Binding of the integrated stress response inhibitor, ISRIB, reveals a regulatory site in the nucleotide exchange factor, elF2B

The Integrated Stress Response (ISR) is a conserved eukaryotic translational and transcriptional program implicated in mammalian metabolism, memory and immunity. The ISR is mediated by stress-induced phosphorylation of translation initiation factor 2 (eIF2) that attenuates the guanine nucleotide exchange factor eIF2B. A chemical inhibitor of the ISR, ISRIB, a bis-O-arylglycolamide, reverses the attenuation of eIF2B by phosphorylated eIF2, protecting mice from neurodegeneration and traumatic brain injury. We report on a cryo-electron microscopy-based structure of ISRIB-bound human eIF2B revealing an ISRIB-binding pocket at the interface between the β and δ regulatory subunits. CRISPR/Cas9 mutagenesis of residues lining this pocket altered the hierarchical cellular response to ISRIB congeners in vivo and ISRIB-binding in vitro, thus providing chemogenetic support for the functional relevance of ISRIB binding at a distance from known eIF2-eIF2B interaction sites. Our findings point to a hitherto unexpected allosteric site in the eIF2B decamer exploited by ISRIB to regulate translation.


Introduction
Stress-induced phosphorylation of the α subunit of eukaryotic translation initiation factor 2 (eIF2α) is a highly conserved mechanism for regulating translation initiation (reviewed in Sonenberg and Hinnebusch, 2009). The resulting attenuation of global protein synthesis and the activation of translation from a rare subset of mRNAs that encode potent transcription factors is the basis of an Integrated Stress Response (ISR) (Dever et al., 1992;Harding et al., 2003). Whilst the ISR has important homeostatic functions that increase fitness, in some circumstances a benefit arises from attenuated signalling in the ISR (reviewed in Pakos-Zebrucka et al., 2016). This failure of homeostasis has encouraged the search for ISR inhibitors, which culminated in the discovery of ISRIB, the first small molecule ISR inhibitor (Sidrauski et al., 2013;Sidrauski et al., 2015a). ISRIB has proven efficacious in certain mouse models of neurodegeneration (Halliday et al., 2015), traumatic brain injury (Chou et al., 2017) and even as a memory-enhancing drug in normal rats (Sidrauski et al., 2013).
ISR inhibition in ISRIB-treated cells is observed despite elevated levels of eIF2(αP) (Sidrauski et al., 2013), indicating that ISRIB's site of action lies downstream of the stress-induced kinases that phosphorylate eIF2α. The cellular target of eIF2(αP) is eIF2B, a large protein complex that possesses guanine nucleotide exchange (GEF) activity directed towards eIF2 (Panniers and Henshaw, 1983). The interaction of eIF2(αP) with eIF2B attenuates the recycling of eIF2 to its active, GTP-bound form, thereby attenuating rates of translation initiation on most mRNAs (Siekierka et al., 1981;Clemens et al., 1982) whilst enhancing initiation on rare mRNAs with peculiar 23-Nov-17 V1.7.3.1F_BioRxiv 5' untranslated regions (Mueller and Hinnebusch, 1986;Harding et al., 2000). Thus the ISR may be monitored by the activity of its target genes.
Genetic and structural observations indicate that the inhibitory effect of eIF2(αP) on eIF2B arises from the engagement by the α subunit of eIF2 at a cavity formed by the convergence of the α, β and δ regulatory subunits of the eIF2B decamer (Vazquez de Aldana and Hinnebusch, 1994;Pavitt et al., 1997;Kashiwagi et al., 2016;Bogorad et al., 2017). This regulatory site is distant from the catalytic γε subcomplex of eIF2B and its consecutive Asn-Phe residues that engage the nucleotide binding γ subunit of eIF2 to effect nucleotide exchange (Kashiwagi et al., 2016;Kashiwagi et al., 2017).
In vitro, addition of ISRIB accelerates eIF2B-mediated GDP dissociation from purified eIF2 Sidrauski et al., 2015b) and somatic cell genetic experiments have revealed that targeting eIF2B's δ regulatory subunit can impart ISRIB resistance to cells Sidrauski et al., 2015b). An interaction of ISRIB with eIF2B is also consistent with the finding that the eIF2B decamer isolated from ISRIB-treated cells (or cell lysates) is more stable in a density gradient than eIF2B from untreated cells (Sidrauski et al., 2015b', figure 3 therein). However, the ISRIB-resistant mutations identified to date cluster at a distance from both the regulatory site engaged by eIF2(αP) and the catalytic site engaged by eIF2γ (Kashiwagi et al., 2016;Kashiwagi et al., 2017). Thus, whilst the bulk of the evidence suggests that ISRIB binds eIF2B to regulate its activity, indirect modes of action are not excluded. Here we apply biophysical, structural and chemogenetic methods to reveal the presence of an ISRIB-binding pocket in eIF2B and provide insight into ISRIB's mode of action.

ISRIB binds directly to purified mammalian eIF2B
In keeping with previously-published observations (Sidrauski et al., 2015b), a stabilizing effect of ISRIB was observed on the eIF2B complex derived from HeLa cells in which the endogenous eIF2B β subunit was tagged ( Figure 1A). This tagged complex was purified from intact cells ( Figure 1B) and used to further study ISRIB's interaction with eIF2B.
To observe ISRIB's interaction with eIF2B directly, we designed a fluorescently labelled derivative (AAA2-101) based on known structure-activity relationships of ISRIB derivatives (Hearn et al., 2016)  Using this compound we measured the effect of added purified eIF2B on the fluorescence polarization (FP) signal. FP increased with increasing concentrations of 23-Nov-17 V1.7.3.1F_BioRxiv eIF2B ( Figure 1C). At the concentrations of eIF2B available for testing, the FP signal was not saturated, thus we have been unable to extract a reliable dissociation constant from this assay. However, unlabelled ISRIB effectively competed for eIF2B in the FP assay, with an EC 50 in the nanomolar range, as observed for ISRIB action in cells (Figure 1-supplement 1B), whilst less active congeners competed less successfully ( Figure 1D and Figure 1-supplement 1C). The FP assay thus likely reports on engagement of the FAM-labelled ISRIB derivative at a site(s) on eIF2B that is relevant to ISRIB action.

Cryo-EM reveals an ISRIB binding pocket in eIF2B
The structures of several eIF2B-related complexes have been solved by X-ray crystallography: the human α 2 dimer (Hiyama et al., 2009), the C. thermophilum (βδ) 2 regulatory sub-complexes (Kuhle et al., 2015) and the complete S. pombe eIF2B decamer (Kashiwagi et al., 2016). However, as fission yeast eIF2B does not respond to ISRIB and as there is an inherent uncertainty if the partial complexes would bind ISRIB, we elected to solve the structure of the ISRIB-bound human eIF2B complex by single particle cryo-electron microscopy (cryo-EM).
We purified endogenous human eIF2B from HeLa cell lysates in the presence of ISRIB and determined the structure of the ISRIB-eIF2B complex at an overall resolution of 4.1-Å (   human catalytic subcomplex was lower compared with the regulatory core, consistent with flexibility of this region in both human and yeast eIF2B (Kashiwagi et al., 2016). Similar to the S. pombe eIF2B crystal structure, we were unable to resolve the catalytically important C-terminal HEAT domain of the ε subunit in our cryo-EM map.
Due to lack of sufficient high-quality cryo-EM images of an apo-eIF2B complex, we were unable to calculate a difference map of eIF2B with and without ISRIB. However, a nearly continuous density with a shape and size of a single ISRIB molecule was conspicuously present at the interface of the β and δ regulatory subunits. The quality of the map in this position provided sufficient detail to confidently model the ISRIB molecule ( Figure 2C). The ISRIB binding pocket is located at the plane of symmetry between the β and δ subunits. In the central part of the pocket, residue H188 of the β subunit is notable, as its side chain is positioned in the vicinity of the essential carbonyl moiety of ISRIB (Hearn et al., 2016). Residue N162 of the β subunit also stabilises the diaminecyclohexane moiety of ISRIB, possibly through hydrogen bonding interactions. More distally in the pocket, various residue side chainsnamely δL179 (the human counterpart of hamster δL180), δF452, δL485, δL487, βV164, βI190, βT215, and βM217 -form the hydrophobic end of the symmetrical pocket that accommodates the aryl groups of ISRIB ( Figure 2-supplement 2C).
The observation that a hamster Eif2b4 L180F mutation (δL179 in the human) disrupts ISRIB action in cells  is consistent with a clash between the bulkier side chain of phenylalanine and the bound ISRIB molecule. The ISRIB binding  8 site identified in the map is also consistent with the lack of a corresponding density in the S. pombe apo-eIF2B map.
Overall the human and yeast structures are highly similar (r.m.s.d of 2.57 Å over 3049 alpha carbons) with no evidence of major domain movements that might be attributed to ISRIB binding in the human complex ( Figure 2 -supplement 2D). The region corresponding to the N-terminal ~166 residues of the human δ subunit remains unresolved (in both yeast and human structures).
However, residues 167-204 of the human protein, which can be reliably placed in the density, assumed a very different conformation from the corresponding segment of the yeast δ subunit ( Figure 2D). It is unclear if this difference reflects speciesdivergence in the structure of eIF2B, a crystallization-induced difference or is a consequence of ISRIB binding. It is noteworthy, however, that this divergent region emerges from the putative ISRIB density at the (βδ) 2 interface and that it contains two residues, human eIF2B δ R170 and V177, whose mutation interferes with ISRIB action in hamster cells, δ R171 and V178 (reported in Sekine et al., 2015).

Chemogenetic analysis of the putative ISRIB-binding pocket
To validate the mode of ISRIB binding revealed by the structural data ( Figure 2C) we randomized the codons, encoding residues predicted by the model to line the ISRIBbinding pocket, in hope of eliciting a correlation between amino acid substitutions and ISRIB's activity in such mutagenized cells. To achieve this, we used CRISPR/Cas9 to target the aforementioned Eif2b2 locus of CHO-K1 cells and provided a repair template randomized at either Eif2b2 N162 , Eif2b2 H188 or Eif2b2 I190 codons ( Figure 3A).  Exposure to histidinol, an agent that activates the eIF2α kinase GCN2, induces an ISR, which can be tracked by a CHOP::GFP reporter in living cells . In a first round of sorting, histidinol-treated, mutagenized pools of cells, were segregated by phenotypic selection into ISRIB sensitive [ISRIB SEN , CHOP::GFP inhibited ("GFPdull")], and ISRIB resistant [ISRIB RES , CHOP::GFP activated ("GFP-bright")] classes. In a second round of sorting untreated populations from the ISRIB SEN and ISRIB RES pools were purged of GFP-bright cells that had acquired a constitutively active ISR phenotype ( Figure 3B).
Using fluorescence activated cell sorting (FACS) we successfully isolated ISRIB SEN cells, in which ISR induction by histidinol was readily counteracted by ISRIB, as well as ISRIB RES cells, in which ISR induction was maintained upon ISRIB treatment ( Figure   3C -compare left and right panels, red traces). To determine if the phenotypicallydistinguished pools were enriched in different mutations, we subjected genomic DNA derived from each population to deep sequencing analysis (Figure 3D and table   S2).
The ISRIB RES pool of cells targeted at Eif2b2 H188 diverged dramatically from the parental sequence ( Figure 3D -middle panel, plum bars). Of a total of 250,617 reads histidine was present in only 6,443 (2.6%), with arginine, glycine, leucine, lysine and glutamine, dominating (24%, 21%, 18%, 8.2% and 6.2%, respectively). Histidine was preserved in the ISRIB SEN pool (269,253 of 328,113 reads, 82%), which was constituted largely of parental alleles that escaped targeting altogether and targeted alleles that had acquired a synonymous mutation ( Figure 3D -middle panel, orange bars). These observations point to the importance of histidine at position 188 of the 23-Nov-17 V1.7.3.1F_BioRxiv β subunit in mediating ISRIB sensitivity. The bias in a favour of certain substitutions in the ISRIB RES pool likely arises by their ability to preserve eIF2B function whilst eliminating responsiveness to ISRIB.
The ISRIB RES pool of cells targeted at Eif2b2 I190 was dominated by tryptophan, methionine and tyrosine (28%, 24% and 15%, respectively) consistent with a role for these bulky side chains in occluding the ISRIB binding pocket ( Figure 3D  Mutagenesis of Eif2b2 N162 was less successful in generating a pool of strongly ISRIBresistant cells ( Figure 3C, panel 2, red trace). This is not a reflection of failure of the homologous recombination strategy, as the incorporation of a silent PAM disrupting mutation enabled us to restrict the sequencing analysis to alleles that had successfully undergone homologous recombination. The high frequency with which the parental asparagine had been retained in both ISRIB SEN and ISRIB RES pools (82% and 44%, respectively) suggests that replacement of this residue is selected against (Eif2b2 is an essential gene), with the one exception being acquisition of a threonine, which was greatly enriched (26%) in the ISRIB RES pool ( Figure 3D -top panel, plum bars).
While these features of the ISRIB RES mutations suggest that they exert their effect by altering the character of an ISRIB-binding pocket, it is impossible to exclude an alternative possibility that they alter eIF2B structure so as to disrupt the communication between an ISRIB binding site that might be located elsewhere and 23-Nov-17 V1.7.3.1F_BioRxiv the relevant catalytic activity of eIF2B. To examine this possibility, we tested pools of ISRIB RES cells for residual responsiveness to ISRIB congeners, reasoning that a mechanism involving an altered binding site might alter the hierarchy of potency of ISRIB congeners, whereas a mechanism involving disruption of an allosteric signal is unlikely to do so.
Compounds AAA1-075B and AAA1-084 are 3-to 11-fold less potent than ISRIB in inhibiting the ISR of wildtype cells (Figure 1-supplement 1B, Figure 4A, figure 4supplement 1A). However, both compounds were relatively more potent than ISRIB in reversing the ISR of the ISRIB RES pool of cells targeted at Eif2b2 H188 , attainting nearly complete inhibition when applied at micromolar concentration ( Figure 4B).
The faint, biphasic response of the mutant ISRIB RES pool to ISRIB is reproducible, but To consolidate the aforementioned findings, we exploited the diversity of ISRIB RES mutations in the Eif2b2 H188X population to select, by a new round of FACS ( Figure   5A), for sub-pools that acquired sensitivity to AAA1-075B or AAA1-084, or retained their sensitivity to ISRIB ( Figure 5B, compare continuous traces) and sequenced their Eif2b2 alleles ( Figure 5C and table S3). As expected, sorting for ISRIB sensitivity enriched, by over 20-fold for those rare wildtype H188 alleles that persisted the pool of ISRIB RES Eif2b2 H188X cells ( Figure 5C -compare plum to orange bars). H188 was also somewhat enriched (about 5-fold) in the pools sorted for their sensitivity to AAA1-075B (075B SEN ) or AAA1-084 (084 SEN ), but unlike the ISRIB SEN these pools were also  enriched for residues other than histidine ( Figure 5C, blue and cyan bars).
Importantly, selecting for sensitivity to these ISRIB congeners enriched for different residues than those found in the original ISRIB RES pool: arginine, glycine and leucine were depleted and replaced by lysine, serine, alanine and threonine ( Figure 5C, compare plum bars to blue and cyan bars).
To examine the sensitivity to ISRIB congeners in individual clones of Eif2b2 H188X mutant cells, we selected for study clones with unambiguous genotypes (Figure 6supplement 1A). CHO cells that had undergone the aforementioned selection scheme but remained homozygous for the parental Eif2b2 H188 allele retained their responsiveness to ISRIB and AAA1-084. The responsiveness of the mutant clones to ISRIB was greatly enfeebled, noticeable in both a shift to the right in the concentration-response curves and in the magnitude of inhibition observed at the highest concentration of compound ( Figure 6A, compare red traces across clones).
The biphasic response to ISRIB, observed in the pools of mutagenized cells ( Figure   4B), was also evident in the clonal populations. A similar shift to the right was noted in the response of the mutant clones to AAA1-084 but the inhibitory effect observed at the higher concentration was only slightly diminished ( Figure 6A, compare blue traces).
To address the effect of ISRIB-resistant mutations in eIF2B on binding of the FAMlabelled AAA2-101, we purified eIF2B from wildtype, Eif2b4 L180F and Eif2b2 H188K CHO cells by exploiting a 3xFLAG-tag knocked into the endogenous eIF2B γ subunit ( Figure   6B). The wildtype eIF2B gave rise to a conspicuous concentration-dependent FP signal in the presence of a FAM-labelled AAA2-101 ( Figure 6C, circular pictograms).
Validity of this FP signal was confirmed by competition with unlabelled ISRIB ( Figure   6 -supplement 1B). However eIF2B purified from the mutant cells failed to give rise to an FP signal ( Figure 6C, square and triangle pictograms), thereby establishing a correlation between ISRIB resistance in cells and defective ISRIB binding in vitro ( Figure 6A, C and Figure 6 -supplement 1C).

Discussion
We have identified the ISRIB binding site at the core of the regulatory complex by determining the cryo-EM structure of a human eIF2B-ISRIB complex. The residues that contact ISRIB can be readily identified in the experimentally-derived density map. Mutation of these residues leads to loss of sensitivity to ISRIB in cultured cells and to loss of ISRIB binding in vitro, but the same mutations exert less of an enfeebling effect on the cellular response to certain ISRIB congeners. Together these structural, biophysical and chemogenetic findings point to a role for ligand engagement at the aforementioned pocket in ISRIB action.
The ISRIB binding pocket straddles the two-fold axis of symmetry of the core regulatory complex and a single molecule of ISRIB appears to engage the same residues from opposing protomers of the (βδ) 2 dimer of dimers. This feature fits well with ISRIB's own symmetry and also nicely explains the ability of ISRIB to stabilize the eIF2B decamer in vitro.
The overall structure of ISRIB-ligated human eIF2B is similar to that of S. pombe eIF2B, crystallized in the absence of ISRIB. This finding argues against large domain movements as the basis of ISRIB action. However, as we do not have information on the apo structure of human eIF2B, we cannot exclude the possibility that ISRIB stabilizes an active conformation that was also fortuitously assumed by yeast eIF2B in the crystal. Alternatively, it is possible that ISRIB stabilized an active conformation of eIF2B that entailed changes in the disposition of the invisible C-terminal HEAT repeats of the catalytic ε subunit or a conformation which, though relatively 23-Nov-17 V1.7.3.1F_BioRxiv enriched, remained too under-populated in the ensemble of cryo-EM particles for detection or was selectively under-represented in the images available for analysis.
It is intriguing however to ignore the potential for species divergence in structure and consider the crystallized yeast eIF2B as the apo structure and the cryo-EMderived model of human eIF2B as the ISRIB-bound, active conformation. This perspective suggests a major difference in the disposition of the largely invisible Nterminal extension of the δ subunit between the yeast (apo structure) and the ISRIBligated human structure. The region of the human δ subunit implicated contains two residues (R170 and V177) that are distant from the ISRIB binding pocket, but nonetheless important to ISRIB action (hamster δ R171 and V178, reported in Sekine et al., 2015). It is tempting therefore to speculate that an ISRIB-induced change in the disposition of this segment of the δ subunit contributes to the ISRIB-induced ISRIB antagonizes the eIF2(αP)-dependent ISR, but in vitro ISRIB enhances eIF2B GEF activity even in the absence of phosphorylated eIF2α Sidrauski et al., 2015b). Therefore, the consequences of any ISRIB-induced conformational change in eIF2B must not be limited to weaker binding of the phosphorylated form of eIF2. A recent study suggests that eIF2B exerts its GEF activity by stabilizing the transitional, nucleotide free (apo) state of eIF2 and that the difference in free energy between the interaction of the apo and nucleotide-bound eIF2 with eIF2B drives catalysis. Phosphorylation inhibits GEF activity by enhancing the interaction of the Nterminal portion of the α subunit of nucleotide bound eIF2 with the regulatory core of eIF2B, thus diminishing the aforementioned free energy gradient (Bogorad et al., 2017). It is tempting to speculate that by enfeebling the interaction of the nucleotide bound eIF2 with eIF2B, ISRIB binding might enhance the free energy difference to facilitate GEF activity.
The consequences of ISRIB binding may not be limited to its effect on GEF activity, as eIF2B also possess a guanine nucleotide dissociation inhibitor (GDI) -displacement activity, whereby eIF2B promotes the dissociation of a stabilizing complex between eIF5 and the GDP-bound eIF2 (Jennings et al., 2013). Enhancement of eIF2B's GDIdisplacing activity could contribute to ISRIB's activity as an ISR inhibitor.
The regulatory and catalytic subunits of eIF2B evolved from ligand regulated protein ancestors. ADP-glucose pyrophosphorylase has a nucleotide bind site that is conserved in its descendant catalytic γ and ε subunits, whereas ribose-1,5bisphosphate isomerase has a phospho-sugar binding pocket conserved in its descendant α, β and δ regulatory subunits (reviewed in Kuhle et al., 2015;Kashiwagi et al., 2017). The ISRIB binding pocket discovered here represents a third conserved feature of eIF2B. It is intriguing to consider that endogenous ligand(s) might exist 23-Nov-17 V1.7.3.1F_BioRxiv that engage the ISRIB binding site to regulate eIF2B in yet to be determined physiological states. KCl, 4mM Mg(OAc) 2 , 0.5% Triton, 5% Sucrose, 1 mM DTT, 2 mM PMSF, 10 μg/ml Aprotinin, 4 μg/μl Pepstatin, 4 μM Leupeptin]. Lysates were cleared at 21,130 g on a chilled centrifuge, 0.5 mL of the supernatant was applied on 5mL of 5-20% Sucrose gradient, that was prepared on SG15 Hoeffer Gradient maker in the cell lysis buffer with respective amounts of sucrose and equilibrated for 24 hours on ice. The gradient was run on a SW50.1 rotor at 40,000 rpm for 14 hours and 20 minutes and was equally fractionated into 13 fractions of 420 µL. Each fraction was then diluted two-fold with lysis buffer without sucrose, proteins were precipitated with 20% TCA for 16 hours at 4℃ and pelleted at 21,130 g for 20 minutes in a chilled centrifuge.

HeLa-derived cell line was maintained in DMEM
Protein pellets were washed twice with ice-cold acetone, air-dried, resuspended in 60 µL of alkaline SDS loading buffer and 5 µL of resuspension was run on a 12.5% SDS-PAGE gel. The protein gel was then transferred onto PVDF membrane and immunoblotted, using primary monoclonal mouse anti-FLAG M2 antibody (F1804, Sigma) and secondary polyclonal goat anti-mouse-IR800, followed by scanning on Odyssey imager (LI-COR Biosciences) and image analysis on ImageJ software.

Protein purification
Human eIF2B was purified from 50 L of HeLa-2C2 cells (3xFLAG-EIF2B2 in/in) grown in suspension at a maximum density of 10 6 cells/mL. Cell pellets (150 grams  Hamster eIF2B from CHO-S7 [Eif2b3-3xFLAG in/in; Eif2b4 wt ], CHO-S9 [Eif2b3-3xFLAG in/in; Eif2b4 L180F in/Δ] or CHO-B3 [Eif2b3-3xFLAG in/in; Eif2b2 H188K in/ Δ] was purified in a way similar to human eIF2B , with following exceptions. 23-Nov-17 V1.7.3.1F_BioRxiv Hamster eIF2B wt cells were harvested from 3.5 L of CHO-S7 suspension culture grown at maximum density of 10 6 cells/mL, washed with room temperature PBS and lysed in two pellet volumes of lysis buffer. Cleared lysate supernatant was incubated with 300 µL of anti-FLAG M2 affinity gel. The resin was washed three times for 5 minutes with 1 mL of lysis buffer containing 500 mM NaCl, then washed three times for 5 minutes with 1 mL of washing buffer [50 mM Tris (pH7.4), 150 mM NaCl, 2 mM MgCl 2 , 0.01% Triton, 1 mM DTT]. Protein was eluted in two resin volumes of washing buffer supplemented with 125 mg/mL 3X FLAG peptide. Hamster eIF2Bδ L180F along with eIF2Bβ H188K mutant cells (with total cell count of 50x10 7 each) were harvested and washed with room temperature PBS, then lysed in three pellet volumes of lysis buffer. Cleared lysate supernatant was incubated with 60 µL of anti-FLAG M2 affinity gel following binding, washing and eluting procedures as described above.
Site-directed random mutagenesis of hamster Eif2b2 was carried out using CRISPR/Cas9 homologous repair using an equimolar mixture of single-stranded oligo deoxynucleotides (ssODN) as repair templates containing combination of all possible 23-Nov-17 V1.7.3.1F_BioRxiv codons (n=64) for each targeted site proving the diversity of substitutions. Each set of ssODNs was transfected along with a sgRNA guide, inserted into a vector containing mCherry marked Cas9, in the following pairs: UK2105 with ssODN-1922or ssODN-1923, UK2106 with ssODN-1924 (see plasmid and primer tables for further description of guide vectors (UK2105, UK2106) (Table S4) and repair templates (Table S5).
For transfection 20x10 3 CHO-S7 cells  were plated 36 hours prior being transfected with equal amounts of 1 µg guide and 1 µg ssODN using standard Lipofectamine protocol, expanded and sorted 48 hours later. Preceding phenotypical sorts a technical "mCherry-positive" sort for successfully transfected cells was done and 240K, 360K and 330K cells were collected for N162X, H188X and I190X templates respectively. Five days later the recovered cell pools were treated with 0.5 mM histidinol and 200 nM ISRIB. 20 hours later pools were sorted for "GFP-bright" (ISRIBresistant) and "GFP-dull" (ISRIB-sensitive) phenotypes (Sort I). 37K, 126K and 240K cells for each of respective ISRIB-resistant pools: N162X, H188X and I190X, and 500K for each of ISRIB-sensitive pools were collected. 10 days later "GFP-dull" sort of untreated pools (Sort II) was carried out, to eliminate clones with a constitutively active ISR. 1 mln cells were collected for each of ISRIB-resistant and ISRIB-sensitive pools.
For a new round of sorting to reselect congener-sensitive pools from ISRIB-resistant Eif2b H188X pool, 1 mln cells were plated 36 hours before being treated with 0.5 mM histidinol and 2.5 µM of ISRIB or 2.5 µM ISRIB congener (AAA1-075B or AAA1-084). 20 hours later cells were sorted for "GFP-dull" (congener-sensitive) phenotype, and 1 mln cells were collected for each respective pool (ISRIB-sensitive, 075B-sensitive and 084-sensitive). Note that newly collected ISRIB-resistant pool didn't undergo any treatments prior being sorted serving as a control population.
Cell pools were then expanded, phenotypically characterized and their genomic DNA obtained for subsequent next-generation sequencing analysis.
Individual clones were genotyped following Sanger sequencing of PCR products amplified from genomic DNA.  (Table S6).

NGS sequencing of Eif2b2 containing targeted mutations
Genomic DNA was prepared from 10 7 cells using Blood and Cell culture Midi Kit  (Table S5).
The sequencing reads were converted to bam files aligned to the ceEIF2b2 locus and then corresponding sam files were aligned to 31 bp templates surrounding each mutation: N162 (NNNGAXGTGATCATGACCATTGGCTATTCT), H188 (CGAAAGAGGAAGTTCNNNGTCATTGTTGCC) and I190 (CGAAAGAGGAXGTTXCATGTCNNNGTTGCCS). Reads that differed by less than three residues outside the degenerate sequence were counted for codons and amino acids at each position using Python scripts.
CHOP::GFP reporter assay in CHO-K1 cells 40x10 3 CHO-K1 cells were plated in 6 well plates. Two days later the culture medium was replaced with 2 mL of fresh medium and compounds added. Immediately before 23-Nov-17 V1.7.3.1F_BioRxiv analysis, the cells were washed with PBS and collected in PBS containing 4 mM EDTA.
Single cell fluorescent signals (10,000/sample) measured by FACS Calibur (Beckton Dickinson). FlowJo software was used to analyze the data.

Electron microscopy
3 μL aliquots of the protein complex was applied on glow-discharged holey carbon grids (Quantifoil R2/2). Grids were blotted and flash-frozen in liquid ethane using a Vitrobot automat (Thermo Fisher). Data acquisition was performed under low-dose conditions on a Titan KRIOS microscope (Thermo Fisher) operated at 300 kV. The dataset was recorded on a back-thinned Falcon II detector (Thermo Fisher) at a calibrated magnification of x 80 000 (resulting in a pixel size of 1.75 Å on the object scale) with a defocus range of 2-3.5 μm. An in-house built system was used to intercept the videos from the detector at a speed of 25 frames for the 1.5 seconds exposures. Data were acquired automatically using the EPU software (Thermo Fisher) over one 24 hours session. Summarized information could be viewed in supplementary materials (Table S1).

Image processing
After visual inspection of the micrographs, 765 images were selected for further processing. The movie frames were aligned with MotionCorr  for whole-image motion correction. Contrast transfer function parameters for the micrographs were estimated using Gctf (Zhang, 2016). 237,486 particles were selected semi-automatically using the e2boxer routine from EMAN2 (Tang et al., 2007). All 2D and 3D classifications and refinements were performed using RELION (Scheres, 2012a, b).
As the S. pombe structure was not available at the time, first reference-free 2D classification on a sample of images, followed by a 3D classification step was done using as a reference the crystal structure of the tetrameric eIF2B (βδ) 2 complex from Chaetomium thermophilum (PDB: 5DBO, Kuhle et al., 2015) low-pass filtered to 60 Å, generated an initial model that confirmed the presence of all five subunits in our sample and also the existence of a C2 symmetry. 23-Nov-17 V1.7.3.1F_BioRxiv The whole particle dataset was submitted to 3D classification using the newly generated initial model as a reference and sorted into 8 classes. One selected class, representing 41,750 particles (~ 17 % of the dataset), was used for 3D refinement. To further increase the resolution statistical movie processing was performed as described (Bai et al., 2013). The reported overall resolution of 4.1 Å was calculated using the gold-standard Fourier shell correlation (FSC) 0.143 criterion (Scheres and Chen, 2012) and was corrected for the effects of a soft mask on the FSC curve using high-resolution noise substitution (Chen et al., 2013). The final density map was corrected for the modulation transfer function of the detector and sharpened by applying a negative B factor that was estimated using automated procedures (Rosenthal and Henderson, 2003).

Model building and refinement
S. pombe eIF2B structure (PDB accession code: 5B04, Kashiwagi et al., 2016) was used as a starting model. A poly-alanine (glycines were immediately added to account for flexibility) model was generated and subunits were individually fitted to the density map in Chimera (Pettersen et al., 2004). This model was then finely fitted using real space refinement and loops found to be divergent between S. pombe and human were rebuilt in Coot (Emsley et al., 2010). All side chains for which density was clearly resolved (up to 37% of non-Ala non-Gly residues in α and β subunits and as low as 8% in ε subunits) were positioned. The ISRIB molecule was then manually fitted in the density located at the regulatory core of eIF2B.
The model was then refined using phenix.real_space_refine (Adams et al., 2010) to optimize both protein and ligand geometry and limit clashes. Finally, REFMAC in CCP-EM (Burnley et al., 2017) was used to further refine the model and automatically generate the map vs. model FSC curves and validate overfitting (as described by Brown et al., 2015). Briefly, the procedure involved initial random displacement of the atoms within the final model and refinement against one of the two half maps to generate the FSC work curve. A cross-validated FSC free curve was then calculated between this refined model and the other half map. The similarity between FSC work and FSC free curves is indicative of the absence of overfitting. Summarized information could be viewed in supplementary materials (Table S1).
A solution of 8b (0.1 g, 0.18 mmol) in THF-H 2 O (2:1) was treated with LiOH.H 2 O (0.037 g, 0.9 mmol) and the resulting mixture was stirred for 3 hours. The reaction was neutralized with 1 M aq HCl and extracted with EtOAc. The organic phase was dried over Na 2 SO 4 , filtered, and evaporated to afford the carboxylic acid derivative 8b (0.09 g, 98%).

Author contributions
A.Z. Conceived and led the project; designed and conducted most experiments; analyzed and interpreted the data; prepared the figures and drafted the manuscript.

Mutations in Eif2b2 codon 188 (β H188X ) alter selectivity for ISRIB congeners
Graphs showing inhibition of the ISR-activated CHOP::GFP signal (induced upon treatment with HIS, 0.5 mM) by ISRIB and two related derivatives, AAA1-075B (75B) and AAA1-084 (84), in ISRIB SEN (A) and ISRIB RES (B) mutant pools of EIf2b2 H188X , revealed by flow cytometry. Shown is a representative from three independent experiments for each of the compounds (replicated in Fig. 4 -Sup. 1 A, B).
Concentration of inhibitor is represented on a log 10 scale. Curve fitting and EC 50 was generated using "agonist vs. response" function on GraphPad Prism.

Reproducibility of the effect of mutations in Eif2b2 codon 188 (β H188X ) on selectivity for ISRIB congeners
Graphs showing inhibition of the ISR-activated CHOP::GFP signal (induced upon treatment with HIS, 0.5 mM) by ISRIB and two related derivatives, AAA1-075B (75B) and AAA1-084 (84), in ISRIB SEN (A) and ISRIB RES (B) mutant pools of EIf2b2 H188X , revealed by flow cytometry. Shown is a representative from three independent experiments for each of the compounds (replicated in Fig. 4 A, B). Concentration of inhibitor is represented on a log 10 scale. Curve fitting and EC 50 was generated using "agonist vs. response" function on GraphPad Prism. Shown are mean ± SD (n=3). Concentrations of eIF2B are represented on a log 10 scale. The fitting curve and EC 50 was generated using "agonist vs. response" function on GraphPad Prism.      Allele frequency (supplement to figure 3D)

Table S3
Allele frequency (supplement to figure 5C)

Table S4
Plasmid list

Table S5
Oligo list Table S6 Cell line list