Research ArticleTUMOR IMMUNITY

GCN2 drives macrophage and MDSC function and immunosuppression in the tumor microenvironment

See allHide authors and affiliations

Science Immunology  13 Dec 2019:
Vol. 4, Issue 42, eaax8189
DOI: 10.1126/sciimmunol.aax8189

Polarization Player

Tumor-associated macrophages (TAMs) and myeloid-derived suppressor cells (MDSCs) suppress T cell functions in the tumor microenvironment (TME). Halaby et al. examine how the serine-threonine kinase general control nonderepressible 2 (GCN2) is a critical driver of Mϕ and MDSC polarization in the TME. Myeloid-lineage deletion of GCN2 caused a shift in TAM and MDSC phenotypes toward increased antitumor responses within the TME due to proinflammatory responses and increased CD8+ T cell expression of IFN-γ. GCN2 was a critical driver of Mϕ polarization and immunosuppression within the TME by promoting translation of the CREB-2/ATF4 transcription factor. GCN2 activity negatively correlated with antitumor responses and overall survival in human melanoma, suggesting further study into therapeutic targeting of this gene.


General control nonderepressible 2 (GCN2) is an environmental sensor controlling transcription and translation in response to nutrient availability. Although GCN2 is a putative therapeutic target for immuno-oncology, its role in shaping the immune response to tumors is poorly understood. Here, we used mass cytometry, transcriptomics, and transcription factor–binding analysis to determine the functional impact of GCN2 on the myeloid phenotype and immune responses in melanoma. We found that myeloid-lineage deletion of GCN2 drives a shift in the phenotype of tumor-associated macrophages and myeloid-derived suppressor cells (MDSCs) that promotes antitumor immunity. Time-of-flight mass cytometry (CyTOF) and single-cell RNA sequencing showed that this was due to changes in the immune microenvironment with increased proinflammatory activation of macrophages and MDSCs and interferon-γ expression in intratumoral CD8+ T cells. Mechanistically, GCN2 altered myeloid function by promoting increased translation of the transcription factor CREB-2/ATF4, which was required for maturation and polarization of macrophages and MDSCs in both mice and humans, whereas targeting Atf4 by small interfering RNA knockdown reduced tumor growth. Last, analysis of patients with cutaneous melanoma showed that GCN2-dependent transcriptional signatures correlated with macrophage polarization, T cell infiltrates, and overall survival. Thus, these data reveal a previously unknown dependence of tumors on myeloid GCN2 signals for protection from immune attack.


Interactions between cancer cells, T cells, and myeloid cells in the tumor microenvironment (TME) are a key determinant in tumor pathophysiology. In general, survival and responses to therapy correlate favorably with the extent of intratumoral CD8+ T cell infiltration (1); however, when the infiltrate is predominately myeloid, therapy responses and survival are reduced (2).

Myeloid function is governed by environmental signals driving commitment to a functionally polarized state (3). In the TME, several factors influence myeloid function including hypoxia (4, 5), pH (6), and nutrient depletion (7). General control nonderepressible 2 (GCN2) is a Ser-Thr kinase found in all eukaryotic organisms that is activated by deacylated transfer RNAs (tRNAs) resulting from amino acid and glucose limitation (8). The principle substrate of GCN2 is the α subunit of eukaryotic translation initiation factor 2α (eIF2α) (9). After phosphorylation by GCN2, eIF2α’s guanosine diphosphate–guanosine triphosphate exchange activity is reduced, abrogating cap-dependent translation. The resulting changes in mRNA translation alter the cellular phenotype regulating metabolism, autophagy, proliferation, and survival (9). In T cells, GCN2 signals are associated with naïve T cell suppression, blocking entry into cell cycle and T cell receptor signal transduction and promoting Foxp3+ regulatory T cell function (10). Although data on GCN2 function in myeloid cells are limited, our laboratory has reported that GCN2 activation modulates macrophage (Mϕ) and dendritic cell (DC) responses in autoimmune disease promoting acquisition of an IL-10+TGF-β+ phenotype (11).

Myeloid-derived suppressor cells (MDSCs) are a mixed population of immature and highly immunosuppressive monocytic and granulocytic lineage cells elicited by cancer-driven myelopoiesis in mice and humans (12, 13). An essential molecular feature of MDSCs is prominent expression of genes involved in the metabolism of l-arginine (i.e., Arg-1 and inducible nitric oxide synthase). MDSCs expanded by tumors are phenotypically distinct from tumor-associated Mϕ (TAM) and DCs, although MDSCs can differentiate into mature myeloid cells at the tumor site (14).

TAMs are a mature myeloid population originating from both monocytes and tissue-resident Mϕ (15). TAMs often exhibit a suppressive phenotype with production of immunoregulatory factors suppressing innate and adaptive immunity and providing stromal support for tumor growth and metastasis (3). Similar to TAMs, MDSCs are potent suppressors of T cell function. In particular, the action of Arg-1 produced by TAMs and MDSCs in the TME has a profound effect on T cells, reducing T cell receptor signal transduction and inhibiting cell cycle progression by down-regulation of cyclin D3 (16, 17). This effect is the result of extracellular l-Arg consumption and subsequent activation of GCN2 in T cells. It is likely that GCN2 activity would affect myeloid behavior in the TME; however, there are no studies we are aware of that test this prediction. Therefore, we investigated the role of GCN2 in myeloid cell function in tumors. We found that GCN2 deletion altered Mϕ and MDSC phenotype, causing an abrogation of suppressive function and enhanced antitumor CD8+ T cell immunity in vivo. This was due to altered gene expression and metabolism limiting polarization in Mϕs and overall function in MDSCs. Thus, the data reveal that GCN2 is an essential driver of myeloid function in the TME shaping the tumor-immune landscape.


Myeloid GCN2 function is required for tumor growth and T cell exhaustion

To test the functional role for myeloid GCN2 in tumor growth, we monitored tumor growth in mice with a myeloid-lineage deletion of GCN2 (B6.Gcn2fl/flxLyz2+/Cre, hereafter referred to as LysM–GCN2 cKO mice) or littermate control mice lacking Cre expression (GCN2fl/fl mice). In control mice, B16F10 (B16) melanoma tumors grew rapidly, reaching a mean volume of 1300 mm3 at day 25 (d25) after implantation (Fig. 1A). In contrast, B16 tumor growth was markedly restricted in LysM–GCN2 cKO mice, where tumor weight and volume were reduced by 75% (Fig. 1A). MC38 colorectal and EL4 lymphoma tumors showed similar changes with a 10-fold reduction in tumor size, suggesting that myeloid GCN2 was required for tumor growth across several tumor types (Fig. 1A). To determine whether reduction of tumor growth in LysM–GCN2 cKO mice was associated with alterations in the cytokine milieu, we measured d17 whole B16 tumor mRNA for expression of cytokines by quantitative polymerase chain reaction (qPCR). Compared with control tumors, tumors from LysM–GCN2 cKO mice showed a 10-fold reduction in expression of the immunosuppressive cytokine interleukin-10 (IL-10) and increased in expression of inflammatory cytokines including Ifng, Il1b, and Tnfa (Fig. 1B). Because tumor size differences at d17 may affect cytokine expression independent of GCN2 function, we also examined inflammatory cytokine expression in small tumors of equal size (75 mm3) from control and LysM–GCN2 cKO tumor-bearing mice. We observed a similar pattern of increased inflammatory cytokine expression in the absence of myeloid GCN2 function (Fig. 1C), indicating that loss of GCN2 increases local inflammation in the TME.

Fig. 1 GCN2 is required for immunosuppression and tumor growth.

(A) B16, MC38, and EL4 tumor growth curves in LysM–GCN2 cKO and GCN2fl/fl mice. n = 7 mice per group. (B) qPCR analysis of cytokine mRNA levels in bulk B16 tumors from LysM–GCN2 cKO and GCN2fl/fl mice collected on d17. (C) B16 tumors of equal size tumors (75 mm3) were collected from mice of the indicated genotype, and cytokine message was analyzed by qPCR. (D) Uniform manifold approximation and projection and phenograph analysis of CD45+ infiltrate in B16 tumors showing the major immune subpopulations identified by CyTOF analysis. Data from four biological replicates were concatenated in the plots. (E) Heatmap depicting median signal intensity (MSI) of indicated markers derived from CyTOF. Rows are scaled by Z score. (F) Heatmap showing MSI of surface markers in TAMs (CD11b+F4/80+), CD8+, and CD4+ T cells, respectively. (G) Plots of the percentage of CD8+ T cells in B16 OVA tumors expressing either one or all four markers indicated as determined by flow cytometry. *P < 0.05, **P < 0.01, ***P < 0.001, ****P < 0.0001 as determined by the unpaired Student’s t test. For heatmaps, each column represents an individual mouse. All experiments were repeated three times with similar results.

To gain better understanding of the impact myeloid GCN2 has on tumor immunity, we performed time-of-flight mass cytometry (CyTOF) analysis of d20 B16 tumors. Phenograph analysis (18) of the CyTOF data revealed 11 distinct intratumoral immune cell populations (Fig. 1D). Most of the immune infiltrate was TAMs resolving as two populations differentiated by expression of CD11c (populations 5 and 9; Fig. 1D). DCs were present but represented a minor component of the infiltrate resolving as two distinct groups based on CD103 expression (populations 6 and 10; Fig. 1D). Tumors from LysM–GCN2 cKO mice showed differences in the composition and activation state of the immune infiltrate compared with tumors from GCN2fl/fl mice; this included an increase in CD8+ T cells, natural killer (NK) cells, and NKT cells and a trend toward decreased numbers of TAMs (Fig. 1D). We observed both granulocytic Ly6g+ MDSC (gMDSC) and monocytic Ly6GloLy6Chi MDSC (mMDSC) (Fig. 1D); however, there were no differences in numbers between the two groups. LysM–GCN2 cKO mice showed no change in bone marrow precursors or splenic Mϕ phenotypes compared with GCN2fl/fl mice (fig. S1, A and B), indicating that changes in cellular composition were specific to the TME.

The cytokine expression data (Fig. 1, B and C) suggested a more inflamed TME. In agreement with this, expression of major histocompatibility complex II (MHCII) and CD86 was increased in TAMs, whereas there was a reduction in expression of CD206 and programmed cell death ligand 1 (PD-L1) in LysM–GCN2 cKO tumors compared with controls (Fig. 1F). Likewise, intratumoral CD4+ and CD8+ T cells had decreased expression of the exhaustion marker programmed cell death 1 (PD-1), the regulatory T cell transcription factors FoxP3, CD39 (19), and the regulatory T cell marker Helios (Fig. 1F) (20). gMDSCs exhibited a loss of suppressive phenotype with reduced expression of PD-L1, CD206, and CD39, whereas mMDSCs appeared more proinflammatory with increased expression of CD86, MHCI, and MHCII (Fig. 1F).

Intratumoral CD8+ T cell exhaustion is associated with tumor growth and resistance to immune destruction. In this vein, it had been hypothesized that increased expression of multiple exhaustion markers in CD8+ T cells is indicative of a deeply exhausted phenotype (21). Because our CyTOF data showed reduced PD-1 expression in intratumoral CD8+ T cells from LysM–GCN2 cKO mice, we examined expression of four key exhaustion markers in CD8+ T cells to determine whether GCN2 affected this deeply exhausted state. In agreement with the CyTOF data, flow cytometry showed that intratumoral CD8+ T cells from LysM–GCN2 cKO mice showed decreased surface expression of PD-1 and LAG3 and decreased expression of TIM3 and TIGIT (Fig. 1G); in LysM–GCN2 cKO mice, intratumoral CD8+ T cells positive for all four exhaustion markers (i.e., the deeply “exhausted” CD8+ T cells) reduced threefold compared with controls (Fig. 1G). Together, these data show that loss of GCN2 in the myeloid infiltrate profoundly affects the TME driving Mϕ inflammatory maturation, which in turn promotes a more robust T cell infiltrate with reduced exhaustion and potentially improved effector maturation.

Myeloid GCN2 function is required for an immunosuppressive transcriptional landscape in melanoma

Myeloid cells account for about 1% of the total cellularity in B16 tumors. Nevertheless, bulk tumor transcriptome analysis showed a substantial impact on RNA expression patterns with >2900 transcripts showing significant changes in abundance [false discovery rate (FDR), <0.05; log fold change (logFC), >±1] in LysM–GCN2 cKO versus control tumors (Fig. 2A). In particular, there was a decrease in expression of genes associated with immunosuppression including Il10 and Arg1, as well as the transcript for the GCN2-responsive gene Asns (Fig. 2A) (22, 23). In contrast, genes involved in T cell migration (Selplg), Mϕ scavenging (Siglec1) and adhesion (Itgam), and tumor inflammation (Sepp1) increased in LysM–GCN2 cKO tumor-bearing mice compared with control tumors (Fig. 2A). Ingenuity pathway analysis (IPA) of diseases and functions predicted that these transcriptional changes in the tumor would result in increased cell death and morbidity, whereas pathways associated with cellular survival and viability were predominant in tumors from GCN2fl/fl mice (Fig. 2B). Thus, the loss of myeloid GCN2 function shifted the TME and gene expression patterns toward inflammation, T cell recruitment, and death of tumor cells.

Fig. 2 Increased inflammation signatures in the absence of myeloid GCN2 function.

(A) Total B16 tumor transcriptome was analyzed on d20 tumors by RNA-seq. Volcano plot shows differential expression comparing LysM–GCN2 cKO or control tumor-bearing mice. Red dashed line marks significance threshold (FDR, <0.01; logFC, >1). (B) Bar graph of IPA diseases and functions for analysis for tumor samples in (A). A Z score of >2 was considered significant. (C) Heatmaps showing differential expression (FDR, <0.01; logFC, >1) of selected proinflammatory or regulatory markers in whole tumors or FACS-sorted TAMs (CD11b+F4/80+), bulk MDSCs (CD11b+Gr1+MHCIIneg), gMDSCs (CD11b+Ly6GhiLy6CnegMHCIIneg), or mMDSCs (CD11b+Ly6ChiLy6GloMHCIIneg). (D) Plot of canonical IPA comparing TAMs (left) or MDSCs (right) isolated from tumor-bearing mice of the indicated genotype. For heatmaps, each column represents an individual mouse. All experiments were repeated two times with similar results.

We examined the impact of GCN2 deletion on specific myeloid populations in the TME. TAMs and MDSCs comprise most of the myeloid cells in the tumor infiltrate (Fig. 1C), and tracer experiments using a tomato reporter (24) showed that TAMs and MDSCs exhibit active CRE-mediated excision, whereas only a minority of DCs showed reporter expression (fig. S1C). Thus, we focused on TAMs and MDSCs for further analysis.

We found that both TAMs and MDSCs from LysM–GCN2 cKO mice showed increased expression of inflammatory chemokines and their receptors, cytokines, Toll-like receptors, and signal transduction machinery (Fig. 2C and tables S1 to S3). This was also associated with reduced expression of key drivers of immunosuppression (e.g., Il10 and Ahr) and regulatory polarization (e.g., Chil3, Cd163, and Marco) in TAMs and Vegfb, Ccl22, and Retnla in MDSCs (Fig. 2C). Examination of fluorescence-activated cell sorting (FACS)–enriched gMDSC and mMDSC populations showed a module of proinflammatory cytokine and chemokine mRNAs up-regulated in a concordant pattern in both populations, indicative of a more proinflammatory phenotype (Fig. 2C). IPA of TAM and MDSC RNA sequencing (RNA-seq) data predicted that GCN2 deletion altered homeostasis, reducing activity of retinoid X receptor (RXR) and epidermal growth factor signaling and increasing cholesterol biosynthesis (Fig. 2D). RXRs are key nuclear receptors controlling the response to cholesterol and influencing immunosuppressive function (25). Cumulatively, these data suggest that GCN2 is a key driver of the immunosuppressive phenotype in TAMs and MDSCs.

To examine the impact of GCN2 loss on the immune transcriptional landscape in the TME, we performed single-cell RNA-seq (scRNA-seq) of the immune infiltrate from LysM–GCN2 cKO and GCN2fl/fl B16 tumor-bearing mice. Similar to the results from CyTOF analysis (Fig. 1C), phenograph analysis of the pooled data from both experimental groups showed that the infiltrate consisted of 11 clusters of cells. Examination of the top 10 expressed genes from each cluster using the ImmGen database ( identified two clusters of CD8+ T cells (hereafter called CD8+ T cell cluster 1 and CD8+ T cell cluster 2), B cells, and innate cells, including two clusters of TAMs, MDSCs, and conventional DCs, plasmacytoid DCs (Fig. 3A and table S4). When we examined the relative population contribution from LysM–GCN2 cKO and GCN2fl/fl B16 tumor-bearing mice, it was observed that most of CD8+ T cell cluster 1 was derived from the LysM–GCN2 cKO group, whereas CD8+ T cell cluster 2 was almost exclusively from the GCN2fl/fl group (Fig. 3B). Similarly, NK and NKT cells increased in the LysM–GCN2 cKO group compared with GCN2fl/fl tumor infiltrates, whereas myeloid cells appeared more equally distributed between the two experimental groups with a slight increase in the GCN2fl/fl group (Fig. 3B).

Fig. 3 The immune transcriptional landscape in LysM–GCN2 cKO and GCN2fl/fl tumors.

d20 B16 tumors were collected, and the CD45+ immune infiltrate was enriched. Then, pooled samples of three mice per group were analyzed by scRNA-seq. (A) Heatmap showing relative expression of the 10 most highly expressed genes across each population cluster identified by phenograph analysis. cDC, conventional DC; pDC, plasmacytoid DC. (B) Overlay t-distributed stochastic neighbor embedding (t-SNE) plot (left) and graph (right) showing frequency of cells in each cluster compared with the total population. (C) Violin plots showing expression of selected cytokines and chemokines in TAM and MDSC clusters. (D) Heatmap showing the most highly expressed transcripts across the CD8+ T cell clusters. (E) t-SNE plot showing expression of Ifng in all clusters examined in (B) (left), total Ifng expression in all clusters (middle), and Ifng expression in each cluster (right). *P < 0.05, **P < 0.01, ***P < 0.001, ****P < 0.0001 as determined by Seurat single-cell analysis.

Upstream regulator pathway analysis of the two TAM and the MDSC clusters predicted that loss of GCN2 would increase activity of several proinflammatory effector cytokines and chemokines (fig. S3A). Analysis predicted that the most up-regulated cytokine pathways in the LysM–GCN2 cKO group included tumor necrosis factor–α (TNF-α) and IL-1β (fig. S3A). We did not observe differences in Tnfa expression in the scRNA-seq dataset (fig. S3B); however, Il1b mRNA abundance was increased in LysM–GCN2 cKO tumors compared with controls (Fig. 3C), suggesting that an increase in IL-1β–dependent inflammation may contribute to the observed antitumor effect of myeloid GCN2 deletion. To more deeply characterize the MDSC infiltrate, we performed scRNA-seq on FACS-enriched CD11b+GR-1+MHCIIlow-neg MDSCs from tumor-bearing mice. Phenograph and heatmap analysis identified six clusters of MDSCs: two gMDSC and four mMDSC (fig. S4, A and B). LysM–GCN2 cKO MDSCs were enriched for gMDSC cluster 2 and mMDSC cluster 1, whereas MDSCs from control tumors were enriched primarily for mMDSC clusters 1 and 4 (fig. S4C). IPA predicted that IL-1β would be most strongly affected by the loss of GCN2 in gMDSC cluster 2 (fig. S4D). Accordingly, this cluster showed the highest overall expression of IL-1β, and loss of GCN2 significantly (P = 0.000002) increased Il1b in gMDSC cluster 2 (fig. S4E), showing that the loss of GCN2 skews the intratumoral MDSC infiltrate favoring a more granulocytic, proinflammatory phenotype.

Because the CD8+ T cell clusters segregated transcriptionally based on GCN2 (Fig. 3B), we predicted that they represent distinct functional states corresponding to the inflammatory environment of the TME. Heatmap analysis of the top 65 genes expressed across both clusters showed that CD8+ T cell cluster 2 (derived from the control tumors) had a resting/naïve phenotype with increased expression of T cell developmental factors (Lef1, Tcf7, Klf2, Socs3, and S1pr1) lacking expression of activation or effector mRNAs (Fig. 3D and table S5). In contrast, CD8+ T cell cluster 1 (derived primarily from the LysM–GCN2 cKO group) showed increased expression of activation and effector mRNAs (Gzmk, Il2rb, Ccl5, and Isg20) (Fig. 3D). In particular, the gene encoding interferon-γ (IFN-γ) was one of the most differentially expressed between the two clusters (Fig. 3D), and intracellular FACS showed increased IFN-γ+ CD8+ T cells in LysM–GCN2 cKO tumors (fig. S3C). When we examined Ifng in all cells, we found that the expression was highest in CD8+ T cell cluster 1 as expected. We also observed Ifng message in NK and NKT cells (Fig. 3E). Intracluster comparisons of Ifng between LysM–GCN2 cKO and GCN2fl/fl tumor-bearing mice showed no differences within the populations (Fig. 3E). However, given the increased presence of these cell populations in LysM–GCN2 cKO tumors (Fig. 3B), the data show that myeloid GCN2 function represents a key barrier suppressing the accumulation of IFN-γ+ immune cells in the TME.

GCN2 is required for regulatory Mϕ polarization and MDSC function

The in vivo data demonstrated that the loss of GCN2 altered TAM and MDSC phenotypes and promoted CD8+ T cell acquisition of effector function. However, to gain a better understanding of the cell-intrinsic role of GCN2, we directly tested the impact of GCN2 loss on polarization and effector activity in vitro. We generated bone marrow–derived Mϕ (BMDM) from C57BL6/J (B6) and C57BL6/J.Gcn2−/− (B6.Gcn2−/−) mice polarized to either a classic inflammatory (iMϕ) or alternatively activated regulatory (aaMϕ) state. Superficially, loss of GCN2 did not appear to affect polarization as flow cytometry showed that PD-L1, CD86, CX3CR1, MHCI, and MHCII levels were comparable between B6 and B6.Gcn2−/− Mϕ (Fig. 4A). However, loss of GCN2 enhanced iMϕ transcriptional characteristics with an increase in expression (Fig. 4B) and production (Fig. 4C) of proinflammatory proteins. In contrast, aaMϕ polarization was attenuated by GCN2 deletion with a decrease in Arg1 and Ccl22 mRNA and IL-10 protein (Fig. 4, B and D). We tested whether deletion of GCN2 would reduce the ability of aaMϕ to suppress T cell proliferation. When CD8+ T cells were added to cognate peptide-pulsed DCs, there was proliferation with most of the T cells undergoing three to four rounds of division after 72 hours of coculture (Fig. 4E). In contrast, when B6 aaMϕs were added, there was a reduction in proliferation with most of the T cells only undergoing one or two rounds of division (Fig. 4E). Loss of GCN2 attenuated this effect, and we observed that most T cells had undergone three to four rounds of division. Thus, the data show that GCN2 is required for the alternatively activated phenotype and plays a direct role in Mϕ polarization.

Fig. 4 Cell intrinsic GCN2 is required for Mϕ polarization and function.

(A) Expression of surface markers in BMDM of the indicated polarization state was determined by flow cytometry. MFI, geometric mean fluorescence intensity. (B) Mϕs were polarized as described in Materials and Methods, and mRNAs for the transcripts indicated were measured by quantitative real-time PCR (qRT-PCR). uMϕ unpolarized Mϕ. (C and D) Culture supernatants from Mϕ cultured as described in (A) were tested for the indicated cytokines by ELISA. (E) aaMϕ cultures of the genotype indicated were tested for their ability to suppress T cell proliferation. OTI CD8+ T cells were labeled with CFSE, and proliferation was estimated by CFSE dilution after 3 days of coculture with OVA257–264 peptide-pulsed CD11c+ DCs ± aaMϕ (Mϕ:T cell ratio, 1:5) via FACS analysis. For the pie charts (bottom), n = 3 biological replicates per group. *P < 0.05, **P < 0.01, ***P < 0.001, ****P < 0.0001 as determined by the unpaired Student’s t test.

To test the impact of GCN2 on MDSC function, we generated MDSCs from bone marrow by in vitro culture with IL-6 and granulocyte-Mϕ colony-stimulating factor as previously described (26). Deletion of GCN2 in MDSCs affected the surface phenotype, with a twofold increase in CD86 expression, a 40% reduction in PD-L1, and a 30% reduction in IL-4Rα expression (Fig. 5A). Moreover, GCN2 deletion reduced expression of a number of MDSC effectors, including a reduction in Arg1 expression (Fig. 5B) and arginase activity (Fig. 5C). Consistent with our in vivo data, B6.Gcn2−/− MDSCs showed increased IL-1β message and protein (Fig. 5, D and E), suggestive of a less suppressive phenotype.

Fig. 5 MDSC function is attenuated in the absence of GCN2.

(A) Quantitation of immune surface markers on MDSCs as determined by flow cytometry. (B) qPCR analysis of indicated transcripts from cultures described in (A). Values are normalized to Bactin. (C) Quantification of arginase activity in MDSC cultures. (D and E) Measurement of IL-1β expression and protein production by qPCR and ELISA, respectively. (F) Representative Western blot showing C/EBPβ isoform LAP and LIP expression in fresh bone marrow cells (FBM) and 4-day MDSC cultures. (G) MDSCs cultures were tested for their ability to suppress T cell proliferation. (H) Ova ± EL4 tumor cells were labeled with CFSE, mixed at a ratio of 1:1, and added to purified OTI cells activated in the presence or absence of MDSCs as described in (G). After 8 hours, numbers of CFSE-positive (high and low) target cells were evaluated by FACS. *P < 0.05, **P < 0.01, ***P < 0.001, ****P < 0.0001 as determined by the unpaired Student’s t test. Experiments were repeated four times with similar results.

CCAAT/enhancer binding protein β (C/EBPβ) is a key transcription factor for MDSC function (26). Because it was reported that GCN2-driven expression of C/EBPβ is required for gluconeogenesis in the liver (27), we hypothesized that GCN2 would be required for expression of C/EBPβ in MDSCs. Analysis by qPCR and immunoblot showed that C/EBPβ was rapidly induced in MDSC cultures (Fig. 5, B and F). However, Cebpb mRNA expression was reduced in the absence of GCN2 correlating with a lack of C/EBPβ protein (Fig. 5F), indicating that GCN2 is required for C/EBPβ induction during MDSC differentiation.

MDSCs are defined by the functional ability to inhibit T cell proliferation and maturation (13). Accordingly, addition of control MDSCs to CD8+ T cell cognate antigen-pulsed DC cocultures strongly suppressed T cell proliferation (Fig. 5G). In contrast, B6.Gcn2−/− MDSCs showed a marked attenuation of their ability to suppress T cell proliferation (Fig. 5G). Moreover, when we tested the ability of the T cells to kill target tumor cells, we found that T cells from cultures containing B6 MDSCs exhibited a complete abrogation of killing activity, whereas T cells collected from cultures containing B6.Gcn2−/− MDSCs killed all target tumor cells in the assay (Fig. 5H), showing that GCN2 is required for functional maturation of MDSCs.

Activating transcription factor 4 is required for GCN2-dependent phenotype in macrophages and MDSCs

Translational induction of activating transcription factor 4 (ATF4) is a key mechanism mediating downstream effects of GCN2 (9). Thus, we asked whether GCN2-dependent effects are contingent on ATF4 induction. We examined the pattern of expression in MDSC cultures and found that ATF4 protein was induced within 1 day of culture initiation, in contrast to B6.Gcn2−/− MDSCs that failed to induce ATF4 (Fig. 6A). Arginase is a key suppressive effector used by MDSCs (13), and we observed a reduction of Arg1 expression in B6.Gcn2−/− MDSC cultures (Fig. 5B). We tested whether increased expression of ATF4 would rescue Arg1 expression and the immunosuppressive phenotype in B6.Gcn2−/− MDSCs. Introduction of an AFT4 expression vector caused a threefold increase in Arg1 expression in B6 MDSC cultures compared with controls (Fig. 6B). Similarly, ATF4 overexpression increased Arg1 expression fourfold in B6.Gcn2−/− MDSCs (Fig. 6B). Overexpression of ATF4 rescued the ability of B6.Gcn2−/− MDSCs to suppress T cell proliferation (Fig. 6C), whereas inhibition of Atf4 by small interfering RNA (siRNA) knockdown reduced the ability of B6 MDSCs to suppress T cell proliferation by 50% (Fig. 6C). These results show that GCN2 is required for ATF4 induction in MDSCs and further that ATF4 is a key driver of MDSC immunosuppressive activity.

Fig. 6 ATF4 is required for acquisition of an immunosuppressive phenotype in aaMϕ and MDSCs.

(A) Representative Western blot showing ATF4 protein in MDSC cultures. (B) Arg1 quantification by qPCR in MDSCs overexpressing ATF4. (C) Quantification of T cell suppression using MDSCs transfected with ATF4 siRNA (ATF4 KD) or ATF4 overexpression vector (ATF4 OE). (D) Representative Western blot showing ATF4 expression levels in aaMϕ stimulated with LPS (50 ng/ml). (E) qPCR for cytokine mRNA expression in uMϕ or iMϕ transfected with ATF4 siRNA. (F) qRT-PCR showing decreased regulatory marker and increased proinflammatory mRNA expression in aaMϕ transfected with ATF4 siRNA. (G) Pie chart showing the percentage of hits in different chromosomal regions identified by ATF4 ChIP-seq analysis. (H) Comparative heatmap showing regions of ATF4 binding enrichment in aaMϕs versus uMϕs. (I) Quantification of glycolysis and oxidative respiration in iMϕ and aaMϕ was determined by seahorse assay. ECAR, extracellular acidification rate. (J) Quantification of glycolysis and oxidative respiration in aaMϕ cultures was assessed by seahorse assay. n = 5 biologic replicated per group. **P < 0.01, ***P < 0.001, ****P < 0.0001 as determined by the unpaired Student’s t test. ns, not significant; FCCP, 2-[2-[4-(trifluoromethoxy)phenyl]hydrazinylidene]-propanedinitrile; OM, oligomycin. All experiments were repeated four times with similar results.

Similar to MDSCs, deletion of GCN2 in aaMϕs reduced ATF4 protein at baseline and after stimulation with lipopolysaccharide (LPS) (Fig. 6D), indicating that ATF4 induction and overall expression are dependent on GCN2. When we knocked down ATF4 in macrophages using siRNA, the cytokine expression profile resembled B6.Gcn2−/− Mϕ, with an increase in expression of the proinflammatory cytokines in iMϕ (Fig. 6E), whereas in aaMϕ, ATF4 knockdown decreased Arg1 and Ccl22 while increasing Il1b expression (Fig. 6F). This implies that in Mϕ, ATF4 induction serves as a feedback mechanism to limit proinflammatory function and promote regulatory polarization.

To gain a better understanding of the transcriptional response driven by ATF4, we investigated genome-wide ATF4-DNA binding in unpolarized macrophages (uMϕ) and aaMϕ. For this, we performed chromatin immunoprecipitation–coupled deep sequencing (ChIP-seq) (28). The analysis identified 91 ATF4-bound genomic sites (FDR, <0.1%) that were enriched in aaMϕ versus uMϕ (table S6). Moreover, ATF4-bound genomic regions were enriched near transcriptional start sites, suggesting that the primary function is regulation of proximal promoter activity (Fig. 6G). Overall, ATF4-bound DNA was low in uMϕ compared with aaMϕ (histogram; Fig. 6H), implying increased function under alternative polarization conditions. Analysis of the transcriptional start sites within 1 kb of the enriched regions identified known ATF4 target genes Asns and Trib3 (Fig. 6H); moreover, most of the enriched ATF4 target gene loci were involved in protein synthesis including aminoacyl-tRNA synthetases (Yars, Cars, Mars, Lars, Tars, Gars, Nars, and Lars), initiation factors/modulators of mammalian target of rapamycin (mTOR) function (Ei4ebp1, Trib3, and Ddit4), metabolic genes (Pck2, Pfkp, Atp9b, and Ndufa4l2), and regulators of inflammation (Map3k3, Il4i1, and Ctsc) (Fig. 6H).

These results imply that GCN2-driven ATF4 activity would affect metabolism. In macrophages, metabolism is closely related to functional polarization, and glycolysis is strongly up-regulated in iMϕ, whereas aaMϕ favor utilization of oxidative phosphorylation (OXPHOS) and fatty acid oxidation (FAO) (29). Much less is known about metabolism in MDSCs, but the emerging literature suggests that both glycolysis and OXPHOS serve key roles in MDSC expansion and suppressive function (30, 31). When the transcriptomes of wild-type TAMs from B16 tumors were examined, we identified expression of autophagy genes (Atg2a, Atg9a, and Atg101) and genes involved in OXPHOS and FAO (Ndufs3, ndufs5, Atp10d, Atp2a3, Cox4l2, Sdhb, Ndufs8, Ndufv1, Cox7a2l, Acss2, Acsf2, and Ppargc11b) (fig. S5A). However, GCN2 deletion reduced autophagy and OXPHOS/FAO transcripts and increased expression of glycolysis genes, suggesting that lack of GCN2 may shift Mϕ metabolism to a more glycolytic state (fig. S5A). In agreement with this, measurement of hydrogen ion flux and oxygen consumption by seahorse showed that B6.Gcn2−/− aaMϕ cultures showed a drop (twofold) in oxidative respiration compared with control Mϕ (Fig. 6I and fig. S5B). In contrast, iMϕ did not induce OXPHOS as robustly as aaMϕ, and we did not observe differences between wild-type and GCN2−/− Mϕ oxygen consumption rates (OCRs) (Fig. 6I). Gene expression in aaMϕ cultures showed that loss of GCN2 affected expression of several genes involved in oxidative respiration but did not affect genes associated with glycolysis (fig. S5C), suggesting that GCN2 may drive OXPHOS metabolism in Mϕ. Supporting this observation, knockdown of ATF4 reduced aaMϕ respiration but had no impact on glycolysis (Fig. 6J). Intratumoral MDSCs showed substantial differences in expression of genes associated with glycolysis and OXPHOS in the absence of GCN2 (fig. S6A), suggesting a shift to glycolytic metabolism. In vitro, B6.Gcn2−/− MDSCs were less energetic compared with controls with a large decrease in OCR but minimal changes in glycolysis [i.e., extracellular acidification rate] (fig. S6B). Cumulatively, the data suggest that GCN2 affects metabolism in both Mϕs and MDSCs and may regulate tumoral immunity, at least in part, by influencing metabolic utilization of the myeloid infiltrate.

MDSC function can be driven by endoplasmic reticulum (ER) stress–induced activation of the kinase phosphorylated extracellular signal–regulated kinase (PERK) (32, 33). Because GCN2 and PERK both target eIF2α (9), we tested whether a GCN2-PERK cross-talk existed that could affect GCN2 function in MDSCs by measuring expression of C/EBP homologous protein (CHOP) (encoded by Gadd153), a stress protein downstream of both GCN2 and PERK. Loss of GCN2 had no impact on ER stress–driven Gadd153 expression after treatment with tunicamycin; likewise, loss of PERK did not reduce Gadd153 expression in the absence of tryptophan (fig. S6C). This suggests that PERK and GCN2 function independently in MDSCs to drive an immunosuppressive phenotype.

ATF4 is required for tumor growth in vivo

Because we showed that ATF4 induction is a key functional node downstream of GCN2 that could be inhibited via siRNA interference, we next asked whether administration of ATF4 siRNA in vivo would affect Mϕ phenotype and tumor growth. We coupled ATF4 siRNA to nanoparticles optimized for in vivo nucleic acid delivery tagged with a fluorescent tracer to follow localization and uptake in vivo. Flow cytometric analysis showed that about 20% of TAMs were positive for the nanoparticles, although there was minimal uptake by T cells or CD45neg stroma (Fig. 7A). TAMs that were positive for the siRNA nanoparticles showed substantial reduction of ATF4 mRNA detectible by qPCR correlating with increased Il1b and Il6 mRNA levels (Fig. 7B), suggesting that knockdown of ATF4 altered TAM polarization in vivo. In mice that received siRNA (si) ATF4 nanoparticles, tumor growth was inhibited, whereas tumors in the siCTRL nanoparticle–treated group grew at the expected rate (Fig. 7C). This outcome suggested that (i) targeting a minority of TAMs in the TME is sufficient to affect tumor growth and (ii) the GCN2-ATF4 circuit is an essential driver of tumor growth that it can be targeted for therapeutically meaningful inhibition of growth in established tumors.

Fig. 7 ATF4 siRNA nanoparticle knockdown decreases B16 tumor.

(A) Plot showing uptake of AF647-tagged siRNA nanoparticles in tumor macrophages (CD11b+F4/80+), T cells (CD3+), or tumor/stromal cells (CD45neg) was determined by FACS 3 hours after siRNA nanoparticle injection. (B) Plots showing ATF4, IL-1β, and IL-6 mRNA levels of AF647+ TAMs were determined by flow cytometry. (C) Plot showing growth curve of B16 tumors in mice treated with siRNA nanoparticles (siRNA-NP) as indicated (left) and plot showing final tumor weights in these mice. i/v, intravenous; s/q, subcutatenous. (D) Growth curve (left) and final tumor weight (right) for mice of the indicated genotype ± administration of anti–IL-1β IgG or irrelevant control IgG as described in Materials and Methods. For (C) and (D), n = 4 mice per group. *P < 0.05, **P < 0.01, ***P < 0.001, ****P < 0.0001 as determined by the unpaired Student’s t test. Experiments were repeated three times with similar results.

On the basis of our observation that IL-1β levels increased in macrophages and MDSCs lacking GCN2 function, we predicted that in myeloid GCN2 KO mice, the decrease in tumor size was caused by higher local IL-1β production. To test this, we treated tumor-bearing LysM–GCN2 cKO and GCN2fl/fl mice with IL-1β blocking immunoglobulin G (IgG). In agreement with our prediction, mice receiving IL-1β blocking antibodies had an increase in tumor growth resulting in tumor size that was similar to control GCN2fl/fl mice receiving irrelevant IgG (Fig. 7D). This shows that IL-1β is a key driver of tumor-restricting inflammation in the absence of GCN2.

GCN2 activity correlates with human tumor Mϕ polarization and outcomes in melanoma

The data presented above strongly support the notion that GCN2 is a key driver of myeloid function in the TME. Thus, we next asked whether predictions generated by mouse modeling could be extended to humans. First, we generated macrophages sorted from the monocyte fraction of healthy donor blood peripheral blood mononuclear cells (PBMCs) (PBMϕ) (fig. S2B) knocking down either GCN2 or ATF4 message by siRNA, examining the impact on the transcriptome under inflammatory or alternative regulatory polarization conditions. Similar to mouse, IPA predicted that knocking down either GCN2 or ATF4 would increase inflammatory pathway activity with the strongest impact on the IL-1β pathway (fig. S7, A and C). In agreement with this, IL1B message increased in the siGCN2 and siATF4 knockdown groups in inflammatory PBMϕ (iPBMϕ) cultures compared with controls (Fig. 8A). In contrast, inflammatory effectors pathways that did not identify the IPA (i.e., TNFA) or those predicted to be affected with a low Z score (i.e., IL6) were not increased in the siGCN2 and siATF4 groups (Fig. 8A). IPA of alternatively activated PBMϕ (aaPBMϕ) transcriptomes that predicted inhibition of either GCN2 or AFT4 did not increase inflammatory pathway activity (fig. S7, B and D). Instead, in aaPBMϕ, attenuation of GCN2-ATF4 signaling was predicted to reduce activity of key immunoregulatory pathways including IL-4, IL-13, and vascular endothelial growth factor (fig. S7, B and D). Moreover, for both iPBMϕ and aaPBMϕ, knocking down ATF4 was predicted to negatively affect the EIF2AK4 (i.e., GCN2) pathway, further implicating ATF4 in GCN2 function in macrophages. Supporting IPA predictions, siRNA knockdown of GCN2 or ATF4 caused a reduction of the key regulatory mRNAs for CCL22, ARG1, and CD206 in a pattern analogous to mice (Fig. 8B). Next, we FACS-sorted Mϕ (fig. S2C) from five melanoma tumors (TAMϕ) and performed RNA-seq comparing TAMϕ with unpolarized PBMϕs, which have low GCN2-ATF4 pathway activity. We interrogated the sequencing data using a list of mouse transcripts differentially expressed in TAMs from LysM–GCN2 cKO tumors to identify tumor Mϕ-specific expression patterns that required GCN2 activity. By this approach, we were able to identify 51 GCN2-dependent transcripts that were common between human TAMϕ and murine TAMs (Fig. 8C). Of these GCN2-dependent transcripts, 33 were enriched in TAMϕ that included key drivers of immunosuppression (IL10, AHR, and SOCS3), autophagy (ATG3 and ATG5), and downstream effectors of GCN2-ATF4 (DDIT3) (Fig. 8C and table S7). Cumulatively, the data indicate that GCN2 function in human macrophages is analogous to mice and is required for Mϕ polarization programs.

Fig. 8 GCN2 controls human Mϕ polarization and influences tumor immune signatures and survival outcomes in melanoma.

(A) qRT-PCR analysis of proinflammatory markers in unpolarized PBMϕs (uPBMϕs) or iPBMϕs transfected with control, GCN2, or ATF4 siRNA, respectively. (B) qRT-PCR analysis of regulatory markers in uPBMϕs or aaPBMϕs as described in (A). (C) Heatmap showing selected differentially expressed genes in TAMs enriched by FACS from melanoma tumors (n = 5) compared with uPBMϕ. Each column represents a separate tumor/donor. (D) Correlation between GCN2u gene signature and immune cell infiltrates using CIBERSORT analysis of TCGA melanoma datasets. Tregs, regulatory T cells. (E) Survival curves of patients with melanoma based on tiered expression levels of the GCN2u gene signature show that positive survival outcome is correlated with higher GCN2u gene expression.

The results above show that GCN2 is a crucial driver of human Mϕ function required for enactment of transcriptional programs controlling acquisition of effector function in vitro. On the basis of this, we predicted that GCN2-dependent transcriptional control would affect disease outcomes in cutaneous melanoma. GCN2 is highly expressed in melanoma; however, the expression level was not prognostic, and there were no indications of increased expression or mutational burden contributing to disease (fig. S8), suggesting that absolute levels of GCN2 or altered function as a result of mutation is not a determinate factor in clinical outcomes. We next took the transcriptional dataset from the in vitro knockdown of GCN2 in aaPBMϕ identifying 60 genes that were induced (GCN2u; P < 0.01; logFC, >+0.5) and 114 genes that were suppressed by GCN2 knockdown (GCN2d; P < 0.01; logFC, <−0.5) to generate a transcriptional signature to probe The Cancer Genome Atlas (TCGA) (tables S8 and S9).

Because GCN2 suppresses inflammatory polarization and IL1B expression, we reasoned that transcripts enriched by GCN2 inhibition would correlate with inflammatory immune characteristics, whereas transcripts that decreased when GCN2 function was reduced would correlate with immunosuppression and a worse prognosis. To test this, we used CIBERSORT-derived estimates of proportions of immune cell subsets in melanoma (34), defined previously in a large pancancer analysis (35). Although we projected that the GCN2d signature would promote regulatory immune cell accumulation, we did not find association with any immune cell signatures in the TCGA melanoma datasets. In contrast, the GCN2u signature correlated with immune cell phenotype in the tumors (Fig. 8D). Positive correlations were identified for inflammatory and activated immune cell signatures including M1 macrophages, CD8+ T cells, activated NK cells, DCs, and activated memory T cells (Fig. 8D), suggesting that GCN2 may actively suppress nascent inflammatory responses in the TME. On the basis of the CIBERSORT results, we hypothesized that the GCN2u signature would correlate with survival in melanoma. To test this, we analyzed 458 patients for overall survival stratifying them into quartiles based on the expression of the GCN2u signature using a stratified Cox proportional hazards model with the American Joint Committee on Cancer pathologic tumor stages as the strata. The patients with the lowest expression of the signature exhibited worse survival times with 50% mortality at 1500 days after diagnosis, whereas those in the top quartile (i.e., 25% of patients that showed highest expression of the GCN2u signature) had the best survival curves with 50% mortality at 4500 days after diagnosis (Fig. 8E). Thus, the data suggest that in human melanoma, GCN2 activity may be a negative prognostic factor for antitumor immunity and overall survival.


Myeloid cells are present at all stages of tumor growth (36), and on the basis of patterns of arginase and indoleamine 2,3 dioxygenase (IDO) expression, it would be predicted that GCN2 would be a prominent modulator of myeloid phenotype (37). In this study, we show that GCN2 is required for the suppressive phenotype and function of TAMs and MDSCs. Loss of GCN2 in the myeloid tumor infiltrate had a profound and complex effect on the overall tumor transcriptional program, growth, and immune infiltrate composition driving inflammation and reduction in tumor growth. These results underscore two key facts: (i) the importance of myeloid cells in controlling tumor immune landscape and (ii) that GCN2 is a fundamental driver of myeloid phenotype in TME. It is hypothesized that GCN2 primarily affects cellular phenotype by promoting ATF4-dependent expression of genes involved in amino acid biosynthesis, autophagy, and cell death (38). Our data agree with this prediction because we found that ATF4 is required for suppressive Mϕ and MDSC identity in vitro. ChIP-seq analysis of ATF4 binding in Mϕ showed that ATF4 preferentially binds to transcriptional targets in aaMϕ versus uMϕ; however, most of the gene targets identified were tRNA-charging enzymes or other genes involved in amino acid synthesis and transport, genes of the mTOR pathway, or metabolism. These results are similar to ATF4-promoter interactions reported for other cell types (39), suggesting that GCN2-ATF4 may elicit a transcriptional response that is conserved across cell types. The data imply that GCN2 does not directly regulate cytokine production but affects myeloid function by regulating metabolism or overall protein production.

Our data from both mouse and human experiments identified IL-1β as a key cytokine suppressed by GCN2 activity. Ravindran et al. (40) reported that GCN2-deficient antigen-presenting cells produced higher levels of IL-1β with increased pathology in dextran sulfate sodium (DSS)–induced colitis. Similarly, we found that in LysM–GCN2 cKO TAMs and MDSCs, IL-1β expression was increased. When we treated LysM–GCN2 cKO with anti–IL-1β antibodies, tumor growth was restored to levels similar to control mice, showing that IL-1β is a major downstream effector pathway suppressed by GCN2. The role of IL-1β in cancer progression is controversial. On the one hand, IL-1β signaling and inflammation have been proposed to be an early driver of tumorigenesis and growth (41). In contrast, evidence suggests that IL-1β secretion by proinflammatory Mϕ was essential for NK cell activation and tumoricidal activity (42). Moreover, a report by Spalinger et al. (43) showed that myeloid knockout of protein tyrosine phosphatase nonreceptor type 2 caused inflammasome activation and IL-1β production in DSS-induced colitis. However, when colorectal tumors were induced in those mice by azoxymethane, tumor growth was reduced compared with controls. This suggests that the role of IL-1β in tumorigenesis is context dependent and that enhanced IL-1β production by myeloid cells may potentiate antitumor T cell responses (43).

The CIBERSORT and TCGA correlation suggests that the impact of GCN2 on the immune environment is primarily via suppression rather that activation of gene programs. Unexpectedly, genes that were dependent on GCN2 for expression did not correlate with immune signatures and had no correlation with survival, in contrast to genes that increased in GCN2 knockdown Mϕ. The transcripts suppressed by GCN2 (as shown in table S9) are not directly involved in immune function per se, suggesting that GCN2 may influence immunity in the human TME indirectly via metabolic control. TAMs from patients with melanoma showed increased expression of a number of transcripts that are affected by GCN2 function in mice, supporting the notion that GCN2 may serve an analogous role in the TME in humans.

GCN2 is downstream of several key immunoregulatory pathways, making it an attractive target for immune modulation therapy. Nevertheless, although our data show that loss of GCN2 function has a major impact on immune function and tumor growth in relatively small, transplantable tumor model systems, it remains to be seen whether targeting GCN2 will be a successful strategy against spontaneous tumorigenesis and more advanced disease. Currently, there are no existing drugs that inhibit GCN2, but an inhibitor was recently reported to block GCN2 activity in the low nanomolar range (median inhibitory concentration, 2.4 nM) (23). This compound had reasonable bioavailability and could target ASNSlo acute myeloid leukemia when used in conjunction with aspariganase (23). This result showed that GCN2 could be targeted therapeutically in vivo. Moreover, because some tumors may rely on GCN2 to adapt to a low-nutrient environment (22), GCN2 inhibition could serve a dual role as both an anticancer and an immuno-oncologic agent. However, caution must be taken because it is unknown what effect systemic GCN2 inhibition may have on normal physiology. In conclusion, our report shows that GCN2 is a key driver of Mϕ polarization in the TME, revealing a key new target for cancer immunotherapy.


Study design

The aim of this study was to determine the impact of GCN2 on myeloid function in the TME. We characterized the effect of myeloid-lineage deletion of GCN2 on tumor growth and immune infiltrates by high-dimensional cytometry (CyTOF), qPCR, and RNA-seq at the bulk and single-cell levels in a mouse model of melanoma. To determine the cell-intrinsic role of GCN2 in myeloid function, we conducted in vitro experiments characterizing the transcriptome by RNA-seq and DNA binding characteristics of its obligate downstream effector ATF4 in Mϕ and MDSC cultures. We also characterized GCN2-dependent effects on metabolism by RNA-seq, qPCR, and seahorse assay. The number of replicates is indicated in the figure legends. No animals were excluded from the studies.


B6, B6.Eif2ak4tm1.2Dron (B6.Gcn2−/−), B6.Eif2ak3flox/flox (Perkfl/fl), LysMCRE+/−, B6.Eif2ak4flox/flox (Gcn2fl/fl), B6.SJL-Ptprca Pepcb/BoyJ (CD45.1), B6.Cg-Gt(ROSA)26Sortm9(CAG-tdTomato)Hze/J(dtTomato), and C57BL/6j Tg(TcraTcrb)1100Mjb/J (OT1) mice were obtained from colonies maintained under specific pathogen–free conditions in the Princess Margaret Cancer Centre animal facility in accordance with the Institutional Animal Care and Use Committee guidelines.

Human melanoma samples

All human tissues were obtained through a protocol approved by the University Health Network institutional review board.

Cell culture

B16, MC38, and EL4 cells obtained from the American Type Culture Collection were grown in complete Dulbecco’s modified Eagle’s medium [10% heat-inactivated fetal bovine serum (FBS), streptomycin (100 mg/ml), and penicillin (100 mg/ml); Gibco]. For splenocyte cultures and cytotoxic T lymphocyte (CTL) activation assays, medium was further supplemented with 0.02 mM 2-mercaptoethanol (Gibco).

BMDMs and MDSCs (mouse) and PBMC-derived macrophages (human) were generated as previously described (26, 44). For further information, see Supplementary Methods.

CTL assay

For lysis target cell preparation, ovalbumin (OVA)–antigen expressing EL4 tumor cells were labeled with 2 μM carboxyfluorescein diacetate succinimidyl ester (CFSE), whereas the Ovaneg EL4 cells were labeled with 0.2 μM CFSE. These two groups of differentially CFSE-labeled target cells were mixed at a ratio of 1:1 and added to cultures with activated, purified OTI cells at an effector:target ratio of 1:1. After 8 hours, cultures were collected, and the CFSE-positive (high and low) cells were evaluated by flow cytometry. Specific lysis was calculated according to the previously described formula: [1 − (ratio of CFSElow/CFSEhigh of naïve cells)/ratio of CFSElow/CFSEhigh of activated cells] × 100 (45).

T cell suppression assays

We used either aaMϕ or MDSCs sorted out of B16-OVA tumors. CD8+ T cells were isolated from OT1+/+Thy1.1+/+ mice using the EasySep Mouse CD8+ T Cell Isolation Kit (catalog no. 19853, STEMCELL Technologies). CD11c DCs were isolated from the spleen using anti-CD11c biotin antibody and the EasySep Positive Selection Kit (catalog no. 18559, STEMCELL Technologies). DCs were pulsed with OVA257–268 peptide (S7951, Sigma-Aldrich) for 6 hours. T cells were labeled with CFSE at a final concentration of 5 μM. They were then incubated with OVA-pulsed DCs at a ratio of 1:10 in the presence or absence of suppressive myeloid cells (MDSCs or aaMϕ). T cell proliferation was determined by detecting CFSE dilution on gated CD3+CD8+Thy1.1+ cells by flow cytometry.

Tumor studies

For the tumor models, mice were injected subcutaneously with 2 × 105 cancer cells and then monitored every other day for tumor growth. For information about nanoparticle-mediated ATF4 knockdown, see Supplementary Methods.

RNA isolation and quantitative real-time PCR

RNA from cells was purified using RNeasy Plus RNA purification kits (QIAGEN) and reverse transcribed using Quantabio qScript cDNA SuperMix. For qPCR reaction, complementary DNA (cDNA) was amplified using the PerfeCTa SYBR Green SuperMix on a CFX Connect real-time PCR detection system (Bio-Rad). Results were analyzed using the accompanying software according to the manufacturer’s instructions.

Tumor sample preparation and flow cytometry

B16 tumors were digested using collagenase IV (100 U/ml) and deoxyribonuclease I (DNase I; 50 U/ml) in complete RPMI 1640 medium at 37°C. For CyTOF experiments, the CD45+ population was enriched using STEMCELL EasySep Biotin Positive Selection Kit (cat. no. 18559) after staining with anti-mouse CD45.2 biotin antibody (catalog no. 60118BT). For analysis of intracellular cytokine production, cells were incubated with GolgiStop (eBioscience) for 4 to 5 hours and then washed and fixed/permeabilized with permeabilization/fixation buffer (eBioscience). For flow cytometric analysis, at least 105 events were collected on the LSRFortessa Flow Cytometer (BD Biosciences). Data were analyzed by FlowJo (Tree Star Inc.).

For human melanoma samples, fresh biopsies were minced and incubated in complete RPMI 1640 medium in the presence of collagenase IV (100 U/ml) and DNase I (50 U/ml). They were incubated in gentleMACS Octo Cell Dissociator (Miltenyi Biotec) at 37°C with gentle dissociation. Medium was then added to wash the cells and neutralize the reaction. Samples were then filtered twice, stained, and sorted as described below.

For sorting experiments, mouse tumor cells were stained with anti-CD45.2, anti-CD11b, and anti–Gr-1 (for MDSCs) and anti-CD45.2, anti-CD11b, anti–Gr-1, and anti-F4/80 (for TAMs). For isolation of human melanoma macrophages, human anti-CD45, anti–HLA-DR, anti-Lin (CD3, CD19, and CD56), anti-CD14, anti-CD11b, and anti-CD163 monoclonal antibodies were used on single-cell suspensions generated as described above, and the cells were sorted on a MoFlo (Beckman Coulter) cell sorter using the gating strategy illustrated in fig. S2C.


Western blot was performed according to previously published methodology (46). For more information, see Supplementary Methods.

RNA sequencing

RNA samples were quantified by Qubit (Life Technologies), and an Agilent Bioanalyzer assessed the RNA quality. All samples had RNA integrity number (RIN) above 8. SMART-Seq v4 Ultra Low Input RNA Kit for Sequencing (#634894, Takara Bio) was used per the manufacturer’s instructions for amplification of RNA and subsequent cDNA synthesis as previously described. For more information, see Supplementary Methods.

RNA-seq analysis

Raw fastq sequencing reads were aligned against the respective reference genome sequence (GRCm38/mm10 or GRCh37/hg19) using the STAR aligned v2.5.0c (47), discarding all nonuniquely aligned reads. For read counting per annotated gene, we used the STAR function “--quantMode GeneCounts,” counting reads matching exons of the Ensembl V75 gene annotation. Further processing was performed with the R Bioconductor package edgeR v.3.14.0 (48) using nonstranded reads. Reads were normalized for intra- and intersample variances using the functions “calcNormFactors” and “estimateTagwiseDisp,” resulting in counts per million for each gene. Differential gene expression analysis was performed with the functions “glmQLFit” and “glmQLFTest,” reporting P value, FDR and log2FCs between any possible pair-wise comparison and gene.

Chromatin immunoprecipitation–coupled deep sequencing

B6 uMϕs or aaMϕs BMDMs were cross-linked using 16% formaldehyde solution and then washed and scraped into phosphate-buffered saline (PBS). Nuclei were then isolated, digested with micrococcal nuclease, and then sonicated to release chromatin material using the SimpleCHIP Enzymatic Chromatin IP Kit (catalog no. 9003, Cell Signaling Technology). Immunoprecipitation using ATF4 antibody (D4B8; catalog no. 11815, Cell Signaling Technology) or anti-rabbit serum was performed overnight. Immunoprecipitated DNA was then purified and ran on a Bioanalyzer 2100 to check sample size and concentration. Libraries were then prepared using the NEBNext Ultra II DNA Library Prep Kit for Illumina (catalog no. E7645, New England BioLabs).

ChIP-seq analysis

Raw fastq data of ChIP-seq were processed with the hic-bench pipeline as previously described (49). For information about the analysis, see Supplementary Methods.

scRNA-seq analysis

Tumors from three GCN2fl/fl or 3 LysM-CRE cKO mice were digested, pooled, and stained with anti-CD45.2 antibody and 4′,6-diamidino-2-phenylindole. Live CD45+ cells were FACS-sorted into buffer (PBS + 2% FBS), washed 2× with PBS + 0.04% BSA, and then mixed with 10X Genomics Chromium single-cell RNA master mix, followed by loading onto a 10X Chromium chip according to the manufacturer’s protocol to obtain single-cell cDNA. Libraries were subsequently prepared and sequenced using the HiSeq 2500 sequencer (Illumina). For further information about the scRNA-seq analysis, see Supplementary Methods.

Enzyme-linked immunosorbent assay

BMDMs were treated with LPS alone (100 ng/ml) for 5 hours (for IL-6 and TNF-α), LPS + 5 mM adenosine triphosphate (for IL-1β), or IL-4 (100 ng/ml) overnight (for IL-10), and then supernatants were collected. Enzyme-linked immunosorbent assays (ELISAs) were performed using kits from Invitrogen (catalog no. 88-7324 for TNF-α, catalog no. 88-7064-88 for IL-6, catalog no. 88-7013-88 for IL-1β, and catalog no. 88-7105-88 for IL-10) according to the manufacturer’s instructions. The plates were read using a BioTek microplate reader at a wavelength of 450 nm.

Seahorse assay

In vitro–differentiated macrophages or MDSCs were plated in an Agilent Seahorse XF96 microplate at a density of 50,000 to 60,000 cells per well. In the case of MDSCs, the wells were precoated with 0.1% gelatin to allow cells to adhere to the bottom. Macrophages were either left unpolarized (uMϕs) or treated with LPS (50 ng/ml) + IFN-γ (iMϕs; 50 ng/ml ) or IL-4 (50 ng/ml) + IL-13 (aaMϕs; 50 ng/ml) overnight. During that time, the Agilent Seahorse XFe96 cartridge was equilibrated overnight in a CO2-free incubator according to the manufacturer’s instructions.

Time-of-flight mass cytometry

Purified unconjugated antibodies were labeled with metal tags at the SickKids Flow and Mass Cytometry Facility using the Maxpar Antibody Labeling Kit from Fluidigm (catalog no. 201300). Alternatively, directly conjugated antibodies were purchased from Fluidigm. CD45-enriched tumor single-cell suspensions were stained with antibodies that did not perform well after fixation (indicated by asterisk in the CyTOF antibody table) for 5 min at room temperature, washed with PBS, and then pulsed with 12.5 μM cisplatin (BioVision) in PBS for 1 min before quenching with CyTOF staining media [Mg+/Ca+ Hanks’ balanced salt solution containing 2% FBS (Multicell), 10 mM Hepes (Corning), and FBS underlay]. Cells were then fixed for 12 min at room temperature with transcription factor fixative (00-5523-00, eBiocience) and permeabilized, and individual samples were barcoded according to the manufacturer’s instructions (Fluidigm 20-Plex Pd Barcoding Kit, 201060), before being combined. Pooled samples were then resuspended in staining media containing metal-tagged surface antibodies and Fc block (CD16/32; in-house) for 30 min at 4°C. Cells were then stained with metal-tagged intracellular antibodies using Transcription Factor Staining Buffer Set (00-5523-00, eBiocience), according to the manufacturer’s instructions. Cells were then incubated overnight in PBS (Multicell) containing 0.3% (w/v) saponin, 1.6% (v/v) paraformaldehyde (diluted from 16%; Polysciences Inc.), and 1 nM Iridium (Fluidigm). Cells were analyzed on a Helios mass cytometer (Fluidigm). EQ Four Element Calibration Beads (Fluidigm) were used to normalize signal intensity over time on CyTOF software version 6.7. Flow cytometry standard files were manually debarcoded and analyzed using Cytobank 6.2. For further information about the analysis, see Supplementary Methods.

Statistical analysis

Means, SDs, and unpaired Student’s t test results were used to analyze the data. Tumor growth was analyzed using two-way analysis of variance (ANOVA). TCGA survival was analyzed by Kaplan-Meier Curve. When comparing two groups, P ≤ 0.05 was considered to be significant.



Fig. S1. Flow cytometric analysis of LysM-GCN2 cKO splenic macrophages and bone marrow precursors.

Fig. S2. Gating strategies for sorting mouse and human tumor associated macrophages and MDSCs.

Fig. S3. Increased proinflammatory signaling pathways in cytotoxic T cell and myeloid clusters of LysM–GCN2 cKO versus GCN2fl/fl tumors.

Fig. S4. Single cell RNA sequencing of mouse intratumoral-MDSCs.

Fig. S5. Altered metabolism in TAMs lacking GCN2 function.

Fig. S6. Altered metabolism MDSCs lacking GCN2 function.

Fig. S7. IPA of upstream regulators in PBMϕs.

Fig. S8. Landscape of EIF2AK4 (GCN2) gene expression and somatic alterations across 10,967 tumors from 32 cancer types in the TCGA PanCan study visualized by (10, 11).

Table S1. Differential expression of genes in whole B16 tumors from GCN2fl/fl versus LysM–GCN2 cKO mice shown in Fig. 2C heatmap.

Table S2. Differential expression of genes from GCN2fl/fl versus LysM–GCN2 cKO TAMs shown in Fig. 2C heatmap.

Table S3. Differential expression of genes from GCN2fl/fl versus LysM–GCN2 cKO MDSCs shown in Fig. 2C heatmap.

Table S4. Top 20 differentially expressed genes in B16 clusters shown in Fig. 3A heatmap.

Table S5. Differential gene expression between CD8 T cell clusters 1 and 2 shown in Fig. 3D heatmap.

Table S6. Identified ATF4 ChIP-seq chromosome features in uPBMϕs versus aaPBMϕs.

Table S7. List of differentially regulated genes and their functions in human melanoma TAMs versus resting, unpolarized macrophages (uMϕs) shown in Fig. 7D heatmap.

Table S8. List of genes that are down-regulated after GCN2 KD in human M2 macrophages.

Table S9. List of genes that are up-regulated after GCN2 KD in human aaMϕ.

Data file S1.

References (5053)


Funding: This work was supported by NIH grants CA190449, AI105500, and AR067763; the Medicine by Design/Canada First Research Excellence Fund; and the Canadian Institutes of Health Research (grant PJT-162114 to T.L.M.). Author contributions: T.L.M. designed and supervised the research. M.J.H., K.H., S.L., B.L.M., M.G., T.D.S.M., and M.T.C. executed the biochemical, cell biological, and in vitro experiments. M.J.H., K.H., and S.L. performed the animal experiments. A.K. and A.T. analyzed the RNA-seq results. M.J.H., K.H., and T.D.S.M. performed the ATAC-seq experiments and analysis. A.C., S.U., V.B., D.H.M., T.J.P., D.D.D., M.O.B., P.S.O., and D.G.B. contributed reagents, human samples, analysis of TCGA datasets, and discussions. M.J.H., K.H., S.L., T.C.G., A.K., M.G., A.C., T.J.P., and T.L.M. prepared figures and conducted statistical analysis, and M.J.H. and T.L.M. wrote the paper. Competing interests: The authors declare that they have no competing interests. Data and materials availability: All materials described in the study are either commercially available or upon reasonable request to the corresponding author under a material transfer agreement (T.L.M.). The RNA-seq and ATAC-seq data for this study have been deposited in NCBI GEO ( and are available under the accession identifier GSE140029.

Stay Connected to Science Immunology

Navigate This Article