according peak pair in the MS spectra generated by the d4/d0-BS3 mixture and (ii) by the quality of the MS/MS spectrum.
Absolute quantification of the eIF2-eIF2B complex. The proteins were digested with RapiGest and analysed by nanoLC-MS/MS (see above). The raw data were searched against Uniprot_yeast database using MaxQuant software 53 v1.2.2.5. The mass accuracy filter was 20 p.p.m. for precursor ions and 0.5 Da for MS/MS fragment ions. Peptides were defined to be tryptic with maximal two missed cleavage sites. Carbamidomethylation of cysteines and oxidation of methionine residues were allowed as variable modifications. iBAQ 54 was implemented in the MaxQuant search. Obtained iBAQ intensities of the eIF2-eIF2B proteins were compared to determine the stoichiometry of the two complexes.
GTP binding. To investigate GTP binding, purified eg-core complex was incubated with excess (10 mM) of 6-Thio-GTP. Ultraviolet-cross-linking was performed on ice at 365 nm for 5 min. The protein complex was analysed before (control) and after ultraviolet cross-linking as described above.
DEPC labelling. Approximately 5 mg of eIF2B complex were incubated with 70 mM DEPC at 37°C. After 1 min, the reaction was quenched by addition of 5 ml of 10 mM Imidazole. The proteins were digested with 0.4 mg Tryspin at 37°C overnight and the peptides were subsequently subjected to LC-MS/MS analysis (see above). Modified sites were identified by database search using MassMatrix Database Search Engine (see above), allowing for DEPC modification at His, Tyr, Ser and Thr residues 55 .
Homology modelling. Homology models for yeast eIF2 (a, b, g) and eIF2B (a, b, d, g, e) were generated using MODELLER web server (https://modbase.compbio.ucsf.edu/scgi/modweb.cgi) and SWISS-MODEL automated protein modelling server (http://swissmodel.expasy.org). eIF2 subunits (a res. 7-267, g res. 96-527 and b res. 127-261) were modelled using the crystal structure of archaeal aIF2 from S. solfataricus, PDB 3CW2 as a template (Fig. 3a). Homologous regulatory subunits of eIF2B a and b were modelled based on a crystal structure of human eIF2Ba (PDB 3ECS) (SWISS-MODEL). For the d-subunit, the best homologous structure based on the crystal structure of 5-methylthioribose 1phosphate isomerase from Bacillus subtilis (PDB 2YVKA) (MODELLER) contained residues 245-536 within the homologous region. As most of the cross-links were identified within the C and N termini, we submitted the CTD (540-651) and NTD (1-244) sequences of the d-subunit for homology modelling (MODELLER) separately. For these, we obtained homologous structures based on the crystal structure of Ribose-1,5-bisphosphate isomerase, Thermococcus kodakaraensis (PDB 3A11A) and (1YA9A) for CTD and NTD, respectively. The CTD structure constitutes part of the RFL domain, while the NTD corresponds mainly to the NaH domain (Fig. 4c).
We used AGP and GlmU as templates to obtain models of eIF2B gand e-subunits. For eIF2Bg, only the structure based on GlmU template reached completion when modelled (SWISS-MODEL). Therefore, we also modelled PL  and LbH (358-578) domains of eIF2Bg separately. Intra-subunit crosslinks identified in this study within eIF2Bg indicate that the arrangement of the PL and LbH domains is more consistent with the AGP-like rather than GlmU domain arrangement in agreement with previous experimental results 27 . In AGP, the PLD and LbH domains have more extensive interactions forming a more compact structure as opposed to the LbH domain extending away from the PL domain in GlmU. The eIF2Be residues 1-520 excluding e-cat were submitted for modelling to MODELLER. This resulted in a homology model of e (residues 30-470) based on the AGP template (PDB 1YP2). For e-cat, we used crystal structure available (PDB 1PAQ) for the yeast catalytic domain 18 .