Introduction

Phosphorylation of eukaryotic translation initiation factor 2 alpha (eIF2α) mediates the response to a variety of stresses by regulating translation, metabolism and cell survival (1, 2). In yeast, GCN2 is the sole eIF2α kinase and triggers the General Control Response to the accumulation of uncharged tRNA and to ribosome stalling (3, 4). Gene duplication gave rise to a family of eIF2α kinases in multicellular organisms linking eIF2α phosphorylation to a variety of other stresses. HRI (EIF2AK1) is activated by haem deficiency, oxidative stress and mitochondrial dysfunction (5–7), PKR (EIF2AK2) responds to cytosolic double stranded RNA during viral infections, while PERK (EIF2AK3) is activated by endoplasmic reticulum (ER) stress (2). Mutations affecting these kinases have been linked to the development of disease and so their manipulation by small molecules is being pursued as a means to develop therapies (2, 8–11).

The eIF2 complex recruits initiator methionyl-tRNA to ribosomes commencing protein synthesis (2). This requires eIF2 to bind GTP, which is subsequently hydrolysed. Bound GTP is replenished by the guanine nucleotide exchange factor (GEF) eIF2B (12), but phosphorylated p-eIF2α binds avidly to eIF2B inhibiting its GEF activity (13). When levels of eIF2-GTP-tRNA-Met subsequently fall, most translation is disfavoured, but some transcripts are translated more efficiently owing to the presence of upstream open reading frames (uORFs) within their 5’ untranslated regions (5’UTRs) (1, 14). ATF4 and CHOP are transcription factors upregulated in this manner during stress to induce genes of the Integrated Stress Response (ISR) (1, 2, 14). Eventually, PPP1R15A (also called GADD34), itself an ISR target, dephosphorylates p-eIF2α to restore the basal state (15, 16).

GCN2 is comprised of five domains—RWD, pseudokinase, kinase, HisRS-like, and C-terminal—and exists constitutively as a homodimer (17, 18). Its kinase activity is stimulated when the RWD domain interacts with colliding ribosomes via the cofactor GCN1 (19–21) or when the HisRS domain interacts with deacylated tRNAs (22, 23). Direct interaction between GCN2 and the ribosome P-stalk can also mediate GCN2 activation during ribosome stalling (24, 25).

Mutations of EIF2AK4, which encodes GCN2, cause the aggressive subtypes of familial pulmonary hypertension, pulmonary veno-occlusive disease (PVOD) and pulmonary capillary haemangiomatosis (PCH) (2, 17). Some pathogenic alleles generate hypomorphic variants of GCN2 that we and others have shown can be activated by small molecule inhibitors that bind to the GCN2 ATP pocket, e.g. neratinib and GCN2iB (9, 11, 26). Type-1 inhibitors bind the ATP pocket of the kinase in its active conformation (DFG-Asp-in, αC-helix-in), while type 2 inhibitors bind the inactive conformation (DFG-out, αC-helix-out) (27). Type 1.5 inhibitors, by contrast, bind an intermediate conformation (DFG-Asp-in, αC-helix-out). Several type 1 and 1.5 kinase inhibitors can activate GCN2 at low concentrations while inhibiting at higher concentrations (9, 11, 26). GCN2iB is a relatively selective type 1.5 inhibitor of GCN2 that activates the kinase at submicromolar concentrations (9, 17). It is hypothesised that drug-induced stabilisation of one GCN2 protomer within a dimer is propagated to the second protomer to cause its activation (11, 28). Several inhibitors of other kinases that are in clinical use or development as anticancer agents activate GCN2 in a GCN1-dependent manner, which may contribute to their therapeutic efficacy (26, 29–31). For this reason, GCN2 activators are being sought as potential anticancer drugs.

The complexity of GCN2 activation, involving direct and indirect detection of amino acid deprivation and ribosome stalling, has hampered efforts to reconstitute GCN2 function in vitro (32–34). We therefore performed a cell-based chemical screen for GCN2-dependent activators of the ISR, followed by in vitro validation using orthogonal approaches. In doing so, we identified three novel, structurally distinct GCN2 activators. Notably, one of these, designated compound 20, activates GCN2 in a GCN1-independent manner.

Results

Chemical screen for GCN2 activators

We generated clonal Chinese Hamster Ovary (CHO) cell lines expressing an ATF4::NanoLuc reporter comprising the 5’ untranslated region (UTR) of human ATF4 mRNA fused to NanoLuc luciferase (Figure 1A) (17). Activation of the ISR could be detected as bioluminescence in response to ISR activators histidinol, tunicamycin and latrunculin A (Figure 1B). A high-throughput screen of the BioAscent library of 123,222 drug-like compounds performed by the ALBORADA Drug Discovery Institute, yielded two separate pools of potential ISR inhibitors and activators. Using this knowledge, we assembled a subset of 6,600 compounds from the BioAscent library enriched for potential ISR activators and screened these in 348-well format at 10µM for 16 hours. To increase sensitivity, screens were performed with test compound alone or combined with the half-maximal effective concentration (EC₅₀) of histidinol. This yielded 153 compounds that activated the reporter at least three standard deviations above baseline (Figure 1C).

Figure 1.

The pipeline used to filter hit compounds is illustrated in Figure 1D. Counter screens were performed against constitutively expressed Nanoluc, to exclude trivial activators of luminescent proteins, leaving 134 hits. Cytotoxicity assays excluded immediately toxic compounds leaving 130. Hits were then re-screened using a second CHO ATF4::Luc2 reporter cell line, shortlisting 89 lead compounds. GCN2-deficient ATF4::NanoLuc CHO reporter cells were generated by introducing the reporter into Eif2ak4-/- CHO cells (24). Thirty compounds were selected based on GCN2 dependency, where stimulation was lost in GCN2 null cells. Liquid chromatography-mass spectrometry (LC-MS) verified compound purity and quality leaving 16 lead compounds of high confidence from the original library stocks.

Orthogonal secondary screen

The ISR can be activated in isolation or in combination with other stress signalling pathways, for example as part of the unfolded protein response (UPR) to ER stress (2). To distinguish between inducers of ER stress and selective activators of the ISR, we used dual reporter CHO cells that express a fluorescent ISR reporter (CHOP::GFP) and a reporter of IRE1 activity (XBP1::Turquoise) (35). Histidinol served as a positive control for ISR activation via GCN2, while tunicamycin was used to induced ER stress and activate the UPR. Eif2ak4^-/- CHO cells were used to assess GCN2 dependence, while Ppp1r15a^-/- cells, which are impaired in the dephosphorylation of eIF2α, were used to amplify the effects of ISR activation. Histidinol selectively activated the CHOP::GFP ISR reporter in wildtype and Ppp1r15a^-/- cells, but not in GCN2 deleted cells, validating this system (Figure 2A). As expected, the ISR was exaggerated in the Ppp1r15a^-/- cells owing to their defective dephosphorylation of eIF2α. Tunicamycin activated both the CHOP::GFP and XBP1::Turquoise reporter in all three genotypes.

Figure 2.

The 16 lead compounds were next examined both in the presence and absence of a submaximal concentration of histidinol. Using this orthogonal approach, three compounds were found to be GCN2 selective activators of the ISR (Figure 2B&C). Those which induced ER stress or showed minimal GCN2 dependence are not shown. Reporter activation was measured by fluorescence activated cell sorting (FACS) after 18 hours of treatment with 30µM compound 18, 20µM compound 20, and 7.5µM compound 21. The three GCN2 selective activators included compound 18, N-(4-{[(furan-2-yl)methyl]sulfamoyl}phenyl)-3-methyl-1-phenyl-1H-thieno[2,3-c]pyrazole-5-carboxamide, containing a furan and two phenyl rings; compound 20, 4’-methyl-N∼2∼-2-pyrimidinyl-4,5’-bi-1,3-thiazole-2,2’-diamine, containing two aminothiazoles; and compound 21, 5-(4-bromophenyl)-2-(methylsulfonyl)pyrimidine (Figure 2B).

Larger quantities of compound 18 were then purchased from TimTec, USA (compound ST51216373); compound 20 was synthesised by ChemBridge, USA (compound 5935953); and compound 21 was purchased from Specs, Netherlands (compound AC-907/25005415). After validating their effects using the assays described above, we next confirmed activation of the ISR by these compounds at the transcriptional level. CHO cells were treated with 30µM of compound 18 or 10µM of compound 20 or 21 for 4 hours and expression of the ISR target genes Chop and Ppp1r15a was measured using qPCR (Figure 2D). All three induced ISR gene expression, with compound 20 eliciting the strongest response.

Cell line differences in GCN2 activator potencies

Neratinib is an ATP-competitive pan-HER kinase inhibitor used clinically in the treatment of breast cancer (36). Recent studies have shown that the sensitivity of glioblastoma cells to neratinib is dependent on GCN2, with GCN2 activation observed at submicromolar concentrations but inhibition at higher concentrations (26). We reproduced this biphasic activation of the ISR reporter in human cervical carcinoma HeLa cells and hamster CHO cells, recording EC₅₀ values for activation of 269nM (95% CI 218-322nM) and 690nM (95% CI 634-772nM) respectively (Figure 3A). Disruption of the EIF2AK4 gene by CRISPR/Cas9 (EIF2AK4^-/-) in HeLa cells abrogated ISR activation confirming the effect to be mediated by GCN2 (Figure 3A, middle panel).

Figure 3.

In CHO cells, compound 18 was the least potent with an EC₅₀ of greater than 30µM in CHO cells without obvious inhibition up to 100µM (Figure 3B). Notably, compound 18 failed to activate the ISR measurably in either HeLa cells or human embryonic kidney 293T cells (Figure 3C). Compound 20 activated the ISR in CHO cells in a biphasic manner with an EC₅₀ of 4.8µM (95% CI 4.0-5.6µM) (Figure 3B). It was active in 293T cells though with a lower potency, EC₅₀ of 17µM, but was inactive in HeLa cells (Figure 3C). Compound 21 also showed biphasic activation of the ISR in CHO cells with an EC₅₀ of 5.0µM (95% CI 3.9-6.5µM) (Figure 3B). Weak activation was observed in 293T cells, but none was seen in HeLa cells (Figure 3C).

GCN2iB, a potent inhibitor of GCN2, abolished histidinol-induced activation of the ATF4::NanoLuc ISR reporter in CHO cells confirming this to be GCN2 mediated (Figure 3D). Supporting a role for GCN2 in mediating the compounds’ effects, GCN2iB substantially reduced ISR reporter expression by each, though this effect was only partial at concentrations of compound 20 above 10µM (Figure 3D). This might indicate GCN2-independent effects or an ability to partially outcompete GCN2iB for binding at higher concentrations.

Compound 20 activates GCN2 in a GCN1-independent manner

GCN1 mediates the interaction between stalled ribosomes and GCN2 (21, 37). Histidinol is a competitive inhibitor of the histidyl-tRNA synthetase that depletes the cell of charged histidyl-tRNA, thus causing ribosome stalling and activation of the ISR in a GCN1-dependent manner (37–39). Using GCN1 deficient human 293T cells transiently expressing the ATF4::NanoLuc reporter, we confirmed that activation of the ISR by histidinol requires GCN1 (Figure 4A). Low concentrations of GCN2iB (5-70nM) had minimal effect on ISR reporter activation, while neratinib robustly triggered the ISR. The stimulatory effects of neratinib were significantly attenuated in GCN1 deficient cells consistent with the observation that GCN1 knockout renders astrocytoma cells resistant to neratinib (26). By contrast, activation of the ISR by compound 20, chosen for having the largest effect size in 293T cells, did not require GCN1 to activate the ISR reporter (Figure 4A). Indeed, loss of GCN1 modestly increased ISR activation at the highest concentration of compound 20 suggesting a mode of action independent of ribosome stalling. To gain confidence in this observation, we inactivated the Gcn1 gene in CHO cells by CRISPR/Cas9. Neratinib-induced activation of the ATF4::NanoLuc ISR reporter was lost in GCN1 deficient CHO cells; however, ISR activation by compound 20 was completely insensitive to loss of GCN1 (Figure 4B).

Figure 4.

No systematic assessment of GCN1 dependence has previously been performed for GCN2 activating drugs. We therefore proceeded to test a panel of ten additional kinase inhibitors that had been reported to activate GCN2 (9, 11, 26). Our panel included examples of multi-targeted tyrosine kinase inhibitors (dovitinib, sunitinib), RAF inhibitors (LY3009120, dabrafenib), inhibitors of the Wee1 kinase involved in G2 cell cycle checkpoint (WEE1-IN-4, AZD1775, Debio 0123), an inhibitor of the Heat Shock Factor 1 pathway (NXP800), and EGFR inhibitors (erlotinib and gefinitinib) (9, 11, 26, 31, 40–42). In all cases, a marked reduction in GCN2 activation by the drugs was observed in GCN1 deficient cells (Figure 4C). Most of these molecules failed to activate the ISR in the absence of GCN1 (sunitinib, NXP800, dovitinib, AZD1775, dabrafenib, gefinitinib, LY3009120), while for three drugs weak activation of the ISR was seen at higher concentrations (WEE1-IN-4, Debio 0123, erlotinib). None showed the dramatic GCN1 independence observed for compound 20 in unstressed cells.

ATP-competitive interaction with the GCN2 kinase domain

Existing activators of GCN2 have been shown to interact with the ATP binding pocket (9, 11, 26). We therefore used a bioluminescence resonance energy transfer (BRET)-based assay to examine the interaction of the three lead compounds with the kinase domain of GCN2 in cells. Briefly, the N-terminal portion of human GCN2 kinase domain fused to NanoLuc luciferase was expressed in either CHO or 293T cells. These were then treated with an ATP-binding pocked tracer. In this system, energy transfer is proportional to tracer binding to the NanoLuc-tagged kinase domain, and so interactions of other molecules that displace the tracer reduce the BRET ratio (Figure 5A) (43). In both CHO and 293T cells, GCN2iB efficiently suppressed the BRET ratio thus validating the system (Figure 5B). All three lead compounds significantly reduced the BRET ratio, though unlike GCN2iB, compounds 18 and 20 were noticeably more effective in CHO than in 293T cells. Although compound 21 suppressed BRET in both cell types, its effect was less dramatic in CHO cells compared to compounds 18 and 20.

Figure 5.

To examine further the interaction of compounds 18 and 20 with GCN2’s kinase domain, we generated in bacteria a recombinant protein comprising Glutathione S-transferases (GST) fused to the kinase domain of human GCN2 (residues 585 to 1005 of NCBI reference Sequence NP_001013725.2). GST was chosen both to facilitate protein purification, and to enhance dimerisation of the kinase domain in order to cause autophosphorylation and constitutive activation. When incubated in vitro with increasing concentrations of ATP and a fixed quantity of His6-tagged bacterially expressed eIF2α N-terminal domain (44), GST-GCN2 kinase domain phosphorylated its substrate on serine 51, as detected by immunoblotting (Figure 5C). Neratinib dose-dependently inhibited eIF2α phosphorylation (Figure 5D, lanes 1-5). The lack of activation by neratinib likely reflected the requirement for autoinhibitory domains of GCN2 to observe this effect. Therefore, in this system inhibition of eIF2α kinase activity was used as a surrogate for target engagement. Compounds 18 and 20 also inhibited eIF2α phosphorylation, while compound 21 appeared inactive in this assay (Figure 5D). High concentrations of ATP were unable to completely overcome the inhibition by compounds 18 or 20, suggesting a mode of action, at least partially, non-competitive in nature (Figure 5E&F).

To test the ability of compound 20 to activate full-length GCN2, we next generated full-length human GCN2 (uniprot ID: Q9P2K8) through baculoviral expression using an N-terminal twin StrepII tag. Despite exerting only a modest effect on GCN2 autophosphorylation at T899 within the activation loop, compound 20 robustly triggered the phosphorylation of full-length eIF2α to an even greater extent than purified tRNAs, a known endogenous stimulus of GCN2 activity (Figure 5G).

Interaction with other ISR kinases

The eIF2α kinases have a common evolutionary origin and consequently the kinase domains share some structural similarity (45). We therefore examined the effect of the compounds on all four eIF2α kinases in vitro. Lower concentrations of ATP (10µM for GCN2, 0.1µM for PERK, 0.15µM for HRI, 0.1µM for PKR) were used to accentuate antagonism at the ATP site. Under these conditions, in contrast to Figure 5G where 100µM ATP was used, activation of GCN2 is not observed with any of the compounds tested. At low ATP, full length, purified GCN2 was efficiently inhibited by GCN2iB with an IC₅₀ on 1.55nM.

None of the screen hits reached 50% inhibition at the highest concentration tested (3µM) though compound 20 and to a lesser extend 18, appear to cause weak inhibition (Figure S1A). No inhibition was observed with the test compounds again purified PERK (Figure S1B). The PERK inhibitor GSK2606414 served as a positive control with an expected IC₅₀ of 0.66nM. Weak inhibition of HRI by compounds 18 and 20 was observed but failed to reach 50% (Figure S1C). In this system, GCN2iB was found to inhibit purified HRI with an IC₅₀ of 2.76nM, similar to that observed against GCN2. By contrast, both compounds 18 and 20 inhibited PKR in vitro: compound 18 with an IC₅₀ of 4.50µM and compound 20 with an IC₅₀ of 3.75µM (Figure S1B-D). GCN2iB inhibited PKR with an IC₅₀ of 3.26µM. Compound 21 was inactive against each of the four eIF2α kinases in this assay.

In silico docking of compounds 18 and 20 to the GCN2 kinase domain and structure-activity relationship (SAR) analysis

We used the SiteMap module from Schrodinger Suite to identify binding sites based on drug interactions with nearby GCN2 residues. Published crystals structures of the GCN2 kinase domain (see methods) show the ATP binding pocket comprises a large solvent exposed region and a buried hydrophobic region that extends beyond the gatekeeper residue (Met802) and towards the IZC-helix (Figure 6). Compound 18 was predicted to occupy a similar pocket to that of the ligand in the PDB 6N3N crystal structure demonstrating a Type 1.5 binding configuration, characterised by the DFG loop being “in” and the IZC-helix being in the “out” conformation (10). The furan motif of compound 18 is predicted to be buried in the allosteric pocket around αC-helix, between gatekeeper Met802 and Leu640 (Figure 6D). The compound’s sulfone was predicted to interact with Phe867 of the conserved DFG-motif. In neratinib, a triad of hydrogen bonding was predicted between the 8-position of the isoquinoline core (Ar-H bond), the amine linker (H bond) and the 2-position of the fluorobenzene ring (Ar-H Bond) with Asp866, similar to the predicted interaction between the amide linkage of compound 18 and Asp866 (Figure 6B). Both neratinib and compound 18 were also predicted to protrude into the solvent exposed space (Figure 6B-E). However, the interaction of neratinib with Cys805 of the hinge region through the nitrile of isoquinoline core was not present in compound 18.

Figure 6.

Due to the metabolic liabilities of furans in medicinal chemistry programs, we sought to explore the importance of this motif for binding and activity using available structural analogues. Replacement of the sulfonamide furan motif with some benzene derivatives including phenolic (BCC0106222), 2-bromobenzene (BCC0113320), 3-acetylbezene (BCC0104214), preserved ISR activation while others (BCC0073242 and BCC0043997) resulted in a loss of activity (Supplementary Table 1). Incorporation of heterocyclic ring systems (BCC0073142, Z56783036) or modification of solvent exposed region (Z88478561, Z226209130, Z11741677) were similarly not tolerated (Supplementary Table 1). However, modifying the furan motif to benzene derivative 2-bromobenzene (BCC0113320) and heterocyclic ring (BCC0077119) preserved GCN2-specific ISR activation.

Unlike compound 18 and neratinib, compound 20 was not predicted to protrude into the allosteric pocket (Figure 6F&G). Analogues extending further towards the IZC-helix and protruding into the solvent exposed region were inactive (e.g., Z134826202 and Z118606822), suggesting that the small size of compound 20 may be advantageous, fitting into the ATP-binding pocket of GCN2 and facilitating stronger activation (Supplementary Table 1). Hydrogen bond interaction was predicted between the NH₂ of the compound 20 aminothiazole with the carbonyl of the hinge region Cys805. Additional interaction was predicted between the linker NH of compound 20 and backbone carbonyl of Asp866 in the DFG-motif. Such 2-aminopyrimidine rings are present in many kinase ligands (46). Rather than this interacting with hinge region residues, this moiety in compound 20 was predicted to form a pi-cation interaction with Lys619 and a hydrogen bond with DFG Asp866, as has been reported for other ligands of kinase ATP pockets (10). Replacement of this pyrimidine ring with 2-pyridine analogues (BCC0084738, BCC0079350 and BCC0083923) was well tolerated, likely due to the conserved hydrogen bonding interaction between the pyridine nitrogen and Asp866. Removal of a nitrogen from the ring system, such as BCC0057014, BCC0089519, BCC0084744, and BCC0084833 all resulted in a loss of activity, possibly due to their lack of interaction with Asp866 (Supplementary Table 1). Due to toxicity issues surrounding 3-aminothiazoles, modifications were made to replace the terminal thiazole ring system with phenyl derivatives (BCC0080152, BCC0075509, BCC0084830, BCC0057431 and BCC0080200).

However, these resulted in a decrease in activity towards GCN2, suggesting that there might be a steric component contributing to activity. Replacing the thiazole with a minimally steric affecting thiophene (BCC0079876) also led to activity loss, possibly due to the twisting of central thiazole ring that impaired hydrogen bonding between pyrimidine nitrogen and Asp866. We were therefore unable to improve upon compound 20’s activity but gained valuable SAR information for future development.

Discussion

The ISR plays a critical role for maintaining homeostasis under diverse stress conditions and its disruption results in a variety of disease states (2). The observation that some pathogenic variants of GCN2 are amenable to pharmacological activation led us to seek novel GCN2 activators. Our chemical screen identified three chemically distinct GCN2 activating molecules, one of which is, to our knowledge, the first to show independence from GCN1 for its activity. This may prove useful as a tool compound for the study of GCN2 in cells and using purified components since it overcomes at least one hurdle to kinase activation.

Compounds 18 and 20 demonstrate weak inhibition of the related kinase PKR. Previous structural analyses have shown the kinase domains of PKR and GCN2 to be highly conserved, particularly at the ATP-binding pocket, which likely contributes to this effect. Indeed, a previous publication reported that other small molecule inhibitors of GCN2, including a series based on a triazolo[4,5-d]pyrimidine scaffold, demonstrated cross-inhibition of PKR (47). Furthermore, the PKR inhibitor C16 has been shown to activate GCN2 in vitro (11). Intriguingly, viral infections or synthetic dsRNA known to activate PKR, have also been shown to trigger GCN2 in some circumstances, though if this relates to structural similarities of their kinase domains has not been studied (11, 48, 49).

A remarkable finding in the current study was the lack of dependence on GCN1 for compound 20 to activate GCN2 in cells. During amino acid starvation, ribosome-associated GCN1 facilitates binding of uncharged tRNAs to GCN2, thus facilitating GCN2 activation (21, 50, 51). GCN1 also functions as a disome detector, enabling GCN2 activation in response to ribosome collisions (37). In genetic screens identifying modifiers of GCN2 activation by a small molecule, such as the dual HER2/EGFR inhibitor neratinib and the WEE1 inhibitor AZD1775, GCN1 has repeatedly been found to be necessary (26, 30, 31). We were able to confirm this GCN1 dependence for neratinib in the present study. It was therefore surprising that GCN2 activation by compound 20 did not appear to require GCN1 in human or hamster cells. While GCN1 is essential for the activation of GCN2 by amino acid starvation in yeast (52, 53), in vitro studies have shown that GCN2 activation by the ribosome P-stalk or by deacylated tRNAs does not absolutely require GCN1 (24, 25). Yet we found compound 20 able to activate purified full-length human GCN2 to a far higher level than was possible with purified tRNAs. Our modelling suggests that unlike neratinib, compound 20 can fit within the ATP-pocket without needing to protrude into the solvent, nor extending deeply into the hydrophobic region beyond the gatekeeper Met802 residue towards the IZC-helix. Indeed, analogues extending in these directions were inactive, perhaps suggesting value to this molecule’s small size. This may prove useful as a tool compound for studies in which GCN1 is unavailable or non-functional.

There are limitations to our study. So far, our experiments have been limited to cultured cells, purified proteins, and in silico modelling. We have yet to test our compounds in an animal model. Although our pipeline excluded molecules showing toxicity, it is likely that prolonged activation of the ISR will have toxic consequences via targets including CHOP and PPP1R15A (16). Our most promising molecule, compound 20, contains a 2-aminothiazole. Although known to be a versatile scaffold, appearing in several clinically useful drugs including anti-inflammatory, anti-microbial and anticancer agents, 2-aminothiazole is sometimes considered a toxicophore owing to toxic consequences following its metabolism, and as a pan-assay interference compound (PAINS), owing to frequent false positives in high-throughput drug screens (54–56). Our use of orthogonal approaches to support its activity as a GCN2 activator, mitigate against it being a non-specific hit. Nevertheless, if this compound were to be developed further, attempts to ‘design out’ this feature by, for example, replacing the sulphur for an oxygen (thiazole to oxazole switch) could be attempted. Indeed, computational modelling reveals an adjacent NH which should form a stronger H-bond to oxygen of an oxazole, which may result in an improved potency to GCN2. Another potential limitation of our study is the use of existing analogues in SAR studies, the so-called ‘analogue by catalogue’ approach. Although less systematic than chemically synthesising further series de novo, it was cost effective and generated useful information on the tolerability of modifications of our lead scaffolds. Finally, our compounds were identified in a screen of CHO cells and appear to be more effective in these cells that in HeLa and 293T models. The absence of detectable activity in HeLa cells does not appear to be a species effect, since some activity was observed in human 293T cells and when using recombinant human GCN2 protein. It is likely therefore to reflect cell-type effects, since cancer cell lines are known to demonstrate dramatic differences in drug efficiency, even when using different strains of the same cell line (57). This should be considered carefully when using these compounds as tools or for further drug development.

In summary, we performed a chemical screen for novel activators of GCN2. We identified three chemically distinct scaffolds, and demonstrated one to be independent of the GCN2 cofactor GCN1.

Methods

Screen molecules

The initial screen used compounds sourced from the BioAscent library and all molecules dissolved in DMSO. Some follow ondin compounds for SAR were sourced from Enamine. Subsequently, compound 18 was purchased from TimTec, USA (compound ST51216373). Compound 20 was synthesised by ChemBridge, USA (compound 5935953). Compound 21 was ordered from Specs, Netherlands (compound AC-907/25005415).

NanoGlo® luciferase assay

pGL4.2 ATF4-Nluc and pGL4.2 ATF4-Nluc-PEST® were described previously (17). In 384-well plate, test compounds, 3mM histidinol or vehicle control were plated as 5x solution in a volume of 5μL in 10% FBS-Nutrient Mixture F-12 Ham supplemented with 1x GlutaMAX™. CHO wildtype and Eif2ak4-/- ATF4::NanoLuc reporter cells were subsequently seeded at a density of 1.5×10³ cells per well in 20μL. After 16 hours at 37°C, NanoGlo® luciferase assay reagent was added to each well and mixed with orbital shaking for 1 minute. After 10 minutes of incubation at room temperature, bioluminescence signal (360-545 nm) was acquired on a Tecan Spark plate reader. Cells carrying a ATF4::nanoLuc-PEST reporter were used for live-cell assay, allowing detection of time-sensitive ATF4 signals (17). Cells (1.25×10⁴ cells per well) were plated overnight on white, clear bottom 96-well plate. Next day, 1μL 100x Nano-Glo® Endurazine™ substrate was mixed with cell culture medium and added to cells and incubated for 2 hours to reach stabilisation. Compounds were added to the wells and luminescence signals were measured using a Tecan CM Spark plate reader at 37°C in a humidity cassette for 48 hours. Data analysed using Graphpad Prism 10.

Bioluminescence Resonance Energy Transfer (BRET)

NanoBRET assay was performed according to user’s manual. Briefly, 4×10⁵ cells/mL (HEK293T or CHO) were transfected with 100ng of GCN2 kinase domain::nanoLuc reporter and 900ng of transfection carrier DNA mixed in 100μL Opti-mem (no phenol red) using lipofectamine LTX, and plated in 384-well white plate (40μL/well, 8×10³ cells). Twenty-two hours after transfection, tracer was added to each well and mixed on an orbital shaker. Test compound (final concentration 10, 30, 100μM), DMSO control (final concentration 1, 3, 10%), and positive control GCN2iB (final concentration 1μM) were added to each well and mixed. After 2 hours of incubation, NanoBRET™ Nano-Glo® substrate-extracellular NanoLuc® Inhibitor reagent was added to each well before bioluminescence measurement using a Tecan CM Spark plate reader (multi-colour measurement mode: wavelength 1, wavelength 445-470nm, 300ms integration; wavelength 2, wavelength 610-635nm, 300ms integration). Data analysed using Graphpad Prism 10.

Cell Viability Assay

Compounds were plated into 384-well white plates and mixed with 20μL of culture medium containing 1.5×10³ cells per well, incubated for 16 hours at 37°C. CellTiter-Glo® Assay reagent was added to each well for bioluminescence measurement using a Tecan CM Spark plate reader. Data analysed using Graphpad Prism 10.

Flow cytometry

CHOP::GFP and XBP1::Turquoise dual-reporter cells were treated with each compound (7.5-20μM) and histidinol (0.5mM) 18-20 hours prior to cell harvesting (35). ER stress was induced using 2μg/mL of tunicamycin. At the time of experiment, cells were trypsinised and collected by centrifugation, resuspended in 4mM EDTA PBS, filtered through a 50μm filter, then analysed using a DB Fortessa flow cytometer. GFP was measured at 530nm and Turquoise was measured at 450nm. Data analysed using FlowJo software.

Quantitative real time PCR (qPCR)

RNA was reverse transcribed to produce complementary DNA using High-Capacity cDNA Reverse Transcription Kit (Thermo, 4368814) according to the manufacturer’s instructions. To perform realIZtime PCR, the PowerUp SYBR Green Master Mix (Thermo A25742) and the CFX96 BioRad realIZtime PCR machine (Biorad, UK) were used. All samples were measured in duplicate and included water control. Fold changes in expression were calculated using the ΔΔCt method. Hamster qPCR primers include CHOP (Fw: GGGAGCTGGAAGCCTGGTAT, Rv: GGGACCCCCATTTTCATCTG), PPP1R15A (Fw: CCTGGTCTGCAAAGTGCTGAT, Rv: CCAGCTCAGTCACTCCCTCTTC),and β-actin (Fw: ACTCCTACGTGGGTGACGAG, Rv: AGGTGTGGTGCCAGATCTTC). Data analysed using Graphpad Prism 10.

Immunoblotting

Cells were washed with PBS on ice and lysed in low-salt buffer as previously described (17). Following SDS-PAGE (7% gel for GCN2 and 14% gel for eIF2α or alternatively, 5-12% gradient gel), proteins were transferred to nitrocellulose membranes then analysed by immunoblotting. Antibodies used: total GCN2 (Gift from Ron lab: NY168); total GCN2 (1:1000, Invitrogen JA03-83); GCN2-phosphorT899 (1:1000, Abcam, ab75836); β-actin (1:5000, Abcam, ab8226); eIF2α phos-S51 (1:1000, Abcam, ab32157); eIF2α phos-S51 (1:500, Cell Signalling Technology 9721); total eIF2α (1:1000, Invitrogen, AHO0802); total eIF2α (Gift from David Ron lab, CIMR); total eIF2α (1:1000, Santa Cruz Biotechnology sc-133132).

Recombinant GCN2 kinase domain synthesis and in vitro kinase assay

The human GCN2 kinase domain (residues 585 to 1005) was tagged on its N-terminus with GST by cloning into the vector pGV67 by Gibson assembly. Protein was synthesised in BL21(DE3) E. coli Rosetta cells and purified using glutathione agarose beads. GCN2 kinase domain (0.3μg) was incubated with 100μM ATP, 1μM eIF2α-NTD and test compounds in reaction buffer (50mM HEPES, pH 7.4, 100mM potassium acetate, 5mM magnesium acetate, 250μg/ml BSA, 10mM magnesium chloride, 5mM DTT, 5 mM β-glycerophosphate). Reactions were incubated at 32°C for 10 minutes, then immediately quenched with SDS sample buffer pre-warmed to 95°C. Samples were then analysed by immunoblotting.

Recombinant full-length GCN2 synthesis and in vitro kinase assay

Expression and purification of recombinant human GCN2 was conducted as described previously (25). Briefly, human GCN2 (uniprot ID: Q9P2K8) was cloned into a baculoviral vector with an N-terminal twin StrepII tag followed by a TEV protease site. GCN2 was then expressed in Sf9 cells grown at 27 °C for 55 hours. Protein was purified using sequential StrepTrap HP column (GE Healthcare Life Sciences 28-9075-47), HiTrap Q HP column (Cytiva 17115401), Hi Load 16/60 Superdex 200 column (Cytiva/GE Healthcare 28989335). Ten nanomolar GCN2 was incubated at 30 °C for 20 min with 2µM eIF2α and 100µM ATP in increasing concentrations (0.00169-100µM) of deacylated tRNA or compound 20. Samples were then quenched via the addition of SDS-PAGE loading buffer and heating to 95 °C for 3 min, then analysed by SDS-PAGE and immunoblotting: total eIF2α (Santa Cruz Biotechnology sc-133132) 1:1000, phospho-eIF2S1 Ser51 (Cell Signalling Technology 9721) 1:500, total GCN2 (Invitrogen JA03-83) 1:1000, pThr899 GCN2 (Abcam AB75836) 1:1000. Images were captured and quantified in Image Studio Ver 5.2. Data analysed using Graphpad Prism 10.

Expression and purification of eIF2α

Expression and purification of recombinant human eIF2α was conducted as described previously (25). DNA encoding full-length human eIF2α (NCBI reference number: NP_004085.1) was inserted into the vector pOPTH with an N-terminal His6 tag followed by a TEV protease site. Protein was expressed in BL21 Star (DE3) cells, at 37 °C for 3 hours then purified using a HisTrap HP Column (Cytiva 17524801) as described for GCN2.

LanthaScreen in vitro kinase activity assay (Figure S1)

In 384-well plate format, 1nM human GCN2 (Carna Bioscience, # 05-153) or 1nM human PERK (Carna Bioscience, # 05-155) or 2nM human HRI (Carna Bioscience, # 05-154) or 3nM human PKR (Carna Bioscience, # 05-156) were incubated with 50mM HEPES pH 7.0, MgCl₂ 10mM, MnCl₂ 5mM, tRNA 4µM (only for GCN2 assays), ATP (10µM for GCN2, 0.1µM for PERK, 0.15µM for HRI, 0.1µM for PKR), GFP-eIF2α 80nM. Reaction time was 20-30 minutes at room temperature, followed by detection time 50-60 minutes. TR-FRET signal between terbium-labeled antibody and GFP-labeled p-eIF2α read with excitation 340nm and emission of the donor (terbium labeled anti-p-eIF2α antibody: 495 nm) and the acceptor (GFP-labeled p-eIF2α: 520nm) on the Tecan Spark plate reader. IC₅₀ values were determined in GraphPad Prism 8.0, using a 4-parameter model: log(inhibitor) vs. response – variable slope.

In silico modelling

Five reported crystal structures for human GCN2 were considered (PDB 63N3, 6N3O, 6N3L, 7QWK, and 7QQ6 (10, 58). Of these, 6N3N and 6N3L share ∼100% sequence homology. While 6N3N has unresolved amino acids residues, these are remote from the ATP binding site and should not impact the binding of the ligands. A visual inspection of 6N3O, 6N3N, and 6N3L highlighted that the activation loop in 6N3L is cut short and does not extend beyond the IZC-helix, affecting the orientation of the DFG loop and its residues (Asp866, Phe867, Gly868). As such, 6N3L was not considered for further analysis. PDB sequences 7QWK and 7QQ6 share ∼75% sequence homology, with 7QWK containing unresolved residues remote from the ATP binding site. As observed with PDB 6N3L, the activation loop containing the DFG residues is cut short in both cases and does not extend beyond the IZC-helix. Inspection of the binding mode of crystallised ligands of 7QQ6 and 7QWK demonstrate that these bind in a Type I binding pocket, compared to Type II binding as observed in 6N3O.

Prior to docking, the protein structure (PDB ID: 6N3N) was put through the Protein Preparation Protocol of Maestro. This protocol added hydrogens, removed co-crystallised water molecules, optimised the receptor’s hydrogen bond network, enumerated bond orders, and performed a restrained minimisation to alleviate backbone clashes with the receptor backbone. The protein was prepared at pH7.4 (59). SiteMap module was used on the prepared protein, with all settings set to default (SiteMap, Schrödinger Release 2024-4, Schrödinger, LLC, New York, http://www.schrodinger.com/). Grid generation was performed using the Glide Docking Protocol in Masestro (Glide, Schrödinger Release 2024-4, Schrödinger, LLC, New York, http://www.schrodinger.com/). This was centred around Met802 and Lys619. High throughput virtual screening (HTVS) setting was used for docking.

Acknowledgements

SJM was supported by the MRC (G1002610, MR/V028669/1 and MR/R009120/1), EPSRC (EP/R03558X/1), Cambridge Biomedical Research Centre (BRC-1215-20014); British Lung Foundation, Asthma+Lung UK, Royal Papworth Hospital, Alpha1 Foundation. GE was supported by philanthropic funding from Rick Medlock and by awards from EPSRC (EP/R03558X/1), BHF (RE/18/1/34212), Evelyn Trust (Grant 22/03), and the Cambridge-Tsinghua collaborative programme on sustainability and emerging technologies. JZ was supported by the Doctoral Training Programme in Medical Research (DTP-MR; Cambridge Trust). GRM is supported by a Baxter Fellowship from the University of Dundee, and a TENOVUS grant T21/01 119615, and Biotechnology and Biological Sciences Research Council (UKRI-BBSRC) Capital Equipment Grant BB/V019635/1. VV supported by MRC iCase Studentship (MR/R01579/1). This research was supported by the CIMR Flow Cytometry Core Facility. We wish to thank Reiner Schulte and Gabriela Grondys-Kotarba for their advice and support in flow cytometry cell sorting. We wish to thank Heather Harding for helpful discussions and reagents. We wish to thank Steve Andrews, Kushal Rugjee, and other colleagues at the ADDI for help with chemical library preparation. Thanks to Apollo Therapeutics Ltd for conducting the LanthaScreen in vitro kinase activity at Selvita S.A, Krakow, Poland.

Additional files

Supplementary Figure 1. Compound activity across eIF2α kinases using LanthaScreen in vitro kinase activity assay Left panels: dose-dependent inhibition of GCN2, PERK, HRI or PKR with positive control drug GCN2iB or GSK2606414. Right panels: effect of concentration ranges of compounds 18, 20, and 21 in this system revealed weak inhibition of GCN2, HRI, and PKR. Low ATP concentrations were used to accentuate inhibition, and this masks the stimulatory effects on GCN2.