GCN2 kinase Activation by ATP-competitive Kinase Inhibitors

Abstract

Small molecule kinase inhibitors represent a major group of cancer therapeutics, but tumor responses are often incomplete. To identify pathways that modulate kinase inhibitor response, we conducted a genome-wide knock-out screen in glioblastoma cells treated with the pan-ErbB inhibitor neratinib. Loss of General Control Nonderepressible 2 (GCN2) kinase rendered cells resistant to neratinib whereas depletion of the GADD34 phosphatase increased neratinib sensitivity. Loss of GCN2 conferred neratinib resistance by preventing binding and activation of GCN2 by neratinib. Several other FDA-approved inhibitors, such erlotinib and sunitinib, also bound and activated GCN2. Our results highlight the utility of genome-wide functional screens to uncover novel mechanisms of drug action and documents the role of the integrated stress response in modulating the response to inhibitors of oncogenic kinases.

Graphical Abstract. Model of GCN2 activation by neratinib.

In the absence of drug, EGFR is on and GCN2 is off, leading to cell survival and cell growth from EGFR signaling. In the presence of neratinib, EGFR is off and GCN2 is on, leading to inhibited cell growth from loss of EGFR signaling and increased cell death from ISR activation by direct GCN2 activation through neratinib binding.

Introduction

The inhibition of oncogenic kinases represents an important component of cancer therapy for many human cancers ¹. Allosteric kinase inhibitors prevent activity by binding to the kinase at a regulatory domain that is distinct from its active site. Conversely, ATP-competitive kinase inhibitors directly bind to the kinase in the active site, blocking ATP and preventing phosphorylation of its substrates. Irreversible ATP-competitive inhibitors permanently occupy the active site by forming covalent bonds with residues (i.e., cysteine) in the ATP-binding pocket ². Type I kinase inhibitors bind to the active conformation of a kinase, while type II inhibitors bind to the inactive conformation of a kinase ². ATP-competitive inhibitors can exert effects on their target kinase beyond displacing ATP from the kinase active site, for example alter protein localization or protein-protein interactions ^1–3.

Glioblastoma, the most common malignant brain tumor in adults, often harbor gain-of-function alterations in the EGFR gene, including gene amplifications and mutations in the extracellular domain of the receptor ^4–6. Clinical trials with first-generation EGFR kinase inhibitors have shown only partial and transient tumor responses in subsets of GBM patients. Mechanisms of resistance remain poorly understood and may include inadequate drug penetration into tumor tissue, the presence of redundant signaling pathways that can rescue tumor cells from the effects of EGFR kinase inhibition, and many others ^7,8.

To identify signaling or metabolic pathways that modulate response to EGFR kinase inhibitors in EGFR mutant GBM and could provide a foundation for combination therapy strategies, we performed a genome-wide clustered regularly interspaced short palindromic repeats (CRISPR) knock-out screen in EGFR mutant GBM cells treated with the irreversible pan-ErbB kinase inhibitor neratinib ⁹. We selected neratinib for this screen because of the increased sensitivity of cancer-associated extracellular EGFR mutants to type II EGFR kinase inhibitors ^10–12.

Results

GCN2 loss confers resistance to neratinib in GBM cells.

We performed genome-wide CRISPR/Cas9-knockout experiments with the single vector TKOv3 CRISPR library ¹³ to screen for genes that permit the growth of EGFR-mutant GBM cells in the presence of the EGFR inhibitor neratinib (Extended Data Fig. 1a). Using an EGFR mutant GBM cell line, SF268 (EGFR, A289V), we found that neratinib treatment resulted in enrichment of single guide RNAs (sgRNA) targeting the amino acid starvation sensor GCN2, encoded by EIF2AK4, compared to vehicle (Fig. 1a, b, Extended Data Fig. 1b).

Fig. 1. Loss of the amino acid starvation sensor GCN2 confers resistance to neratinib.

a, Representative graph of a CRISPR screen showing the second highest sgRNA (per gene) frequencies of SF268 cell line treated with 1.5 μM neratinib versus vehicle for 21 days. sgRNA frequencies are represented by log₂ normalized sgRNA counts. Effect is the fold difference between the normalized neratinib and vehicle frequencies. The top 5 sgRNAs with the highest effect are labeled. b, Bar graph showing the average frequency of 4 distinct GCN2 targeting sgRNAs from two experiments (Exp #1, Exp #2) at either day 0 or vehicle/neratinib treated. Significance determined with two sided student t-test. c, Effect of GCN2-KO on neratinib treatment measured by SF268 cell growth and assessed by trypan blue exclusion after 6 days. Immunoblot confirming GCN2-KO on right side. Significance was determined with a two sided student t-test. d, Effect of GCN2-KO or control (EGFP) on neratinib sensitivity in TS895 cells. e, Survival curve for in vivo experiment. Significance values calculated with Mantel-Cox test by comparing the neratinib treated and vehicle treated curves. e,f, Data are presented as mean values ± SD of n=3 biologically independent samples.

To validate the results of our CRISPR/Cas9 library screen, we generated single cell GCN2 knock out (KO) clones in two EGFR mutant GBM cell lines, SF268 and SKMG3 (EGFR A289D) using a two vector CRISPR/Cas9 system ¹⁴. We observed that both GCN2-KO GBM cell lines were resistant to neratinib compared to the parental cell lines (Fig. 1c, Extended Data Fig. 1c). We also examined the contribution of GCN2 to neratinib sensitivity in a patient-derived GBM tumor sphere line (TS895). Loss of GCN2 rendered TS895 cells resistant to neratinib relative to TS895 cells transduced with a control sgRNA, namely a sgRNA targeting enhanced green fluorescent protein (EGFP), in vitro and in vivo (Fig. 1d, e).

Neratinib activates the GCN2/ATF4 signaling pathway.

The GCN2 serine/threonine kinase senses amino acid starvation through binding to uncharged transfer RNAs (tRNA)^15–17. GCN2 subsequently upregulates activating transcription factor 4 (ATF4) and activates the integrated stress response (ISR)^18,19 (Fig. 2a). In our initial analysis, we had noted that loss of GCN1L1, which assists in the loading of uncharged tRNAs onto GCN2 ^20,21, was also associated with neratinib resistance (Fig. 1a). To determine whether other members of the ISR might also alter neratinib sensitivity, we performed additional CRISPR/Cas9 screens with neratinib in EGFR mutant SKMG3 GBM cells and then analyzed the results of the multiple independent screens in SKMG3 and SF268 GBM. GBM cells growing in the presence of neratinib showed significant enrichment of sgRNAs targeting ATF4, the downstream effector of GCN2 kinase signaling (Fig. 2b, c, Extended Data Fig. 1d,e). We confirmed that deletion of ATF4 was sufficient to confer neratinib resistance in GBM cells (Fig. 2d, Extended Data Fig. 1f).

Fig. 2. Sustained ATF4 expression induces cell death.

a, Schematic of the ISR signaling pathway. b, Gene candidates identified from two neratinib SF268 CRISPR screen experiments (shown in Fig. 1b,c). sgRNAs are sorted by normalized z-score. c, Gene candidates identified from two CRISPR screen experiments with neratinib in SKMG3 GBM cells after 14 days. d, Growth curve of parental and ATF4-KO cells lines treated with neratinib after 6 days. e, Immunoblot of ISR pathway members in SF268 cells treated with 1 μM neratinib for 4 hours. f, Induction of amino acid transporters in SF268 parental and GCN2-KO cells treated with 675 nM neratinib for the indicated times. g, Growth curve comparing neratinib sensitivity of TS895 GADD34-KO to EGFP control after 6 days. h, Immunoblot of SF268 Parental, GADD34-KO, and GADD34-KO rescued with CRISPR resistant GADD34* after 4 hour 675 nM neratinib treatment. i, SF268 growth curves of GADD34-KO and GADD34-KO + GADD34* after 3 day neratinib treatment. j, Immunoblot of ISR pathway members in H3255 cells treated with 1 μM neratinib for 4 hours. d,g, Data are presented as mean values ± SD of n=3 biologically independent samples.

Since our results suggested that the GCN2/ATF4 signaling pathway contributed to growth inhibition by neratinib, we sought to determine whether neratinib influenced the activity of this pathway directly or indirectly. Immunoblot of whole cell lysates from neratinib treated GBM cells indeed showed increased phosphorylation of GCN2 at Threonine 899, an established biochemical readout for GCN2 kinase activation, and upregulation of ATF4 protein levels (Fig. 2e). Experiments in GCN2-KO cells showed that ATF4 induction by neratinib required GCN2 (Extended Data Fig. 2a). Time course experiments demonstrated that neratinib induced GCN2 autophosphorylation and ATF4 expression within one hour, followed by the expression of the ATF4 regulated amino acid transporters xCT and ASCT2 expression (Fig. 2f, Extended Data Fig. 2b).

In contrast to our findings with GCN2, we observed that loss of GADD34 (encoded by PPP1R15A), a negative regulator of the ISR ²² (Fig. 2a), was associated with increased ATF4 expression and increased sensitivity to neratinib in our genome-wide screen (Fig. 2b, c, Extended Data Fig. 1e). To confirm that neratinib treatment was more deleterious to cells lacking the GADD34 phosphatase, we generated GADD34-KO cells and compared the neratinib sensitivity of these cells to SF268 GADD34-KO cells reconstituted with a CRISPR resistant GADD34 allele. Consistent with the findings of our screen, GADD34 deleted GBM cells showed greater growth inhibition and induction of Poly (ADP-ribose) polymerase (PARP) cleavage, a biochemical readout for apoptosis, in response to neratinib (Fig. 2g, h). Reconstitution of GADD34 in GADD34-KO cells abrogated induction of PARP cleavage by neratinib and conferred neratinib resistance (Fig. 2h, i).

We hypothesized that GADD34 inactivation enhanced neratinib sensitivity by increasing phosphorylation of eukaryotic initiation factor alpha (eIF2α). To confirm that increased eIF2α phosphorylation is deleterious for EGFR mutant GBM cells, we generated SF268 cells expressing a constitutively active eIF2α (S51D) allele under the control of a doxycycline (dox) inducible promoter. Treatment of these cells with doxycycline induced expression of ATF4 (Extended Data Fig. 3a) and markedly impaired growth (Extended Data Fig. 3b). We also determined that GCN2 induction by neratinib was associated with increased cell death in in the EGFR-mutant (L858R) lung cancer cells (Fig. 2j).

While GCN2/ATF4 signaling can promote apoptosis, this usually occurs in response to prolonged hyperactivation of the ISR. We therefore examined the durability of ATF4 induction by neratinib through the measurement of ATF4 regulated genes, including several metabolite transporters (Extended Data Fig. 4a, b). Gene set enrichment analysis showed that the amino acid deprivation response gene set was one of the most significantly enriched gene sets at both 6 hours and 72 hours following treatment with neratinib (Extended Data Fig. 2c, d). Gene set enrichment analysis of RNA from orthotopic TS895 GBM tumors treated with neratinib also revealed induction of the amino acid deprivation pathway (Extended Data Fig. 2e). Corresponding experiments in GCN2-KO cells showed that induction of ATF4 and ATF4-regulated genes by neratinib was GCN2-dependent (Extended Data Fig. 4f). Of note, unlike the ER stress inducer tunicamycin ²³, neratinib did not induce an ER stress response as determined by the lack of induction of ER stress pathway members, PERK phosphorylation, ATF6 cleavage, or induction of CHOP (Extended Data Fig. 4f, g).

GCN2 induction by neratinib is uncoupled from EGFR inhibition.

We next performed a series of experiments to dissect the relationship between EGFR kinase inhibition and GCN2 induction. We first determined the dose-response relationship between EGFR inhibition and induction of GCN2/ATF4 by neratinib. While we observed near complete EGFR inhibition by neratinib at 25 nM, activation of GCN2/ATF4 occurred at higher neratinib concentrations (75–225 nM) (Fig. 3a). We then tested whether inhibition of canonical EGFR effector molecules, alone or in combination, would result in similar activation of the GCN2/ATF4 signaling axis as treatment with neratinib. Treatment of GBM cells with MK2206 and trametinib, small molecule inhibitors of the serine-threonine kinase AKT and MEK, respectively, did not induce phosphorylation of GCN2 or induction of ATF4 (Fig. 3b). We observed that levels of EGFR expression in a panel of cell lines appeared unrelated to the ability of neratinib to induce ATF4 expression in these cells (Fig. 3c). Lastly, we found that ectopic expression of neratinib-resistant alleles of EGFR (C797S) and HER2 (C805S) rendered cells resistant to EGFR inhibition, but had no effect on the induction of ATF4 protein (Fig. 3d) or ATF4 transcript levels by neratinib (Fig. 3e). Taken together, these data suggest that EGFR inhibition and induction of the GCN2/ATF4 signaling by neratinib are uncoupled.

Fig. 3. Induction of GCN2 by neratinib is uncoupled from EGFR inhibition.

a, SF268 parental cells treated with increasing doses of neratinib for 4 hours. b, GCN2/ATF4 signaling in SF268 cells treated with the indicated combinations of the MEK inhibitor trametinib (1μM), the allosteric AKT inhibitor MK2206 (1μM), or neratinib (1 μM) for 4 hours. c, Immunoblot of EGFR null (CHO-K1, NR6), EGFR low (NHA), mutant EGFR (SF268, SKMG3) cell lines treated with 1 μM neratinib for 4 hours. d, ATF4 induction in TS895 cells with wild-type EGFR/HER2 or neratinib resistant EGFR C797S/HER2 C805S alleles treated with indicated doses of neratinib for 4 hours. e, Relative ATF4 mRNA levels measured by qRT-PCR in the indicated TS895 cell lines treated with neratinib for 1 day normalized to vehicle.

Neratinib directly binds to and activates GCN2.

We next explored the possibility that induction of the GCN2/ATF4 pathway by neratinib was due to decreased amino acid abundance. We measured intracellular amino acid abundance in the presence and absence of neratinib using gas-chromatography mass spectrometry (GCMS). Cells starved of glutamine or leucine exhibited the expected changes in metabolite abundance, but we did not observe significant changes in metabolite abundance following neratinib treatment (Fig. 4a). These results suggest that the acute activation of GCN2 signaling by neratinib is not driven by metabolite starvation.

Fig. 4. Neratinib directly activates GCN2.

a, Intracellular amino acid levels measured by GCMS from SF268 cells treated with 675 nM neratinib, glutamine deprivation, or leucine deprivation. Time point and replicate are indicated to the right. b, Upper, schematic of drug binding ATP competition assay. Lower, immunoblot of avidin capture (ATP bound) and eluate (neratinib bound) protein from the neratinib-GCN2 ATP competition assay. c, Activity assay showing ³²P labeling of recombinant GCN2 by autophosphorylation and eIF2α by GCN2 kinase activity with [γ−³²P]ATP. d, Activity assay showing autophosphorylation of purified GCN2 at Threonine 899 in the presence of neratinib.

Our observation that GCN2 activation by neratinib occurred in the absence of amino acid starvation and was uncoupled from EGFR inhibition raised the question of whether neratinib might directly bind to and activate the GCN2 kinase. Electroporation of recombinant GCN2 into GCN2-KO cells restored neratinib induced phosphorylation of eIF2α at serine 51 (Extended Data Fig. 5), demonstrating that the GCN2 protein was sufficient to mediate neratinib-induced eIF2α phosphorylation.

To determine whether neratinib directly binds to the GCN2 kinase domain, we performed an ATP-displacement assay. We incubated recombinant GCN2 with uncharged tRNAs, neratinib, and a desthiobiotin-ATP probe, which biotinylates GCN2 when ATP has access to the kinase domain and then performed avidin capture of biotinylated GCN2. Consistent with our model, neratinib reduced biotinylation of GCN2 in a dose-dependent manner and increased elution of unbiotinylated GCN2 (Fig. 4b).

To determine the effect of neratinib on GCN2 kinase activity, we performed an in vitro [γ³²P]-ATP labeling assay and observed that neratinib dose-dependently enhanced GCN2 autophosphorylation and phosphorylation of the GCN2 substrate eIF2α in the presence of uncharged tRNAs (Fig. 4c). To confirm that neratinib induced GCN2 phosphorylation at the biologically relevant phosphorylation site (threonine 899, T899), we repeated the GCN2 activity assay with unlabeled ATP and a T899 P-GCN2 specific antibody (Fig. 4d).

Multiple ATP-competitive kinase inhibitors activate GCN2.

A query of the NIH Library of Integrated Network-based Cellular Signatures (LINCS) KINOMEscan data ²⁴ showed that neratinib binds to the GCN2 kinase domain with a dissociation constant (Kd) of ~100nM (Fig. 5a, Supplementary Table 1), lower than the reported plasma C_max levels in a neratinib clinical trial ²⁵. It also showed binding of several other kinase inhibitors to the GCN2 kinase domain. To determine whether the reported binding of kinase inhibitors to the GCN2 kinase domain activates the ISR, we measured the effects of a panel of EGFR inhibitors on GCN2/ATF4 signal activation. Canertinib, another irreversible ATP-competitive inhibitor of ErbB receptor family members, was reported not to bind to the GCN2 kinase domain (Fig. 5a) and did not induce ATF4 in SKMG3 cells (Fig. 5b). Among first-generation, reversible ATP-competitive EGFR inhibitors, erlotinib was reported to bind GCN2 at lower Kd than gefitinib (Fig. 5c). Treatment of SF268 cells with erlotinib resulted in robust induction of GCN2 autophosphorylation and ATF4 induction, whereas induction of GCN2/ATF4 signaling by gefitinib occurred at higher drug concentrations (Fig. 5c).

Fig. 5. Several ATP-competitive kinase inhibitors bind to and activate GCN2.

a, Bar graph of GCN2 (R585-T1018, kinase domain) Kds for selected inhibitors reported in the LINCS KINOMEscan dataset. Previously reported C_max for drug plasma levels are represented as red points. b, Immunoblot of SKMG3 cells after 4 hour neratinib or canertinib treatment. c, Immunoblot of SF268 cells treated with the reversible EGFR inhibitors erlotinib or gefitinib for 4 hours. d, Superposition of aligned EGFR/Neratinib (red) and GCN2/dovitinib (green) kinase domain co-crystal structures with ATP (pink mesh). e, Immunoblot of SF268 cells treated with the FLT3 inhibitor dovitinib for 4 hours.

We next determined whether ATP-competitive inhibitors of other kinases, which were reported to bind to the GCN2 kinase domain in LINCS KINOMEscan, also activated the GCN2/ATF4 pathway. Sunitinib which is widely used for the treatment of KIT-mutated cancers ²⁶ was reported to bind the GCN2 kinase domain (Fig. 5a) and induced phosphorylation of eIF2α in KIT-mutated gastrointestinal stromal tumor cell line (GIST-T1) and SF268 GBM cells (Extended Data Fig. 6a). GCN2 loss only conferred resistance to sunitinib at concentrations that induced the GCN2-ATF4 signaling (Extended Data Fig. 6b).

To explore how various classes of kinase inhibitors (i.e., type I and II, reversible and irreversible) might activate GCN2, we analyzed previously published GCN2 crystal structures. Both human and yeast GCN2 have been shown to form constitutive dimers in an inactive antiparallel conformation ^15,17,27. Upon activation, GCN2 undergoes a conformational change resulting in re-organization into a parallel active dimer. We aligned the kinase domains of the recently published co-crystal structure of dovitinib bound to GCN2 ²⁷ and EGFR bound to neratinib ²⁸. Dovitinib is an ATP competitive multi-kinase inhibitor reported to binds to the GCN2 kinase domain in the LINCS KINOMEscan (Fig. 5a). Our alignment data with dovitinib and neratinib suggests that neratinib fits into the GCN2 kinase domain and would be predicted to result in occlusion of the ATP binding pocket (Fig. 5d). Similar to our biochemical results with neratinib, treatment of SF268 GBM cells with dovitinib induced GCN2 autophosphorylation, phosphorylation of eIF2α (serine 51), and induction of ATF4 (Fig. 5e).

Neratinib binding stabilizes the active GCN2 dimer conformation

To determine how dovitinib might increase GCN2 kinase activity despite apparently occluding the ATP-binding pocket of GCN2, we expanded our in vitro kinase activity assay to a broader range of dovitinib concentrations. We observed that dovitinib enhanced GCN2 kinase activity at lower concentrations, but inhibited it at higher concentrations (Fig. 6a). We observed a similar, bell-shaped effect on GCN2 kinase activity in our ³²P labeling assay with neratinib (Fig. 6b). We also observed that neratinib was able to induce GCN2 activation in the absence of uncharged tRNAs (Fig. 6c).

Fig. 6. Mechanism of GCN2 activation by kinase inhibitors.

a, Activity assay showing ³²P labeling of recombinant GCN2 and eIF2α treated with high dose dovitinib and b, neratinib. c, Recombinant GCN2 activity assay showing GCN2 T899 autophosphorylation and substrate eIF2α S51 phosphorylation induction by neratinib in the presence and absence of starved tRNAs. d, Model of GCN2 kinase activation at intermediate concentrations of neratinib with higher concentrations leading to kinase inhibition.

Taken together, this data supports a model that neratinib binds to one of the two GCN2 monomers, induces a conformational change that stabilizes uncharged tRNA binding, and this conformational change favors activation of the GCN2 dimer, resulting in increased enzymatic activity through the unbound monomer (Fig. 6d). Further increases in kinase inhibitor levels lead to bound drug in both kinase domains, ultimately resulting in reduction of GCN2 kinase activity.

Taken together, our experiments demonstrated that induction of the GCN2 kinase occurs in response to several, widely used tyrosine kinase inhibitors and contributes to their antitumor activity. Since this finding had emerged from our genome-wide CRISPR library screen, we repeated this screen with the EGFR inhibitor, osimertinib, which did not induce GCN2 kinase activity in GBM cells (Extended Data Fig. 7a). Consistent with our hypothesis and in contrast to our results with neratinib, sgRNAs targeting GCN2 or other ISR pathway members were not enriched in cells growing in the presence of osimertinib (Extended Data Fig. 7b). To confirm that induction of GCN2 can augment the effects of EGFR inhibition in EGFR mutant cancer cells, we performed a synergy study using the EGFR inhibitor, osimertinib, and the GCN2 activator, dovitinib, in SF268 GBM cells and their SF268 GBM cells with GCN2-KO cells. Consistent with our model, we observed synergy between dovitinib and osimertinib at intermediate concentrations of dovitinib in the parental but not the GCN2-KO cells, suggesting that GCN2 activation sensitizes EGFR dependent GBM cells to EGFR inhibition (Extended Data Fig. 7c).

Discussion

Our study introduces a mechanism of GCN2 kinase activation by ATP-competitive kinase inhibitors and demonstrates the contribution of GCN2 activation to the antitumor activity of the pan-ErbB inhibitor neratinib. The family of ErbB receptor tyrosine kinases includes four members (i.e., epidermal growth factor receptor (EGFR), HER2, HER3, HER4) which are aberrantly activated in a variety of human cancers ²⁹. ErbB kinase inhibitors include reversible inhibitors such as erlotinib and gefitinib or irreversible inhibitors such as neratinib, osimertinib, and canertinib ^10,30,31. Our results in EGFR mutant glioblastoma cells show that in the absence of EGFR inhibition, EGFR signaling drives cell survival and cell growth, while the GCN2/ATF4 signaling axis is inactive. In the presence of neratinib, EGFR signaling is inactive and GCN2 signaling is active resulting in increased cell death and inhibited cell growth.

Genome-wide screens have the potential to expand our understanding of drug mechanism-of-action ³². We used this approach to evaluate the broad target landscape of kinase inhibitors ³³ and found that induction of the GCN2/eIF2α/ATF4 signaling axis contributes to the antitumor activity of neratinib in EGFR mutant GBM cells. Activating transcription factor 4 (ATF4) controls the expression of a wide range of adaptive genes and allows cells to endure periods of stress, but can also promote cell death under persistent stress conditions ³⁴. Our findings are reminiscent of the observation that treatment of V600E BRAF mutant melanoma cells with the BRAF inhibitor vemurafenib induces eIF2α phosphorylation and knockdown of ATF4 reduced vemurafenib-induced apoptosis ³⁵. While peak plasma levels of neratinib and other kinase inhibitors used in our study are sufficiently high to induce GCN2 in patients ^25,36–38, further studies with on-treatment tumor biopsies are needed to determine whether induction of ATF4, or lack thereof, is associated with the depth or duration of clinical treatment response to neratinib or other kinase inhibitors that activate the GCN2/ATF4 axis.

While binding of several kinase inhibitors to the isolated GCN2 kinase domain has been reported ²⁴, the induction of GCN2 kinase activity by these inhibitors and contribution to their antitumor activity was unexpected. Unlike prior examples in the literature, where ATP-competitive inhibitors were found to behave as agonists of their direct kinase target ^2,39, such as RAF ⁴⁰ and AKT ⁴¹, we describe direct effects on an unrelated kinase and signaling pathway. We hypothesize that the binding of certain kinase inhibitors to the GCN2 kinase domain promotes a conformational change of the dimer from the auto-inhibited antiparallel conformation to an active parallel conformation ^15,17,27, resulting in increased affinity of GCN2 for uncharged tRNAs and triggering GCN2 activation in the absence of amino acid starvation. This model is reminiscent of the transactivation of RAF dimers and increased ERK signaling in response to certain RAF inhibitors ^42,43. It is intriguing to speculate that inhibitor-stimulated kinase priming of the GCN2/ATF4 axis could perhaps be engineered into future kinase inhibitors to augment their antitumor activity.

Methods

Cell Lines and Reagents

Adherent lines were grown in DMEM 10% FBS (Omega scientific, FB-11). Neurospheres were grown in NeuroCult NS-A Pro-liferation Kit (Stem Cell Technology) supplemented with human EGF and bFGF (20 ng/ml each). Human embryonic kidney HEK-293T were authenticated by SNPs analysis. HEK-293FT cells were kindly provided by the Varmus Lab (Thermo). SF268 was obtained from National Cancer Institute. SKMG3 cells were provided by Conforma Therapeutics. NHA cells were kindly provided by Dr. Russell Pieper (UCSF). NR6 cells were kindly provided by Dr. Harvey Herschman (UCLA). GIST-T1 cells were kindly provided by Dr. Ping Chi (MSKCC). HEK-293T and CHO-K1 cell line was purchased from ATCC. GBM tumor spheres TS895 were derived at the MSKCC Brain Tumor Center according to MSKCC IRB guidelines. DNA fingerprinting was used for authentication of all glioma cell lines; no further validation was performed. All the cell lines were routinely checked for mycoplasma contamination by PCR analysis. Antibodies to P-GCN2 (ab75836/1000x), P-EIF2S1 (ab32157/2500x), GCN2 (ab134053/1000x) were purchased from Abcam. Antibodies to GCN2 (3302S/1000x), ATF4 (11815S/1000x), eIF2α (5324S/5000x), P-EGFR (3777S/1000x), EGFR (4267L/2500x), P-AKT T308 (13038/1000x), P-AKT S473 (4060S/2500x), AKT (4691L/5000x), P-ERK (4370L/1000x), ATF6 (65880/1000x), CHOP (5554S/1000x), 4F2hc (47213/1000x), xCT (12691S/1000x), ASCT2 (8057S/1000x), vinculin (13901S/5000x), Parp-CL (5625S/1000x), P-cKIT (3073S/1000x), cKIT (3074S/2500x) were purchased from Cell signaling. The antibody to GADD34/PPP1R15A (10449–1-AP/1000x) was purchased from Proteintech. The antibody to P-PERK (PA540294/1000x) was purchased from Thermo Fisher. The inhibitors Osimertinib (S7297), Canertinib (S1019), Erlotinib (S7786), Gefitinib (S1025), MK2206 (S1078), Trametinib (S2673), Sunitinib (S7781), Vemurafenib (S1267) were purchased from Selleckchem. The inhibitor neratinib is from Puma Biotechnology.

Plasmids

The lentiCRISPRv2 (Addgene #52961) and lentiGuide-Puro (Addgene #52963) vectors were cloned with guide RNAs that target respective genes (Supplementary Table). The lentiCRISPRv2 and lentiGuide-Puro vector were digested with BsmBI and ligated with annealed oligonucleotides. Overexpression GADD34 plasmids were generated by Gibson assembly (NEB E2611S) or Infusion cloning (Clontech 638909) into lentiCas9-Blast (Addgene #52962). Doxycycline inducible EIF2S1 plasmids were generated by Gibson assembly into pCW-Cas9 (Addgene #50661). Mutant EIF2S1 (S51D) was generated by site directed mutagenesis (NEB E0554S). All plasmids were Sanger sequence verified.

Cell line transfections and infections

For transduction of CRISPR/Cas9 into SF268 and SKMG3 cells, lentivirus was generated by transfection of CRISPR/Cas9 plasmids with pPAX2 and pMDG.2 packaging plasmids into HEK-293T cells using calcium phosphate. Viral particles were collected 48 hours after transfection. Virus was concentrated with lenti-X concentrator (Clontech #631232). Transductions were performed for 2 hours with ViraDuctin lentivirus transduction reagents (Cell BioLabs # LTV-201). Transduced cells were selected 48 hours post-transduction with blasticidin (2–7 μg/ml), G418 (500–1000 μg/ml), and puromycin (2–5 μg/ml) according to the plasmid antibiotic resistance. Clonal isolations were performed by serial dilutions (1 cell/well).

CRISPR screens

Pooled CRISPR screens were performed with the single vector TKOv3 CRISPR library (Addgene #90294). The TKOv3 plasmid library was amplified and the screen was performed as previously described ¹⁴. TKOv3 library lentivirus was produced by co-transfection of HEK-293FT cells with lentiviral vectors pPAX2 and pMDG.2 with TKOv3 lentiCRISPR plasmid library, using lipofectamine 3000 (Thermo #L3000008) transfection reagent. Virus was collected after 48 hours. The multiplicity of infection (MOI) was determined at 72 hours by comparing percent survival in puromycin selection. A total of 3.5 X 10⁷ cells were transduced with TKOv3 lentiviral library (71,090 gRNAs) at an MOI of ~0.3 to achieve ~500x coverage of the library after selection. The library transduction was carried out by spinfection at 1000 x g for 2 hours at 33°C. Cells were selected with puromycin (InvivoGen #ant-pr-1) at 2 μg/ml for 7 days. Selected cells were split and treated with vehicle (DMSO) or specified drug treatment for indicated times. Media and drug were refreshed every 3 days. A total of 3.5 X 10⁷ cells were collected at day 0 and either at day 14 or day 21 for total genomic DNA extractions. PCR was performed to amplify sgRNA regions followed by PCR addition of adapters and barcodes for Next Generation Sequencing (NGS). Samples were multiplexed on the Illumina NextSeq500 for single-end sequencing. Raw reads were aligned to the guide library using Bowtie. Fold change of the sgRNA counts are relative to the sgRNA ratios of the library plasmid pool. Significance (BH corrected) and normalized Z scores for all neratinib CRISPR screens was generated using the drugZ software ⁴⁴.

RNA-sequencing Analysis

Computational analysis was performed on multiple sets of bulk RNA-seq data. The 1 × 10⁶ cells were treated with the indicated drugs and performed in duplicate. Total RNA was extracted using the RNeasy Mini Kit (Qiagen, 74104) and used for poly-A library preparation (MSKCC core). Polyadenylated RNA-seq was performed using the standard Illumina Truseq kits and samples were sequenced using HiSeq2000 with one to three samples per lane. All RNA-seq analyses were performed using conventional RNA-seq analysis tools in R. All RNA-seq reads were aligned to the human reference genome (GRCh38/hg38) using salmon ⁴⁵. Differential expression analysis was performed with the DESeq2 package ⁴⁶. The differential expression of each treatment group was compared to the vehicle control group. The pathway analysis was performed with the GSEA software ^47,48. Gene sets for canonical pathways (c2) and gene ontology (c5) were downloaded from the Molecular Signatures Database (MSigDB).

GCN2 kinase activity assay

Total RNA for GCN2 ligand was collected from 4 hour glutamine starved SF268 cells followed by RNA extraction and purification. Purified recombinant GCN2 (Abnova #P5554, Sigma #14–934M) and EIF2S1 (Sigma #SRP5232) were purchased. The kinase activity assay was carried out essentially as previously described¹⁶. In short, 0.05 μg of purified GCN2 was incubated with 0.2 μg EIF2S1, 3 μCi of [γ−³²P]ATP (3000Ci/mmol, PerkinElmer), 50 ng of total RNA from Gln starved cells, 0.5 μg of bovine serum albumin in 25 μL of kinase assay buffer for 20 minutes at 30°C. The samples were resolved by 3–8% SDS PAGE and transferred to nitrocellulose membrane. Loading controls were detected by standard immunoblotting. The [³²P]-GCN2 was imaged with autoradiography film (Denville) exposed with intensifying screens at −20°C. Site specific detection of GCN2 autophosphorylation and eIF2α phosphorylation were carried out with but with unlabeled ATP at 50mM and detected with the T899 P-GCN2 (ab75836) and S51 P-EIF2S1 (ab32157) antibodies.

Recombinant GCN2 Electroporation

Recombinant GCN2 (Sigma #14–934M) was diluted in PBS to 1 μg/μL. Electroporation was performed with the Neon Transfection System (Thermo Fisher MPK1025). Cells were prepared by washing with PBS and resuspending the cells in electroporation Buffer R at 8 × 10⁶ cells/mL. For each electroporation reaction, 8 × 10⁵ cells were mixed with 2 μL of GCN2 protein or PBS. The cell/protein mixture was loaded into a 10 μL Neon Pipette Tip and electroporated with 1400V, 20 ms, 2 pulses. Electroporated cells were transferred to growth medium and allowed to recover for 3 hours after electroporation. Following recovery, cells were treated with vehicle or neratinib for 4 hours then harvested and lysed for western blotting.

ATP binding assay

Recombinant GCN2 (Sigma #14–934M) binding assays were performed in 25 μL reactions with 186 nM GCN2 protein, 25 ng of total RNA from 4 hour glutamine starved SF268 cells, 46.5 μM or as specified of ActivX desthiobiotin-ATP probe (Thermo 88311), and indicated concentrations of drug in kinase buffer (50 mM Tris-HCl, 30 mM MgCl2, 1 mM DTT). The mixtures were then incubated at 25°C for 20 minutes. Pierce high capacity streptavidin agarose beads (Thermo 20357) were then added to the reaction and allowed to pull down biotinylated proteins for 1 hour at 25°C on a rotator. Recombinant GCN2 protein will be biotinylated unless the ATP-binding pocket is occupied by neratinib. Beads were washed 3 times and eluted with 2x Laemmli sample buffer. Pulldowns were then analyzed by immunoblot.

qPCR

Total RNA was extracted using the RNeasy Mini Kit (Qiagen, 74104) and first-strand cDNA was synthesized with the qScript cDNA SuperMix (Quantabio, 95048). qPCR was performed with Taqman Absolute Blue qPCR Low ROX reagents (Thermo AB4319A). Fold changes in expression were calculated using the ΔΔC_t method. qPCR probes to ATF4(Hs00909569_g1) and RPLPO (4326314E) were purchased from Thermo Fisher. Results were plotted as average of all quadruplicate technical replicates from duplicate experiments.

GC-MS Analysis of Metabolites

SF268 cells were cultured in DMEM (or DMEM without glutamine/leucine) with vehicle or neratinib for indicated times prior to metabolite extraction. For harvesting metabolites, cells were rinsed with 4° C PBS and metabolites were extracted from cells with overnight incubation of 80:20 methanol:water containing 2 μM deuterated 2-hydroxyglutarate (D-2-hydroxyglutaric-2,3,3,4,4-d5 acid; deuterated-2HG) as an internal standard for the GCMS at −80° C. Samples were then vortexed and centrifuged at 21,000 G for 20 min at 4° C. Metabolite extracts were dried in a vacuum evaporator (Genevac EZ-2 Elite). Dried metabolites were resuspended by the addition of 50 mL of methoxyamine hydrochloride (40 mg/mL in pyridine) and incubated at 30° C for 90 min with agitation. Metabolites were derivatized by the addition of 80 mL of N-methyl-N-(trimethylsilyl) trifluoroacetamide (MSTFA) + 1% 2,2,2-trifluoro-N-methyl-N-(trimethylsilyl)-acetamide, chlorotrimethylsilane (TCMS; Thermo Scientific) and 70 mL of ethyl acetate (Sigma) and incubated at 37° C for 30 min. Samples were diluted 1:2 with 200 mL of ethyl acetate, then analyzed using an Agilent 7890A GC coupled to Agilent 5977 mass selective detector. The gas chromatography was operated in splitless mode with constant helium carrier gas flow of 1 mL/min and with a HP-5MS column (Agilent Technologies). The injection volume was set to 1 mL and the gas chromatography oven temperature was ramped from 60° C to 290° C over 25 min. Peaks representing compounds of interest were extracted and integrated using MassHunter vB.08.00 (Agilent Technologies) and then normalized to the internal standard (deuterated-2HG) and total protein extraction. The peaks were manually inspected and verified relative to known spectra for each metabolite. Metabolite abundance is represented as log₂ fold change over the vehicle average. Dendrograms for the heatmap were produced with hierarchical clustering.

Western Blotting

Protein extracts were prepared by using cell lysis buffer (CST #9803S) with protease (Sigma #539134) and phosphatase inhibitors (Sigma #524625). Cells were rinsed with PBS and sonicated. Equal amounts of total protein were separated on NuPAGE Bis-Tris gels or NuPAGE Tris-Acetate gels (Life Technologies), transferred to nitrocellulose membrane and incubated with the indicated primary antibodies.

Drug Synergy

Cells were seeded in 96 well plate with the density of 4,000 cells per well cultured in 200 uL of media. Cells were treated with the indicated drug doses for 6 days. Each treatment group was seeded in duplicate. Viability was measured with CellTiterGlo (Promega #G7570) and measured with the GloMax luminometer according to the manufacturers protocol. Bliss synergy scores⁴⁹ were analyzed with the SynergyFinder software.

Assessment of Cell Growth

Cells were seeded on 5-cm dishes in triplicate and allowed to attach overnight. Cells were then treated with the indicated drug doses for 6 days, unless specified otherwise. After treatment, attached cells were harvested and counted by trypan blue exclusion on a ViCell Cell Viability analyzer. Cell count is expressed as the percentage of trypan blue negative cells over the average percentage of trypan blue negative cells in the vehicle group.

Subcutaneous Tumor Xenografts

Mice were restrained to expose the flank where the hair was removed with an electric razor, and the injection site was disinfected with 70% ethanol. 10⁶ cells in 100 μL of growth media and Matrigel (BD, 356237) at a 1:1 mixture, was injected under the skin of the flank. Mice with tumors that exceeded 1.5 cm in diameter were sacrificed and removed from the study in the survival curve. Control mice were given vehicle and treatment mice were given 1000 mg/kg of neratinib by oral gavage.

Orthotopic Tumor Xenografts

To evaluate the transcriptomic effects of neratinib in a xenograft model, ICR-SCID mice were inoculated orthotopically (AP = 0 mm, ML = 2.0 mm, DV = 3.0 mm) with TS895 cells. The tumors were monitored by magnetic resonance imaging (MRI) measurement until tumor volumes were estimated to be approximately 3 mm³. Mice were dosed with an intraperitoneal injection of 0.05 mg EGF and either vehicle or neratinib at 40 mg/kg for 3 hours. Tumor tissue was lysed and total RNA was extracted and prepared for bulk RNA-sequencing.

Protein structure analysis

Crystal structures of EGFR complexed with neratinib (PDB: 3W2Q) ²⁸, EGFR complexed with an ATP analog (PDB: 2GS6) ⁵⁰, and GCN2 complexed with dovitinib ²⁷ were aligned and visualized with the Schrodinger suite software.

Statistics and Reproducibility

Two-sided unpaired t-test was used where indicated. Bar graphs and error statistics in the text are reported as the mean ± standard deviation, unless specified otherwise. All statistical analyses were performed in R and all IC₅₀ calculations were performed in Prism 8.0. Experiments were performed with at least two biological replicates, with similar results obtained each time. The number of animals to perform our in vivo studies was chosen to ensure 90% power with 5% error based on observed standard deviation from previous studies.

Extended Data

Extended Data Fig. 1. CRISPR screen resistance candidates are significant and consistent.

a, Schematic of genome-wide CRISPR screens using the TKOv3 library. b, Representative graph of a CRISPR screen showing the second highest sgRNA (per gene) frequencies of SKMG3 cell line treated with 2 μM neratinib versus vehicle for 14 days. c, Effect of GCN2-KO on neratinib treatment measured by SKMG3 cell growth and assessed by trypan blue exclusion after 6 days. Immunoblot confirming GCN2-KO on right side. d, Spearman correlation matrix of all CRISPR screen experiments showing spearman coefficients. e, false discovery rates of ISR pathway genes from drugZ analyzed CRISPR screens. f, Immunoblot of ATF4 induction by 4 hour 675 nM neratinib in parental and ATF4-KO SF268 cells.

Extended Data Fig. 2. Induction of the ISR by neratinib is GCN2 dependent.

a, Immunoblot of SKMG3 parental and GCN2-KO and SF268 parental and GCN2-KO cells treated with 675 nM neratinib for 4 hours. b, ISR induction time course in TS895 cells EGFP or GCN2-KO cells treated with 675 nM neratinib for indicated times.

Extended Data Fig. 3. Constitutive activation of eIF2α induces cell death.

a, Immunoblot of SF268 cells were transduced with doxycycline (dox) inducible wild-type (WT) or constitutively active (S51D) eIF2α and treated with 5μg/mL dox for indicated times (upper band, tagged recombinant eIF2α). b, effect of dox induced WT or S51D eIF2α on cell growth. Data are presented as mean values ± SD of n=3 biologically independent samples.

Extended Data Fig. 4. Induction of the amino acid starvation response by neratinib requires GCN2.

a, Volcano plots and gene set enrichment analysis results showing differential expression of genes in SF268 cells treated with 675 nM neratinib or vehicle for 6 hours and b, 72 hours with gene labels corresponding to the enriched krige amino acid gene set. c, gene set enrichment analysis results showing differential expression of genes in SF268 cells treated with 675 nM neratinib or vehicle for 6 hours and d, 72 hours. e, gene set enrichment plot for the amino acid deprivation gene set for intracranial TS895 tumors treated with neratinib 40 mpk for 3 hours. f, Plot comparing the enrichment of ER stress and amino acid deprivation pathways between SF268 parental and GCN2-KO cell lines treated with 675 nM neratinib for 6 hours. g, Comparison of ISR signaling in the presence of neratinib (1 μM)[lanes 1 and 2 also shown in panel a], glutamine (Gln) starvation, and ER stress by tunicamycin (5 μg/mL) for 4 hours.

Extended Data Fig. 5. GCN2 is sufficient for neratinib induced eIF2α phosphorylation.

a, Schematic of recombinant GCN2 electroporation rescue. b, immunoblot of eIF2α phosphorylation for SKMG3 cells electroporated with recombinant GCN2 then treated with 675 nM neratinib for 4 hours. c, image quantification of eIF2α phosphorylation, normalized to vinculin.

Extended Data Fig. 6. GCN2 loss confers resistance to GCN2 activating KIT inhibitor.

a, Immunoblot of ISR signaling in GIST-T1 (Left) and SF268 parental/GCN2-KO (Right) treated with the KIT inhibitor sunitinib for 4 hours. b, Bar graph comparing the effect of sunitinib on SF268 parental and GCN2-KO cell growth after 6 days. Data are presented as mean values ± SD of n=3 biologically independent samples. Significance was determined with a two sided student t-test.

Extended Data Fig. 7. GCN2 activation by dovitinib synergizes with osimertinib.

a, Immunoblot of SF268 cells treated with either neratinib (Ner), osimertinib (Osi), or canertinib (Can) for 4 hours. b, Scatterplot of a CRISPR screen showing the second highest sgRNA (per gene) frequencies of SF268 cell line treated with 3 μM osimertinib versus vehicle for 21 days. ISR pathway members and PTEN are labeled. c, Synergy between GCN2 activation by dovitinib (0 – 1.5μM) and EGFR inhibition by Osimertinib (0 – 2μM) in SF268 parental (left) and GCN2-KO (right) cells. Synergy assessed with Bliss synergy scores.