Main

T cell differentiation is a complex process of integrating numerous signals from the environment, engaging transcriptional machinery and making epigenetic changes to support that new functional program1. For CD8+ T cells, this results in acquisition of various fates: short-lived, cytotoxic effector programs and long-lived, self-renewing memory programs2. However, a fate alternative to effector or memory differentiation is induced in pathologies in which antigen persists, notably chronic infections and cancer: T cell exhaustion. Under persistent inflammatory cues, T cells progressively lose polyfunctionality and renewal capacity, failing to control infection or malignant spread3. T cells progressively upregulate co-inhibitory molecules, ultimately reaching a terminally differentiated state expressing high levels of programmed cell death (PD)-1 and multiple co-inhibitory and co-stimulatory markers (LAG-3, Tim-3, TIGIT and 4-1BB). Exhausted cells are defined by several transcription factor networks including Eomesodermin, BATF, partner-less nuclear factor of activated T cells (NFAT)1, Tox and the transcriptional repressor Blimp-14,5,6,7,8, exploiting an altered epigenetic landscape defining the dysfunctional lineage9,10,11. Exhaustion has an evolving definition12 and can be reliably produced only in vivo, severely limiting the capacity to identify drivers of this fate.

Differentiation is typically linked to immunologic cues received by a T cell: co-stimulatory or cytokine signals supporting one fate or another. However, the metabolic composition of the environment, nutrient sensors detecting that milieu and the intrinsic metabolic state of a T cell are also crucial in dictating the ultimate outcome of differentiation13. It is critical to understand metabolic signals, because T cell activation and differentiation occur in various metabolically distinct tissue environments. In cancer, elevated tumor cell metabolism can dramatically alter the nutrient milieu that T cells experience14. We have previously shown that T cells infiltrate tumors at a severe metabolic disadvantage: repressed capacity for glucose uptake and loss of functional mitochondrial mass, concomitant with the development of exhaustion15. We and others have also shown that correction of this faulty metabolic state (through a variety of approaches) can invigorate immunity and enable immunotherapeutic outcomes16,17,18,19,20.

Although it was previously thought that anti-PD-1, for instance, blocks PD-1 signaling on the most terminally exhausted T cells, this ‘direct’ model has been called into question, suggesting that successful PD-1 blockade probably acts on less differentiated, ‘progenitor’-like T cells21, and high proportions of tumor-infiltrating, terminally exhausted T cells predict resistance to anti-PD-1 (refs. 21,22). Furthermore, although metabolic correlates, such as the capacity of a tumor cell to consume oxygen and generate hypoxia, can predict resistance to PD-1 blockade18, these environmental stressors have been shown to have disparate effects on T cell function. Hypoxia-inducible factor 1α (HIF-1α) and its negative regulators have been previously implicated in T cell activation23, and expansion of therapeutic cells under hypoxic conditions enhances tumor-killing capacity24. However, hypoxia is also clearly immunosuppressive, both in isolation and in vivo25. So, although metabolic stress and T cell exhaustion are linked, what remains unclear is whether T cell exhaustion promotes a program that causes a change in cellular metabolism, or whether metabolic insufficiency and stress directly contribute to the exhausted T cell’s fate.

In the present study, we show that, although hypoxia is a common metabolic stress present in tumor microenvironments, terminally exhausted, intratumoral T cells differentially experience more hypoxia, suggesting that those metabolic stresses have a role to play in their biology. Although hypoxia alone cannot induce T cell dysfunction, hypoxia alters how T cells respond to other signals. We developed a system to interrogate how persistent antigen stimulation and hypoxia have a role to play in T cell exhaustion, revealing that continuous antigenic stimulation under hypoxia results in rapid, severe T cell dysfunction consistent with exhaustion. This in vitro system allowed for interrogation of the biology underlying T cell exhaustion, suggesting that metabolic stress originating in the mitochondria can accelerate terminal differentiation, and that targeting these processes may improve immunotherapy for cancer.

Results

Exhausted intratumoral T cells experience severe hypoxia

Although hypoxia is common within tumors, it is unclear whether subsets of intratumoral T cells experience more hypoxia than others. We first phenotyped CD8+ tumor-infiltrating lymphocytes (TILs) in 8- to 10-mm (day 14) B16 melanoma tumors along the ‘spectrum’ of exhaustion (Extended Data Fig. 1a). Terminally exhausted T cells were defined as expressing high and sustained levels of PD-1 and co-expressing Tim-3 (Fig. 1a), possessing high LAG3 and Tox expression, low TCF1 expression, and failing to secrete interleukin 2 (IL-2) on restimulation (Fig. 1b–e). Antigen-specific responses were measured by transferring gp100-specific Pmel-1 T cells into B16-bearing mice, restimulating cells once they had reached terminal exhaustion (Extended Data Fig. 1b). Compared with lymph node (LN)-resident Pmel-1 T cells, tumor-resident, gp100-restimulated T cells were far less polyfunctional12,21 (Fig. 1f). Mice bearing B16 tumors were infused with pimonidazole, a hypoxia tracer, before sacrifice, revealing that terminally exhausted T cells experience the highest degree of hypoxia, using anti-pimonidazole and anti-HIF-1α antibodies, compared with other subsets (Fig. 1g,h). Thus, hypoxia is a major metabolic component of a tumor’s metabolic landscape, and terminally exhausted T cells experience elevated hypoxia compared with other subsets.

Fig. 1: Terminally exhausted CD8+ tumor-infiltrating T cells experience high levels of hypoxia.
figure 1

a, Flow cytogram of PD-1 and Tim-3 staining of wild-type (WT) CD8+ LNs and TILs. tdLN, tumor-draining LN. bd, Groups are labeled PD-1 low in LNs (L), and the TIL groups are PD-1, PD-1int, PD-1hi and PD-1hiTim-3+ for Lag-3 expression (n = 9 mice) (b), Tox expression (n = 5 mice) (c) and TCF1 expression (n = 5 mice) (d) as in b. MFI, mean fluorescence intensity. e, IL-2 production with 16 h of PMA/ionomycin stimulation (final 5 h with a protein transport inhibitor) (n = 7 mice). f, TNF versus IFN-γ production (n = 10) in gp100-specific Pmel T cell LNs and TILs with 16 h of gp100 stimulation (final 5 h with a protein transport inhibitor). g, Histogram overlays of HIF-1α staining in WT CD8+ LNs and TILs based on PD-1 and Tim-3 staining. h, Hypoxyprobe (anti-pimonidazole) representative staining and tabulation after in vivo pimonidazole injection into mice 1 h before sacrifice (n = 14 mice). Data represent three to five independent experiments. P < 0.05, P < 0.01, P < 0.001, P < 0.0001 by one-way ANOVA with Dunnett’s multiple comparisons test (be,h) or unpaired Student’s t-test (f). Error bars indicate the s.e.m.

T cells continuously stimulated under hypoxia appear exhausted

To determine whether hypoxia can drive exhaustion, we sought to model exposure to this stress in vitro. Although HIF-1α activation has been previously associated with upregulation of co-inhibitory molecules, expansion of T cells under hypoxic conditions results in increased cytolytic T cell function, suggesting that hypoxia’s inhibitory effects may depend on the presence of other signals23,24. As exhausted T cells were originally described in the context of persistent antigen, we hypothesized that hypoxia has distinct effects under conditions of continuous activation. OT-I T cell antigen receptor (TCR), transgenic (Tg) T cells were activated with cognate peptide overnight in atmospheric oxygen (21% O2, ‘normoxic’) conditions and then split into multiple conditions: cultured alone, in the presence of B16 cells or in he presence of B16-expressing ovalbumin (B16OVA). This co-culture occurred in normoxia or under conditions of average tumor hypoxia (1.5%) for 5 d with IL-2 (Extended Data Fig. 1c). T cells remained functional when co-cultured with antigen-expressing tumor cells under normoxia or with antigen-negative tumor cells under hypoxia, but persistent antigen under hypoxic conditions drove elevated PD-1 and Tim-3, a loss of polyfunctionality in cytokine production and poor expansion (Extended Data Fig. 1d-f).

OT-I T cells express an extremely avid TCR, and the tumor:T cell co-culture system produces other stresses, because tumor cells consume nutrients and secrete cytokines and toxic metabolites, exacerbated under hypoxic co-culture. To mitigate these and other unknown confounders coming from a tumor:T cell co-culture, we further reduced the system using purified polyclonal CD8+ T cells and ‘off-the-shelf’ stimulatory magnetic beads (coated with anti-CD3/anti-CD28). Notably, we used flow cytometrically sorted CD44+ cells for this assay to better mimic tumor-infiltrating T cells’ activated phenotype, although similar results were obtained using bulk CD8+ and naïve (CD62LhiCD44lo) T cells (data not shown). CD8+ T cells were stimulated with beads overnight in the presence of IL-2 and IL-12, then washed and split into four conditions. Cells were expanded with IL-2 (‘acute’ activation) or co-cultured with beads and IL-2 (‘continuous’ activation) for 5 d, with the medium changed regularly to prevent nutrient depletion. This post-activation expansion step occurred under normoxic or hypoxic conditions in round-bottomed plates, keeping cells and beads in close contact for continuous stimulation (Fig. 2a). These manipulations produced phenotypes observed in tumor:T cell co-culture: T cells maintained functionality in the presence of hypoxia or continuous stimulation alone, but a combination of both stressors produced an exhausted-like state—high co-inhibitory molecule expression (Fig. 2b–d), expression of CD39 (Fig. 2e), Tox (Fig. 2f)—and, importantly, one possessing severely limited polyfunctional cytokine production (Fig. 2g). Consistent with previous data, hypoxia alone enhanced T cell function, as a population but also on a per-cell basis. This dysfunctional state was maintained even after removal from stress conditions: 5 d of expansion, in or out of hypoxia, could not rescue the dysfunctional phenotype (Extended Data Fig. 2a). The combination also induced a metabolic state that we previously associated with exhausted T cells: mitochondrial respiratory capacity was severely repressed15 (Fig. 2h).

Fig. 2: Continuous activation under hypoxia induces an exhausted-like dysfunctional state in CD8+ T cells.
figure 2

a, Experimental scheme. CD44hi CD8+ T cells are sorted from B6 mice, then activated with anti-CD3/anti-CD28-coated magnetic beads + 25 U ml−1 of IL-2 and 10 ng ml−1 of IL-12 in normoxia (20% O2). After 24 h, cells were washed and expanded in IL-2, but placed into various culture conditions (removing beads or continuous co-culture with beads, under normoxia or 1.5% O2 hypoxia). b, Flow cytograms of PD-1 and Tim-3 staining in live CD8+ T cells generated as in a accompanied by quantification (n = 6). cf, Quantification of Lag-3 (n = 7; c), Tigit (n = 6; d), CD39 (n = 5; e) and Tox (n = 6; f) staining as in b. Stim, stimulation. g, Flow cytograms and tabulation of TNF and IFN-γ production after 16 h of PMA/ionomycin restimulation of live CD8+ T cells generated as in b (n = 5). h, Mitochondrial spare respiratory capacity (difference between basal OCR values and maximal OCR values after FCCP uncoupling) of T cells generated as in a (Acute Norm, n = 4; Cont Norm, n = 4; Acute Hypox, n = 3; Cont Hypox, n = 4). Graph represents one of three independent experiments. i, Flow cytometric quantification of OT-I T cells generated as in a, then adoptively transferred into B6 mice infected intraperitoneally with 1 × 106 plaque-forming units of vaccinia-OVA (VVOVA; n = 3 independent in vitro experiments, transferred into multiple mice). All data represent three to six independent experiments. P < 0.05, P < 0.01, P < 0.001, P < 0.0001 by one-way ANOVA with Dunnett’s multiple comparisons test. Error bars indicate the s.e.m.

To assess persistence and renewal capacity of these cells, OT-I T cells were acutely activated or continuously stimulated under hypoxia before transfer into mice infected with OVA-expressing vaccinia virus. Continuous stimulation under hypoxia limited the ability of T cells to respond to antigen in vivo (Fig. 2i), indicating limited renewal capacity, consistent with phenotypes observed in tumor-infiltrating exhausted T cells3,21.

We next assessed the effects of these stressors on proliferation. Use of proliferation dyes showed that continuous stimulation or hypoxia had modest impacts on proliferation (Extended Data Fig. 2b). However, these manipulations did impact cellular accumulation. Continuous stimulation increased cell death in late cell divisions, exacerbated under hypoxia. By day 6, continuous stimulation under hypoxia showed decreased accumulation (population doublings) compared with control conditions, in agreement with previous work3,21 (Extended Data Fig. 2c,d). Analysis of various markers per division suggested that the phenotypes observed were not driven by examining less proliferative or undifferentiated cells, but rather that cells were different on a per-division basis (Extended Data Fig. 2e–h). Thus, continuous stimulation under hypoxia drives T cells to proliferate and express co-inhibitory molecules, but be prone to cell death, in agreement with reports of exhausted T cells expressing Ki67 but not accumulating21,26.

Next, we determined cell-intrinsic consequences of continuous stimulation under hypoxia. BNIP3, a known target of hypoxic signaling, was elevated under hypoxia and mTORC1 (mammalian target of rapamycin complex 1) was activated with continuous stimulation (Fig. 3a). RNA sequenced from these four conditions was used to compare transcriptomes between cultures and previously published data. Principal component analyses showed all four populations were transcriptomically distinct (Fig. 3b). Analysis of differentially expressed genes (DEGs) within our dataset showed that continuous stimulation under hypoxia did not induce a combination of the ‘continuous stimulation’- and ‘hypoxia’-induced genes, but rather drove a distinct transcriptional profile (Fig. 3c and Supplementary Table 16). Gene Ontology suggested that shared gene clusters between continuous stimulation conditions were associated with negative regulation and metabolic changes (clusters 3 and 5) (Fig. 3c). Genes associated with continuous stimulation under hypoxia conditions (clusters 1 and 7) were predominantly driven by antiviral and antiproliferative genes, in line with previous exhaustion data3,27 (Fig. 3c).

Fig. 3: Continuous activation under hypoxia induces distinct intracellular programs from either stressor alone.
figure 3

a, Representative immunoblot of mTOR and HIF-1 signaling components in T cells stimulated in the four in vitro conditions in Fig. 2. b, Principal component (PC) analysis based on transcriptomic data from RNA-seq data of CD8+ T cells incubated in the four conditions as in Fig. 2 (each group n = 3). c, Heatmap and unsupervised hierarchical clustering of the average expression of 767 DEGs from pairwise comparisons (log2(fold-change) > 2, adjusted P < 0.05). Gene Ontology analysis of genes defining each cluster is represented in the accompanying bubble plot. d,e, Leading edge plots of GSEA of acute/normoxic versus continuous/hypoxic culture RNA-seq compared with published transcriptional profiles of progenitor versus terminally exhausted T cells in B16 TILs. FDR, false discovery rate. NES, normalized enrichment score. Heatmaps for this GSEA appear in Extended Data Fig. 3. All data represent three independent experiments.

Source data

Gene set enrichment analysis (GSEA) was performed using differentially expressed genes derived from published transcriptomes of less differentiated ‘progenitor-exhausted’ T cells versus PD-1hiTim-3+ ‘terminally exhausted’ T cells from B16 melanoma21 (Supplementary Table 7), overwhelmingly suggesting that continuous stimulation under hypoxia produced a transcriptional state associated with terminal exhaustion (Fig. 3d,e and Extended Data Fig. 3a,b). Analysis of key immunologic genes within each dataset revealed that, although continuous activation or hypoxia could each induce some transcriptional changes associated with exhaustion, the combination promoted sustained expression of a number of exhaustion-specific genes (Ccl3, Adora2b, Lag3, Tnfrsf9, Nr4a2, Prdm1) and the repression of stemness and survival genes (Tcf7, Bcl2, Il2) (Extended Data Fig. 3a,b). Thus, although T cells can respond adequately to continuous stimulation or hypoxic stress, the combination results in differentiation to an exhausted-like dysfunctional fate.

HIF-1α is dispensable for hypoxia-induced dysfunction

Although hypoxia induces a well-described transcriptional program via the von Hippel–Lindau (VHL)–HIF-1α machinery, HIF-1α-deficient T cells (from Hif1af/fCd4Cre mice) still upregulated inhibitory receptors PD-1 and Tim-3, and lost polyfunctionality during continuous stimulation with hypoxia (Extended Data Fig. 4a,b). Although our transcriptomic data (Fig. 3) suggested that hypoxic signaling was still intact (both hypoxic culture conditions induced expression of HIF targets such as glucose transporters), there were notable differences in both the breadth and the degree of upregulation when cells were experiencing acute versus continuous stimulation (Extended Data Fig. 4c,d). Consistent with a repressed mitochondrial program, hypoxic conditions produced T cells that were much more avid to take up the fluorescent glucose tracer 2-(N-(7-nitrobenz-2-oxa-1,3-diazol-4-yl)amino)-2-deoxyglucose compared with the acute/normoxia condition (Extended Data Fig. 4c). Thus, another non-HIF factor induced by either continuous stimulation or hypoxia might alter metabolic sufficiency and promote terminal exhaustion.

Blimp-1 represses PGC-1α under continuous stimulation

Persistent antigen is crucial for altered differentiation toward exhaustion, but its metabolic consequences are less clear. As continuous stimulation alone did not produce overt metabolic repression, persistent stimulation may activate a transcriptional repressor preventing metabolic reprogramming. Blimp-1 (encoded by Prdm1) is such a repressor, highly upregulated in terminally exhausted CD8+ T cells in B16 melanoma (Fig. 4a) and under conditions of continuous activation under hypoxia (Fig. 4b and Extended Data Fig. 3b). We have previously associated peroxisome proliferator-activated receptor-γ coactivator (PGC)-1α (encoded by Ppargc1a), a transcription coactivator coordinating mitochondrial biogenesis and antioxidant activity, with avoidance of exhaustion in T cells15. T cells continuously activated under hypoxia in vitro also repress PGC-1α expression (Fig. 4c). We next determined whether Blimp-1 could repress PGC-1α, finding that enforced Blimp-1 expression could repress the activity of a Ppargc1a promoter construct in 293T cells, without affecting viability (Fig. 4d and Extended Data Fig. 5a). CD8+ T cells from Prdm1f/fCd4Cre mice were resistant to dysfunction through continuous activation under hypoxia (Fig. 4e,f), failing to repress Ppargc1a expression under continuous activation (Fig. 4g).

Fig. 4: Continuous activation causes Blimp-1-mediated repression of PGC-1α.
figure 4

a, Blimp-1 staining in CD8+ LNs and TILs based on PD-1 and Tim-3 expression (n = 14 mice). b, Blimp-1 staining in in vitro–cultured CD8+ T cells stimulated continuously (C) under hypoxia, normalized in quantification to acute (A) stimulation in normoxia (n = 7). c, Ppargc1a (encoding PGC-1α) messenger RNA expression in cells cultured as in Fig. 2 (each group n = 3). d, Luciferase assay of 293T cells co-transfected with a plasmid containing the mouse PGC-1α promoter (Ppargc1ap) driving luciferase and mouse Prdm1 (encoding Blimp-1) at indicated ratios (n = 3). RLU, relative light units. e, PD-1 and Tim-3 staining in CD8+ T cells isolated from Prdm1f/fCd4cre mice or littermate controls stimulated continuously under hypoxia (n = 5). AN, acute normoxia; CH, cont hypoxia. f, IFN-γ production in T cells from e after 16 h of PMA/ionomycin restimulation (last 5 h with a protein transport inhibitor; n = 3). g, Ppargc1a mRNA expression in WT or Blimp-1-deficient T cells acutely or continuously stimulated in vitro (n = 5 mice per group). h, MitoTracker FM staining in PD-1hiTim-3+ TIL, CD8+ T cells from B16-bearing animals with an inducible CD8-specific deletion of Blimp-1 (Prdm1f/f E8I-Cre-ERT2 Rosa26-LSL-TdTomato (termed Prdm1iKO); WT n = 6 mice, Prdm1iKO n = 10 mice). i, IL-2 and TNF production in CD8+ TILs after 16 h of PMA/ionomycin restimulation, as in h (WT n = 9 mice, Prdm1iKO n = 9 mice). All flow data represent three to six independent experiments. P < 0.05, P < 0.01, P < 0.001, P < 0.0001 by one-way ANOVA with Dunnett’s multiple comparisons test (a–d), paired Student’s t-test (e and f), one-way ANOVA with repeated measures (g) and unpaired Student’s t-test (h and i). Error bars indicate the s.e.m.

We next determined the requirement by Blimp-1 for metabolic repression in vivo. Transfer of naive, Blimp-1-deficient OT-I TCR, Tg T cells into mice bearing B16OVA tumors (Extended Data Fig. 5b) showed that Blimp-1 was required to repress mitochondrial mass in tumor-infiltrating T cells (Extended Data Fig. 5c). Previous reports have showed that Blimp-1-deficient T cells do not effectively differentiate into exhausted T cells, distorting our analysis using T cells that are already Blimp-1 deficient6,28. To determine the role for Blimp-1 in T cells already terminally exhausted (and part of the endogenous anti-tumor T cell response), we utilized an inducible, CD8-specific Cre (E8I-CreERT2 generated by D. Vignali) bred to Prdm1f/f animals29. These mice, termed Prdm1iKO, bore a Rosa26-LSL-TdTomato recombination reporter, which showed that most CD8+ T cells in the tumor had Cre activity after tamoxifen treatment (Extended Data Fig. 5d). We injected B16 cells into untreated mice and let tumors reach 5 mm in diameter (thus harboring terminally exhausted CD8+ T cells). These tumor-bearing Prdm1iKO mice were then treated with tamoxifen for 5 d, and recombined, TdTomato+ PD-1hiTim-3+ TILs were immunometabolically analyzed (Extended Data Fig. 5e–g). Tamoxifen treatment did not alter proportions of co-inhibitory molecule-expressing, terminally exhausted TILs, and the vast majority activated the Rosa26 reporter and deleted Blimp-1 (Extended Data Fig. 5e–g). Loss of Blimp-1 in PD-1hiTim-3+ TILs resulted in recovery of mitochondrial mass and polyfunctionality as read-out by IL-2 and tumor necrosis factor (TNF) production (Fig. 4h,i). Notably, consistent with our in vitro data, the HIF-1α axis was unchanged by Blimp-1: HIF-1α-deficient CD8+ T cells in vivo still upregulated Blimp-1 in the hypoxic tumor microenvironment, and Blimp-1-deficient CD8+ T cells in vivo still retained HIF-1α expression (Extended Data Fig. 5h,i). Thus, continuous activation upregulates Blimp-1, altering how T cells experience metabolic stress via repression of PGC-1α-mediated metabolic reprogramming.

ROS drive exhaustion by acting as a phosphatase inhibitor

We next sought the underlying mechanism driving dysfunction induced through inadequate responses to hypoxia. To understand how PGC-1α represssion was affecting TIL function, we retrovirally overexpressed PGC-1α (PGC-1αOE) in Pmel T cells and transferred them into B16-bearing mice, analogous to a system we previously used in OT-I T cells to enhance adoptive cell therapy of B16-OVA15. RNA-sequencing (RNA-seq) analysis of PGC-1αOE, tumor-infiltrating, Pmel-1 T cells revealed expectedly increased genes involved in mitochondrial metabolism (Supplementary Table 8). GSEA using progenitor and terminal exhaustion datasets (analogous to analysis in Fig. 3) showed that PGC-1α overexpression prevented expression of genes associated with terminal exhaustion (Extended Data Fig. 6a,b). Notably, PGC-1α overexpression did not result in T cells bearing a ‘progenitor-exhausted’ signature, suggesting that metabolic reprogramming resulted in altered rather than slowed differentiation (Extended Data Fig. 6a,b). Thus, consistent with our previous data, PGC-1α overexpression resulted in mitochondrial reprogramming and differentiation away from the exhausted phenotype.

Pathway analysis also delineated pathways upregulated by PGC-1α. Consistent with their enhanced effector function, ‘graft-versus-host disease’ genes were upregulated most highly (Extended Data Fig. 6c, pathway 1). As PGC-1α functions through engaging mitochondrial biogenesis and upregulation of antioxidant enzymes, regulation of ROS was highly enriched in PGC-1αOE compared with controls (Extended Data Fig. 6c, pathway 2). Indeed, tumor-infiltrating, PGC-1αOE, Pmel-1 T cells showed decreased mitochondrial ROS on tumor infiltration (Fig. 5a), suggesting that PGC-1α acts, in part, to mitigate ROS on tumor infiltration.

Fig. 5: Continuous activation under hypoxia increases mitochondrial ROS, which is sufficient to drive an exhausted-like dysfunctional program.
figure 5

a, MitoSOX (mitochondrial superoxide) staining of adoptively transferred, B16-infiltrating, Pmel, Thy1.1+ T cells transduced with a PGC-1α retroviral expression vector (PGC-1αOE) or empty vector (EV) control (EV n = 8 mice, PGC-1αOE n = 10 mice). b, MitoSOX staining in CD8+ LNs and TILs based on PD-1 and Tim-3 expression (quantification normalized to LN cells; n = 16 mice). c, MitoSOX staining in continuous stimulation under hypoxia, normalized to acute stimulation in normoxia (each group n = 6). d, MitoSOX staining of CD8+ T cells activated overnight and expanded in IL-2 and 0.04 μM mitochondrial complex III inhibitor antimycin A (AA), 0.4 μM complex I inhibitor rotenone (Rot) or simultaneous combination of the two for 3 d (n = 7). e, As in d but using the cellular ROS detector DCFDA (n = 5). f, Percentage of PD-1hiTim-3+ cells as in d cultured for 6 d (n = 4). g, IFN-γ and TNF production in T cells generated as in d for 7 d, after stimulation with anti-CD3/anti-CD28 (n = 4). h, MitoSOX staining of CD8+ T cells activated overnight and then expanded in IL-2 and antimycin A, 10 mM NAC or simultaneous combination of the two for 3 d (each group n = 4). Ctrl, control. i, IFN-γ and TNF production in T cells in h after stimulation with anti-CD3/anti-CD28 overnight (n = 3). j, Percentage of PD-1hiTim-3+ cells in cells as in h (n = 6). k, MitoSOX MFI (control n = 5, NAC n = 5). l, IFN-γ and TNF production (control n = 5, NAC n = 6). m, PD-1 and Tim-3 staining (control n = 4, NAC n = 7); km are in continuous stimulation under hypoxia ±10 mM NAC. All data represent three to seven independent experiments. P < 0.05, P < 0.01, P < 0.001, P < 0.0001 by one-way ANOVA with Dunnett’s multiple comparisons test (bj) or unpaired Student’s t-test (a and km). Error bars indicate the s.e.m.

Mitochondrial dysfunction and low oxygen tension can promote ROS through several mechanisms, including reverse electron transport-producing superoxide at mitochondrial complex I (ref. 30). Examination of endogenous B16 TILs showed that terminally exhausted T cells harbor significantly higher amounts of mitochondrial (mt)ROS (Fig. 5b), in line with previous findings31,32,33. Continuous activation under hypoxia in vitro also produced high levels of mtROS (Fig. 5c), suggesting that ROS may be a driver of T cell exhaustion. To interrogate effects of ROS in isolation, we peptide activated OT-I Tg CD8+ T cells for 24 h, then expanded the cells in low concentrations of antimycin A (0.04 μM), a mitochondrial complex III inhibitor established to induce mtROS34. Antimycin A–induced mtROS production (as well as the consequent increase in cellular ROS) in a manner dependent on mitochondrial complex I activity (inhibited by the complex I inhibitor rotenone) (Fig. 5d,e). Notably, rotenone alone at higher doses can transiently induce mtROS35 (Extended Data Fig. 6d). In the present study, low, nontoxic doses of the drugs were used to culture T cells for several days. Indeed, activating T cells (24 h) and then expansion in the presence of antimycin A resulted in an exhausted-like dysfunction: high co-inhibitory molecule expression and decreased polyfunctionality (interferon-γ (IFN-γ) and TNF production) (Fig. 5f,g). We observed small but significant increases in TNF single producers under antimycin A treatment insensitive to rotenone treatment. Antimycin A treatment had similar proliferative consequences as continuous stimulation under hypoxia: the proliferative capacity of T cells was modestly affected but accumulation as measured by overall fold expansion was decreased (Extended Data Fig. 6e,f). Remarkably, addition of rotenone, which when added to antimycin A treatment collapses the entire electron transport chain, rescued the dysfunction, suggesting that the exhaustion observed was not caused by loss of mitochondrial function but rather due to mitochondrial stress and subsequent ROS (Fig. 5d–g and Extended Data Fig. 6e,f).

To further address the role of ROS driving dysfunction, we employed N-acetylcysteine (NAC), a cell-permeable antioxidant that neutralizes ROS (Fig. 5h). NAC could prevent dysfunction induced by either antimycin A or continuous stimulation under hypoxia (Fig. 5i–m). NAC treatment only modestly improved accumulation effects observed from antimycin A or continuous stimulation under hypoxia (Extended Data Fig. 6g,h), suggesting that NAC’s effects were not due to alterations in survival but rather to altering differentiation.

To further delineate how progressive loss of functional mitochondria might induce a dysfunctional state, we used a parallel approach in which mitochondrial activity could be progressively repressed during T cell expansion. We generated Rho0 (lacking detectable mtDNA) primary T cells through culture of OT-I T cells in the presence of low-dose (50 ng ml−1) ethidium bromide (supplemented with pyruvate and uridine; Extended Data Fig. 7a). This well-established experimental technique selectively depletes mtDNA (encoding a subset of mitochondrial genes, including subunits of the electron transport chain), producing cells lacking mtDNA but retaining mitochondrial mass36. Rho0 OT-I T cells harbor high levels of ROS (Extended Data Fig. 7b), despite having no detectable mitochondrial OXPHOS activity, mtDNA or various respiratory components (Extended Data Fig. 7c–e). Notably, T cells withstand ethidium bromide treatment, having no expansion defect after 10–12 d of culture (Extended Data Fig. 7f), consistent with Tfam-deficient T cells37. Rho0 T cells expressed elevated levels of PD-1, Tim-3 and TIGIT, failing to kill target cells when assayed in vivo (Extended Data Fig. 7g,h). Importantly, generation of Rho0 T cells in the presence of NAC rescued cytokine polyfunctionality and lowered expression of Tox (Extended Data Fig. 7i-k). Taken together, our data suggest that heightened ROS production induced by mitochondrial dysfunction drives a phenotype consistent with T cell exhaustion.

Next, we explored how ROS drives dysfunction. ROS are crucial components of cellular biology, and active participants in critical chemical reactions38. In T cells, ablation of mtROS via genetic deletion of mitochondrial complex III subunit Rieske iron sulfur protein (RISP) results in failed NFAT signaling, so we reasoned, conversely, that persistently elevated ROS might drive continuous NFAT signaling39, a transcriptional circuit known to drive exhaustion. Notably, ROS and their cellular byproducts are potent inhibitors of tyrosine phosphatases38, and indeed culture of T cells with antimycin A resulted in persistent and elevated tyrosine phosphorylation (Fig. 6a), which was blocked with rotenone or NAC (Fig. 6a,b). Immunoblot analysis revealed that the absolute amounts of tyrosine phosphorylation increased, but also several different proteins were differentially phosphorylated under antimycin A-induced ROS, in a manner alleviated by NAC treatment (Fig. 6b). Phosphotyrosine signaling is a major player in T cell biology downstream of many cell surface interactions, including TCR signaling. Use of TCR-reporter Nur77-GFP mice showed that antimycin A treatment during expansion induced green fluorescent protein (GFP) expression in the absence of any TCR signals; although lower than signal induced by continuous stimulation conditions, GFP from antimycin A treatment matched the phenotype of T cells cultured with the tyrosine phosphatase inhibitor orthovanadate (50 μm) (Fig. 6c and Extended Data Fig. 8a). Phosphotyrosine cascades promote nuclear accumulation of NFAT1, and ROS induction alone (via actinomycin A) drove increased NFAT1 in a rotenone-inhibited manner (Fig. 6d), consistent with previous data suggesting that exhausted T cells harbor persistent NFAT1 signaling3,7. Expanding CD8+ T cells in orthovanadate produced a similar dysfunctional state: heightened co-inhibitory molecule expression, decreased cell accumulation and cytokine production, and expression of Blimp-1 and Tox (Extended Data Fig. 8). Thus, heightened ROS produced under exposure to continuous stimulation under hypoxia promotes a phenotype of chronic signaling, which may reinforce transcriptional machinery driving exhaustion. This may underlie the inability to ‘rescue’ terminally exhausted T cells even after a persistent antigenic signal has been removed.

Fig. 6: Persistent mtROS increases elevation of phosphotyrosine signaling and NFAT localization.
figure 6

a, Global phosphotyrosine staining in T cells activated overnight and then expanded in IL-2 and 0.04 μM mitochondrial complex III inhibitor antimycin A, 0.4 μM complex I inhibitor rotenone or a simultaneous combination of the two for 3 d (each group n = 5). b, Phosphotyrosine staining of CD8+ T cell lysates from cells activated overnight and then expanded in IL-2 and 0.04 μM mitochondrial complex III inhibitor antimycin A, 10 mM NAC or a simultaneous combination of the two for 3 d. c, GFP expression of Nur77-GFP CD8+ T cells in acutely activated, continuously activated, antimycin A or 50 μM sodium orthovanadate culture conditions. d, NFAT localization in representative control and antimycin A–expanded T cells (day 6) as well as actin and DNA staining (control n = 55, antimycin A n = 75, rotenone n = 70, antimycin A + rotenone n = 67). Scale bars = 2 μm. In the quantification graph (right), each dot represents one cell. All data represent three to six independent experiments. P < 0.01, P < 0.001, P < 0.0001 by one-way ANOVA with Dunnett’s multiple comparisons test (a and d). Error bars indicate the s.e.m.

Source data

Mitigating ROS or hypoxia alleviates T cell exhaustion

Our data suggested that elevated ROS could drive terminal exhaustion, and PGC-1α overexpression in TCR Tg T cells reduced ROS. As PGC-1α promotes multiple downstream programs, we asked whether specifically mitigating ROS in a T cell–intrinsic manner could also protect against dysfunction. Pmel-1 T cells were transduced with Gpx1, a glutathionine peroxidase and known target of PGC-1α capable of acting on many ROS species40, before transfer into B16-bearing animals. Similar to PGC-1α-transduced cells, GPX1-overexpressing T cells were resistant to ROS accumulation in tumors (Fig. 7a). GPX1-overexpressing TIL T cells retained functionality, producing increased IFN-γ (Fig. 7b). Thus, reducing ROS in a cell-intrinsic manner protects T cells from tumor-induced exhaustion.

Fig. 7: Reducing ROS or hypoxia exposure alters T cell differentiation to exhaustion and improves response to immunotherapy.
figure 7

a, DCFDA staining of adoptively transferred, B16-infiltrating, Pmel Thy1.1+ cells. T cells retrovirally transduced with GPX1 or vector control (EV n = 8, GPX1OE n = 8). b, IFN-γ production from cells as in a after 16 h of gp100 restimulation (last 5 h with protein transport inhibitor; EV n = 8 mice, GPX1OE n = 9 mice). c, Oxygen consumption rate of B16-F10 or Ndufs4-deficient B16 (ND4) mice (each group n = 4). d, Hypoxyprobe staining in CD8+ T cells infiltrating WT or Ndufs4-deficient B16-F10 mice after intravenous pimonidazole injection 1.5 h before sacrifice (WT n = 7 mice, Ndufs4-deficient n = 7 mice). e, PD-1 and Tim-3 staining in CD8+ TILs from WT or Ndufs4-deficient B16-F10 mice, accompanied by quantification of terminally exhausted or progenitor-exhausted cells (WT n = 6 mice, Ndufs4-deficient n = 8 mice). f, TNF and IFN-γ production in CD8+ TILs from WT or Ndufs4-deficient B16-F10 mice after 16 h of PMA/ionomycin restimulation (last 5 h with protein transport inhibitor; WT n = 9 mice, Ndufs4-deficient n = 6 mice). g, Hypoxia staining of B16 tumor sections of mice treated with vehicle or 10 mg kg−1 axitinib for 3 d. Scale bar = 500 μm. h, Hypoxia staining of T cells from LNs and TILs of axitinib- or vehicle-treated, B16-bearing mice (vehicle n = 8 mice, axitinib n = 8 mice). i, PD-1 and Tim-3 staining in CD8+ TILs from axitinib- or vehicle-treated B16-bearing mice (vehicle n = 8 mice, axitinib n = 8 mice). j, IFN-γ and TNF production in PMA/ionomycin-restimulated CD8+ LNs and TILs in B16-bearing mice treated with axitinib or vehicle control (vehicle n = 8 mice, axitinib n = 8 mice). k, Tumor growth (left) and survival (right) of B16-bearing mice treated three times weekly with axitinib or vehicle alone or with combination anti-PD-1 + anti-CTLA-4. Therapy began when mice had a palpable tumor (day 5, arrow). Right: survival curve for mice (control n = 7, axitinib n = 9, αPD-1 + αCTLA-4 n = 8, axitinib + αPD-1 + αCTLA-4 n = 10). Data represent two or three independent experiments. CR, complete response. P < 0.05, P < 0.01, P < 0.001, P < 0.0001 by unpaired Student‘s t-test (af and i), one-way ANOVA (h and j) or log(rank) test (k). Error bars indicate the s.e.m.

We next determined whether the trajectory of T cell differentiation toward exhaustion could be altered through changes in the tumor microenvironment. As hypoxia seemed to drive an altered differentiation state through generation of ROS, we sought to use tumor cells that would consume less oxygen and produce less hypoxia. Thus, we implanted mice with B16 melanoma engineered via CRISPR–Cas9 to lack a critical structural component of mitochondrial complex I (Ndufs4) or parental B16-F10 cells (Extended Data Fig. 9a). Consistent with our previous findings targeting Ndufs4 using RNA interference18, Ndufs4-deficient tumors had no appreciable oxygen consumption in vitro and created less hypoxia in vivo (Fig. 7c,d). Also, consistent with RNA interference, Ndufs4-deficient cells grew comparably in vivo (Extended Data Fig. 9b).

We next assessed the phenotype of endogenous infiltrating CD8+ T cells of parental and Ndufs4-deficient (less hypoxic) tumors. Consistent with our previous findings, there was a modest but significant decrease in infiltrating CD8+ T cells in Ndufs4-deficient B16 mice (Extended Data Fig. 9c). Of these endogenous infiltrating CD8+ T cells, a smaller proportion of them progressed to terminal exhaustion (Fig. 7e). However, although these cells still expressed multiple co-inhibitory molecules, PD-1hiTim-3+ T cells infiltrating Ndufs4-deficient tumors showed increased polyfunctionality (Fig. 7f).

We purified and sequenced RNA from endogenous, terminally exhausted TILs from parental or Ndufs4-deficient B16 tumors. RNA-seq of PD-1hiTim-3+ T cells from Ndufs4-deficient versus control tumors revealed a distinct transcriptional profile: decreased expression of genes involved in exhaustion and increased expression of effector or memory-like genes, including regulation of ROS and hypoxia (Extended Data Fig. 9d–f and Supplementary Table 9). Thus, in less hypoxic environments, T cells have an altered differentiation pattern reminiscent of effector/memory-like cells, even within compartments typically defining terminally exhausted T cells (PD-1hiTim-3+).

We next asked whether hypoxia mitigating therapy could alter T cell differentiation to be beneficial for immunotherapy. Beyond genetic approaches, hypoxia is difficult to target in vivo. Although tumor cell energetics certainly drive the altered metabolic landscape in cancer, vascular endothelial growth factor (VEGF)-dependent angiogenesis and changes in tumor vasculature also contribute to generation of a hypoxic environment41. Axitinib is a tyrosine kinase inhibitor with nanomolar affinity for VEGF receptors VEGFR1, -2 and -3. Although anti-angiogenics were designed to ‘starve’ a tumor, inhibiting angiogenesis completely at high doses, low doses of anti-angiogenics can correct tortuous vasculature and lower tumor hypoxia, a strategy used in renal cancer in combination with immunotherapy42.

Consistent with previous studies, treating B16-bearing mice with low-dose (10 mg kg−1) axitinib lowered total tumor and T cell hypoxia as measured by pimonidazole (Fig. 7g,h). Intratumoral T cells were phenotypically less exhausted (Fig. 7i) and more polyfunctional (Fig. 7j), suggesting that, by targeting VEGFRs and lowering hypoxia, T cells might be more responsive to immunotherapy. Treatment of tumor-bearing mice with low-dose axitinib in vivo sensitized B16-bearing mice to cytotoxic T-lymphocyte-associated protein (CTLA)-4 and PD-1 blockade, decreasing tumor burden and improving survival (Fig. 7k). By targeting the hypoxic nature of the tumor microenvironment, T cells do not differentiate to terminal exhaustion and retain responsiveness to checkpoint blockade.

Discussion

The term ‘exhaustion’ has been used since initial reports in chronic infection, but drivers of the phenotype (save antigen persistence) have remained nebulous12. Inflammation, cytokines, regulatory T cells and nonimmune signals have been proposed3, but few approaches have been able to create exhausted-like T cells in isolation. We and others have found metabolic underpinnings to exhaustion: loss of mitochondrial capacity, glucose metabolism defects and exposure to hypoxia15,16,17,18,32,43,44. In the present study, we show that metabolic insufficiency directly induces T cell exhaustion, rather than merely characterizing exhausted T cells. Our findings do not necessarily suggest that metabolic insufficiency is the sole agent of T cell exhaustion, but highlight the need to consider the metabolic status of a T cell and its environment as modifiers of immunologic signals. In both PGC-1α-overexpressing T cells and in less hypoxic tumors, although T cells may appear exhausted ‘on the surface’ (increased co-inhibitory molecule expression), if metabolically sufficient, they may carry a dramatically superior phenotype.

Hypoxic exposure during continuous stimulation produced cells resembling exhausted T cells: high co-inhibitory molecule expression, low metabolic sufficiency and polyfunctionality, and a transcriptome with features of terminally exhausted TILs. Admittedly, our transcriptional profiling did not suggest we had generated real exhausted TILs: comparison of the entire transcriptome (beyond DEGs) did not overlap with transcriptional profiles from tumor-infiltrating T cells. This finding is not surprising, given the relatively simple signals provided to T cells in vitro but serves as a platform for future work. Comparing transcriptional profiles of in vitro exhausted T cells with those found in tumors can identify immunologic or metabolic modulations to improve the assay.

Our study further implicates Blimp-1 in the process of exhaustion. Similar to many exhaustion-related transcription factors, Blimp-1 deletion does not ‘rescue’ T cells from exhaustion but prevents the phenotype or affects survival. An inducible, CD8-specific deletion strategy in cells that had reached terminal exhaustion revealed that Blimp-1 continually represses the plasticity (metabolic and otherwise) of exhausted cells. Identification of Blimp-1’s interactome and genomic binding will be key for ongoing studies.

Much of the prior work in T cell differentiation has been conducted by manipulating the PHD–VHL–HIF-1α axis genetically rather than using hypoxia directly. Regardless, it was surprising that Hif1a deficiency had no effect on the dysfunctional phenotype induced through in vitro continuous activation under hypoxia, emphasizing that low oxygen tension elicits signals beyond the HIF-1α machinery. Hypoxia can directly affect mitochondrial function, causing reverse electron transport and buildup of ROS30. However, it is unlikely that Hif1α is completely uninvolved in T cell exhaustion. Loss of Hif1α’s negative regulator Vhl drives increased co-inhibitory marker expression, but improved cytokine functionality in the setting of persistent viral infection23. As Hif1α is not only stabilized under low oxygen tension but activated through other inputs such as T cell activation45, future studies will dissect Hif1α-dependent responses to hypoxia and continuous activation in vivo.

Although it was surprising that it was not loss of mitochondrial activity but the presence of dysfunctional mitochondrial products that ultimately drove dysfunction, healthy mitochondria probably produce beneficial intermediates for long-lived T cell function. Mitochondria produce energetic currency (ATP) but also buffer calcium, produce biosynthetic intermediates and generate reactants critical for epigenetic changes such as histone acetylation and demethylation46. Furthermore, ROS also serve as modulators of cellular signaling in T cells: as a phosphatase inhibitor, stress-induced ROS created a cellular environment mimicking constitutive signaling, which may serve to ‘feed forward’ and potentiate differentiation. Although these mechanisms may have evolved induction of tolerance in hypoxic tissues, if ROS is the central driver, other metabolic and physiologic states may play a role. ROS can be generated in a variety of cellular conditions, including hyperoxia, inflammation and nutrient stress47,48,49. Furthermore, ROS can also inhibit dioxygenase reactions of demethylases, alter metabolic enzymes and induce DNA damage, so, although phosphatase inhibition is certainly ROS mediated, ROS elevation in T cells probably contributes to exhaustion in multiple ways.

During the review of the present study, a similar approach was reported (OT-I T cells co-cultured with B16OVA cells) exploring metabolic consequences of persistent stimulation in vitro50, bearing remarkable parallels to our own: persistent antigen alters T cell mitochondrial sufficiency, mitochondrial dysfunction drives T cell exhaustion and ROS neutralization can slow terminal differentiation. Our data place transcriptional repression via Blimp-1 as a major mechanistic determinant of metabolic reprogramming downstream of persistent stimulation and provide additional insight into the interplay between metabolic reprogramming and altered differentiation in vivo. Furthermore, our study highlights the role of hypoxia as a critical ‘accelerator’ of this process, suggesting that mitigation of hypoxia may be an actionable strategy to improve immunotherapy. We posit that exposure to hypoxia, during initial activation or in response to immunotherapy, may exacerbate the development of exhaustion rather than a long-lived fate like memory. Our previous studies, showing that hypoxic tumors are less sensitive to immunotherapy in both mice and humans, support this hypothesis18, and in the present study we show that mitigation of hypoxia with VEGFR inhibition also improves immunotherapeutic outcomes. Of course, axitinib, as a tyrosine kinase inhibitor, may act on other molecules present within tumor microenvironments51,52. VEGFRs can be expressed even by T cells regulating their function. Our data suggest a key mechanism of action of VEGF antagonism during immunotherapy (US Food and Drug Administration approved in renal cell carcinoma and under investigation in other cancers) may be limiting exposure of tumor-infiltrating T cells to metabolic stress. Finally, our study highlights the importance of accounting for metabolism in the design of immunomodulatory therapies and supports a role for metabolic reprogramming and modulation to improve immunotherapy for cancer.

Methods

Mice

Animal work was done in accordance with the Institutional Animal Care and Use Committee of the University of Pittsburgh. All mice were on a C57BL/6 background, used at age 6–8 weeks, were both male and female, and housed in specific pathogen-free conditions before use. C57BL/6, SJ/L (Thy1.1), Cd4Cre, Tg(TcraTcrb)1100Mjb/J (OT-I), B6.Cg-Thy1a/Cy Tg(TcraTcrb)8Rest/J (Pmel) and Hif1af/f mice were obtained from the Jackson Laboratory. Prdm1f/f mice were a gift from A. Poholek (University of Pittsburgh). E8I-CreERT2 GFP x Rosa26-LSL-TdTomato mice and spleens from Tg(Nr4a1-EGFP/cre)820Khog/J (Nur77-GFP reporter) mice were a gift from D. A. A. Vignali (University of Pittsburgh).

Cell lines, antibodies and other reagents

B16-F10 and 293T mice were obtained from the American Type Culture Collection. B16OVA (MO5) mice were obtained from P. Basse and L. Falo (University of Pittsburgh). Platinum-E (Plat-E) was obtained from L. Kane (University of Pittsburgh). Ndufs4-deficient B16-F10 mice were generated using transient transfection of Cas9–GFP and Ndufs4-targeted guide RNA, single-cell sorting of GFP+ cells, in vitro culture and screening using extracellular flux analysis, and selection of a clone that had lost GFP expression and lacked Ndufs4 protein. Vaccinia-OVA was generated by J. R. Bennink53 and provided by J. Powell (Johns Hopkins University). Zombie viability dye and anti-CD8 (53.6.7), anti-Tim-3 (RMT3-23), anti-PD-1 (catalog no. 29F.1A12), anti-Lag3 (catalog no. C9B7W), anti-TIGIT (catalog no. 1G9), anti-CD39 (catalog no. Duha59), anti-IL-2 (catalog no. JES6-5H4), anti-TCF1 (catalog no. W16175A), anti-CD44 (catalog no. IM7), anti-Thy1.1 (catalog no. OX-7), anti-IFN-γ (catalog no. XMG1.2) and anti-TNF (catalog no. MP6-XT22) antibodies were obtained from BioLegend. Anti-Hif1α (catalog no. 241812) was obtained from R&D Systems. Anti-NFAT1 antibody (catalog no. 25A10.D6.D2) was obtained from Abcam. Cell Trace Violet, CFSE, MitoSOX Red Mitochondrial Superoxide Indicator, CellROX Deep Red Reagent, CM-H2DCFDA, MitoTracker Deep Red FM, phalloidin, 4′,6-diamidino-2-phenylindole (DAPI), Dynabeads Mouse T-Activator CD3/CD28 for T cell expansion and activation, ethidium bromide solution, anti-Tox (TXRX10) and anti-Blimp-1 (5E7) antibody were obtained from Fisher. Uridine, tamoxifen, oligomycin, carbonyl cyanide-4-(trifluoromethoxy)phenylhydrazone (FCCP), 2-deoxyglucose, rotenone, antimycin A, phorbol 12-myristate 13-acetate (PMA) and ionomycin were obtained from Sigma. Golgiplug was obtained from BD Biosciences. NDUFS4 (E-4) and β-actin (C4) were obtained from Santa Cruz Biotechnology. Phosphotyrosine antibody (P-Tyr-100) was obtained from Cell Signaling. SIINFEKL peptide was obtained from AnaSpec. IL-2 and IL-12 were obtained from PeproTech. In vivo hypoxia staining was detected with an antibody to pimonidazole, obtained from Hypoxyprobe. Sodium orthovanadate was obtained from New England Biolabs. Axitinib was obtained from Cayman Chemical. In vivo antibodies anti-PD-1 (clone J43), anti-CTLA-4 (clone 9H10) and isotype controls were obtained from Bio X Cell.

Flow cytometry

For live flow, metabolic dyes were loaded into cells by culturing T cells in serum-free medium at 37 °C for 20 min, followed by antibody staining in fluorescence-activated cell sorting (FACS) buffer (phosphate-buffered saline (PBS) + 2% fetal bovine serum (FBS)) on ice. For intercellular transcription factor staining, surface staining was done with FACS buffer on ice, followed by 4% paraformaldehyde (PFA) fixation (to preserve Tim-3 staining), then by Foxp3 fix/permeabilization (Fisher) and intercellular antibodies were stained overnight using perm wash. For cytokine staining after stimulation, surface staining was done with FACS buffer on ice, followed by 4% PFA fixation (to preserve Tim-3 staining, if included), then Cytofix/Cytoperm (Fisher) and intercellular antibodies were stained overnight using perm wash.

Metabolism assay

Using a Seahorse XFe96 Bioanalyzer (Agilent), in vitro cultured CD8+ T cells (1.0 × 105 per well) were plated on Seahorse culture plates in medium consisting of minimal, unbuffered Dulbecco’s modified Eagle’s medium supplemented with and 10 mmol l−1 of glucose, 1 mmol l−1 of pyruvate and 2 mmol l−1 of glutamine. Basal oxygen consumption rates (OCRs) were taken for 30 min. Cells were stimulated with 2 μmol l−1 of oligomycin, 0.5 μmol l−1 of FCCP, 10 mmol l−1 of 2-deoxyglucose and 0.5 μmol l−1 of rotenone/antimycin A to obtain maximal respiratory and control values.

Immunoblotting analysis

Immunoblotting was performed as previously described54. Actin (C4) and PGC-1α (H-300) antibodies for immunoblots were obtained from Santa Cruz Biotechnology, p-AKT (S473) (D9E), pS6 (S235/236) (D57.2.2E), BNIP3 no. 3769, LCK no. 2752 and p-Tyr-100 no. 9411 from Cell Signaling, and CV-ATP5A, CIII-UQCRC2 and CII-SDHB from Total OXPHOS Rodent WB Antibody Cocktail from Abcam. Immunoblots were detected via standard secondary detection and chemiluminescent exposure to film. Digitally captured films were analyzed densitometrically using ImageJ software.

T cell isolations from LNs and tumors

LN T cells were isolated from 6- to 8-week-old B16-bearing mice and mechanically disrupted. To obtain single-cell suspensions of TILs, we injected whole tumors repeatedly using 20G needles with 2 mg ml−1 of collagenase type IV, 2 U ml−1 of hyluronidase (Dispase) and 10 U ml−1 of DNase I (Sigma) in buffered RPMI with 10% FBS, and incubated them for 25 min at 37 °C. Tumors were mechanically disrupted between frosted glass slides and filtered to remove particulates, then vortexed for 30 s. In some experiments, mice were injected intravenously with pimonidazole (80 mg kg−1, Hypoxyprobe) in PBS 1.5 h before sacrifice. Pimonidazole was visualized using anti-pimonidazole antibodies (Hypoxyprobe) after 10 min of 4% PFA fixation, followed by Foxp3 Fix/perm (eBioscience) permeabilization.

T cell transduction and retroviral expression

PGC-1α was originally generated by B. Spiegelman, obtained from Addgene (plasmid no. 1026)55, and cloned into a murine stem cell virus (MSCV)-driven retroviral expression vector, which also encodes an internal ribosome entry site (IRES)–GFP cassette, from D. A. A. Vignali. GPX1 was obtained from Origene (catalog no. NM_008160) and cloned into an MSCV-driven retroviral expression vector that also encodes an IRES–amitrine cassette, from D. A. A. Vignali. The vectors were transiently transfected into a Plat-E Retroviral Packaging Cell Line. Freshly isolated CD8+ Pmel T cells were stimulated with 5 μg ml−1 of anti-CD3 + 2 μg l−1 of anti-CD28 in the presence of 50 U ml−1 of IL-2 for 24 h. Retroviral supernatants were harvested, and filtered, and supplemented with 6 μg ml−1 of polybrene. Pmel T cells were spinduced with the retroviral supernatant for 120 min at 2,000 r.p.m. Cells were expanded and sorted by fluorescent marker before adoptive transfer.

Real-time PCR

RNA was isolated via RNeasy Plus Mini Kit (QIAGEN). Complementary DNA was reverse transcribed using the High Capacity cDNA Reverse Transcription Kit (Applied Biosystems). MtDNA was isolated via Trizol prep using the genomic fraction. Real-time PCR was performed with primers for Ndufs4 (forward: 5′-AACGGATCCACAGCCGTA-3′, reverse: 5′-AGTCCTCGGGCCATGATT-3′), mt-dloop1 (forward: 5′-AATCTACCATCCTCCGTGAAACC-3′, reverse: 5′-TCAGTTTAGCTACCCCCAAGTTTAA-3′), mt-rnr2 (forward: 5′-CACTGCCTGCCCAGTGA-3′, reverse: 5′-ATACCGCGGCCGTTAAA-3′) and PGC-1α (forward: 5′-TCAAGCCAAACCAACAACTTTATCT-3′, reverse: 5′-GGTTCGCTCAATAGTCTTGTTCTCA-3′), and normalized to 18S ribosomal 5 (forward: 5′-GTAACCCGTTGAACCCCATT-3′, reverse: 5′-CCATCCAATCGGTAGTAGCG-3′); quantification was performed using the ΔΔCt method.

In vitro T cell cultures

For the bead-based continuous stimulation under hypoxia assay, LN and spleen CD8+ T cells were isolated from 6- to 8-week-old mice, mechanically disrupted and sorted on CD8+ CD44hi via Beckman Coulter MoFlo Astrios. Cells were then activated at 2 × 104 T cells per 96-well round-bottomed plates with an equivalent number of CD3/CD28 washed dynabeads, 25 U ml−1 of IL-2 and 10 ng ml−1 of IL-12 in 200 μl of complete RPMI + 10% serum (starting point = day 0). Cells were activated for 24 h, then the beads were magnetically removed and the cells divided into four conditions: no dynabeads in regular incubator (acute activation in normoxia), no dynabeads in 1.5% oxygen hypoxia chamber (acute activation in hypoxia; BioSpherix, ProOx Model C21), and with 200,000 dynabeads in the regular incubator (continuous activation in normoxia), and with 200,000 dynabeads in a 1.5% oxygen hypoxia chamber (continuous activation in hypoxia). Cells were cultured in 25 U ml−1 of IL-2 in 300 μl of complete RPMI + 10% serum in 96-well round-bottomed plates (starting culture conditions = day 1). After 48 h (day 3), cells were split in half (for example, 10 wells per group are now 20 wells per group), with fresh medium + IL-2 used to replace the old medium. Bead numbers were kept consistent per well. After 48 h (day 5), cells were split in half again, with fresh medium + IL-2 used to replace the old medium. Bead numbers were kept consistent per well. After 24 h (day 6), cells were assayed after the beads had been removed.

For the antimycin A–based assay, spleen and LN preparations from OT-I mice were stimulated with 250 ng ml−1 of SIINFEKL peptide and 50 U ml−1 of IL-2 for 24 h. Cells were washed, expanded tenfold into fresh medium with IL-2 and cultured either in medium + 25 U ml−1 of IL-2 alone, or with 0.04 μM antimycin A, 0.4 μM rotenone, 0.4 μM rotenone + 0.04 μM antimycin A or 0.04 μM antimycin A + 10 mM NAC. Cells were cultured for 7 d in these conditions, then assayed.

For the in vitro T cell + tumor cell-based assay, spleen and LN preparations from OT-I mice were stimulated with 250 ng ml−1 of SIINFEKL peptide and 50 U ml−1 of IL-2 for 24 h. Cells were washed, then plated alone, 1:1 with 1 × 106 B16 or 1:1 with 1 × 106 B16OVA (in 10-cm plate, with 20 ml of medium), and placed in normoxia (atmospheric O2) or hypoxia (1.5% O2; BioSpherix, ProOx Model C21), all with 25 U ml−1 of IL-2. After 48 h, T cells were counted and replated at 1:1 with fresh tumor cells: 1 × 106 B16 or 1 × 106 B16OVA (in 10-cm plate, with 20 ml of medium), and were continued in normoxia or hypoxia with 25 U ml−1 of IL-2. After another 48 h, the cells were assayed.

For sodium orthovanadate (phosphatase inhibitor)-based assay, spleen and LN preparations from OT-I mice were stimulated with 250 ng ml−1 of SIINFEKL peptide and 50 U ml−1 of IL-2 for 24 h. Cells were washed, expanded tenfold into fresh medium with IL-2 and cultured in either medium + 25 U ml−1 of IL-2 alone or the presence of 50 μM sodium orthovanadate for 5 d, and then assayed.

For ethidium bromide-based assays, spleen and LN preparations from OT-I mice were stimulated with 250 ng ml−1 of SIINFEKL peptide and 50 U ml−1 of IL-2 for 24 h. Cells were washed, expanded tenfold into fresh medium with IL-2 and cultured in medium, 25 U ml−1 of IL-2, 50 mg ml−1 of uridine, plus or minus 50 ng ml−1 of ethidium bromide for 12 d, and then assayed.

Luminescence assay

The PGC-1α plasmid was originally generated by B. Spiegelman and obtained from Addgene (plasmid no. 8887). The Prdm1 (Blimp-1) plasmid was a gift from A. Poholek (University of Pittsburgh)56. The 293T cells were co-transfected with the plasmids, then, next day, the Luciferase Assay System (Thermo Fisher Scientific, catalog no. E1500) was used to assay luminescence, according to the manufacturer’s instructions.

Gene expression profiling by RNA-seq

In vitro cultured CD8+ T cells or TILs were dissociated from tissue as described above, then sorted based on CD8, PD-1 and Tim-3 expression. The cDNA was prepared from ~1,000 cells using the SMARTer Ultra Low Input RNA Kit for Sequencing, v.3 user manual (Clontech Laboratories). Sequencing libraries were prepared using the Nextera XT DNA Library Preparation kit (Illumina), normalized at 2 nM using Tris-HCl (10 mM, pH 8.5) with 0.1% Tween-20, diluted and denatured to a final concentration of 1.8 nM using the Illumina Denaturing and Diluting libraries for the NextSeq 500 protocol Revision D (Illumina). Cluster generation and 75-bp paired-end, dual-indexed sequencing were performed on the Illumina NextSeq 500 system. RNA-seq was analyzed using standard methods, including alignment to the genome using HISAT2, gene expression values (transcripts per million) calculated using Subread and identification of DEGs using DEseq2 with a cutoff of twofold and P < 0.05.

Microscopy

NFAT1 localization was analyzed via confocal microscopy. Exhausted T cells were generated in vitro according to the above protocol, then T cells were adhered to glass slides via poly(l-lysine) coating. Cells were then fixed with 4% PFA for 30 min, followed by Foxp3 fix/permeabilization (Fisher) for 30 min. Cells were then stained overnight with anti-NFAT1, followed by phalloidin and DAPI next day before being mounted using ProLong Diamond Antifade Mountant (Fisher). Cells were imaged with an Olympus IX81 spinning disk confocal and analyzed using ImageJ software. Analysis was performed by determining NFAT1 mean fluorescent intensity in the nuclear mask versus whole-cell mask on a per-cell basis.

In vivo tamoxifen administration

Prdm1f/f × E8I-Cre-ERT2 GFP × Rosa26-LSL-TdTomato mice (and Cre-negative littermate controls) received 2.5 × 105 B16 cells intradermally on day 0. After 9 d (when tumors were around 5 mm in size), mice received 1 mg of tamoxifen suspended in corn oil IP for 5 consecutive days. On day 6, mice were sacrificed, tissue was processed as described above and TILs were assayed.

In vivo TIL analysis and survival with axitinib, anti-PD-1 and anti-CTLA-4

Mice were injected intradermally with 2.5 × 105 B16 melanoma cells. When tumors were palpable (typically day 5), mice began axitinib and/or anti-PD-1 plus anti-CTLA-4. Mice were treated with 10 mg kg−1 of axitinib (Cayman Chemical) or vehicle control (10% dimethysulfoxide, 10% Tween-80 and 80% water containing 30% captisol), and/or 200 μg of anti-PD-1 plus 200 μg of anti-CTLA-4 intraperitoneally three times per week. In survival curves, mice were removed from the study when the tumor burden reached 15 mm in any direction. In some experiments, mice were injected intravenously with pimonidazole (80 mg kg−1, Hypoxyprobe) in PBS 1.5 h before sacrifice. Pimonidazole was visualized using anti-pimonidazole antibodies (Hypoxyprobe) after 10 min of 4% PFA fixation, followed by Foxp3 Fix/perm (eBioscience) permeabilization.

Statistics

The P values were calculated in GraphPad Prism using one-way analysis of variation (ANOVA) with Dunnett’s multiple comparison test, unpaired Student’s t-test or paired Student’s t-test, as indicated in the figure legends. Values of P < 0.05 were considered significant. Values of P < 0.05 were ranked as: P < 0.05, P < 0.01, P < 0.001 and P < 0.0001.

Reporting Summary

Further information on research design is available in the Nature Research Reporting Summary linked to this article.