Melanoma Evolves Complete Immunotherapy Resistance through the Acquisition of a Hypermetabolic Phenotype

T-cell checkpoint blockade antibodies are now approved for the treatment of a majority of adult cancers either alone or in various combinations (1). Despite this remarkable progress, most patients with cancer still show intrinsic or naturally acquired resistance to immune checkpoint blockade leading to treatment failure. In addition, certain immunologically “cold” tumors, such as pancreatic ductal adenocarcinoma, have no appreciable response to these therapies (2–4). Before we can identify biomarkers of checkpoint antibody response, or rationally plan to circumvent immunotherapy resistance, we must understand the underlying processes responsible for lack of and/or loss of response.

The mechanisms employed by tumors to escape host immunity have been extensively studied even prior to clinical adoption of checkpoint blockade (5–9). Most of the initial research addressing checkpoint blockade resistance mechanisms focused on the upregulation of alternative immune checkpoints such as TIM3, TIGIT, and VISTA (10–12). Mutational load, or lack thereof (13–15), loss of interferon response (16, 17), and copy-number loss of components of the antigen presentation machinery (18, 19) by tumor cells are also described as mechanisms of resistance to αPD-1 and αCTLA-4 monotherapies. However, these pathways fail to account for the majority of nonresponders. In addition, little is known of the transcriptomic states of tumor cells which favor immunotherapeutic sensitivity versus resistance. To address this critical gap in knowledge, we established an immunotherapy-resistant mouse model of melanoma. We employed an approach rooted in both “cancer immunoediting” theory (9), postulating that immune pressure fosters suppressive adaptations by tumors, and on the in vivo serial passage approach pioneered by Fidler and colleagues (20–22). Mice were treated for B16 melanoma using a triple checkpoint blockade and vaccination approach which cures ≥75% (23), and progressive tumors were then isolated and inoculated into new mice that received the same therapy. These serial immune selections were performed until a completely checkpoint blockade–resistant melanoma emerged. Using gene expression analysis and immunogenomics, flow cytometry, confocal imaging and metabolic assays, we showed that resistant tumors acquired a “hypermetabolic” phenotype characterized by coordinated upregulation of the glycolytic, oxidoreductase, and mitochondrial oxidative phosphorylation (OxPhos) pathways, thus creating a hostile metabolic microenvironment in which cytotoxic CD8⁺ T cells are energetically starved and rendered dysfunctional. Heterologous overexpression of the key genes driving enhanced metabolic fitness in the resistant melanoma back into the parental B16 line conferred substantial resistance to triple checkpoint therapy, thus validating their role in immunotherapy resistance. In a cohort of patients with melanoma who failed both CTLA-4 and PD-1 blockade, we identified similar metabolic enhancement in progressing versus responding patients, suggesting clinical relevance of these findings. Overall, our data demonstrate that melanoma is capable of evolving resistance to simultaneous blockade of CTLA-4, PD-1, and PD-L1 through acquisition of a hypermetabolic state characterized by enhanced glycolysis, oxidoreductase activity, and mitochondrial OxPhos.

Male C57BL/6J and B6.129S7-Rag1tm1Mom/J mice were purchased from The Jackson Laboratory. Mice were housed in our pathogen-free facility which is fully accredited by the Association for Assessment and Accreditation of Laboratory Animals Care. All experiments were performed according to protocols approved by the Institutional Animal Care and Use Committee.

Anti–CTLA-4 (9H10), anti–PD-1 (RMP1-14), and anti–PD-L1 (10F.9G2) were purchased from BioXCell and administered intraperitoneally

A cohort of 9 melanoma patients was included in the analysis. Surgical samples were from patients with metastatic melanoma treated with ipilimumab (αCTLA-4) and/or pembrolizumab or nivolumab (αPD-1) at the UT MD Anderson Cancer Center (Houston, TX) between April 2014 and September 2015 on IRB protocol 2012-0846 prior to therapy or at time of progression (Supplementary Table S1). Clinical responses were assessed on the basis of RECIST 1.1 (24). This study was designed and monitored in accordance with sponsor procedures in compliance with the ethical principles of Good Clinical Practice, International Conference on Harmonization guidelines, the Declaration of Helsinki, and applicable local regulatory requirements. All patients provided written, informed consent. The protocol, any amendments, and informed consent forms were reviewed and approved by the institutional review boards/independent ethics committees.

The B16/BL6 cell line was originally obtained from Fidler (MD Anderson Cancer Center, Houston, TX) in 2012. The B16-sFlt3L-Ig (FVAX) and B16-tdTomato cell lines have been described previously (25). The cells were maintained in RPMI media with 10% FBS. Panc02 cells, originally obtained from the lab of Dr. Elizabeth Jaffee (Johns Hopkins University, Baltimore, MD) in 2010, were maintained in DMEM supplemented with 10% FBS. Cells were used within 2 weeks of thaw from their master tumor banks. These cells were tested Mycoplasma negative but not authenticated in the past year.

To harvest mouse tumor single-cell suspensions, tumors were removed postsacrifice and were treated with 0.25 mg mL⁻¹ collagenase H (Sigma-Aldrich and 25 U mL⁻¹ DNase (Roche Diagnostics) for 20 minutes at 37°C; the dissociated cells were then passed through a sterile 70 μm filter (Thermo Fisher Scientific, catalog no. 352350) The resulting dissociated cells were collected by centrifugation and washed twice in PBS. The cells were then cultured and/or used for flow cytometry analysis and/or flow sorting.

We implanted 15 mice with 25,000 B16/BL6-td cells subcutaneously in 100 μL of PBS and treated them with a combination of three T-cell checkpoint blockade antibodies. On days 3, 6, and 9, postimplantation, mice were injected subcutaneously with 1 × 10⁶ irradiated (150 Gy) FVAX cells resuspended in 100 μL of PBS on the contralateral flank and an intraperitoneal injection of combination αCTLA-4 (100 μg), αPD-1 (250 μg), and αPD-L1 (100 μg) diluted to 100 μL or 200 μL of PBS. Mice developing tumors regardless of treatment were euthanized when tumors reached 200 to 500 mm³ and their tumors harvested. Tumors from nonresponder mice were pooled and a cell line (3I-F1) was generated and maintained in RPMI media with 10% FBS. This 3I-F1 cell line was used to inoculate a new set of 15 mice followed by the same immunotherapy regimen. For this second cycle and all subsequent cycles, only 10,000 cells were implanted to increase the rigor of the screen. We repeated these serial passages until ≥90% of the animals became resistant to therapy. B16 melanoma cell lines were thus named 3I-F1, 3I-F2, 3I-F3, and 3I-F4 (completely resistant). For the untreated control group, we implanted 5 mice with parental tumor cells and with tumor cells from each cycle of selection.

For Panc02 studies, C57BL6 mice were subcutaneously implanted with 5 × 10⁵ control vector, phosphoglycerate mutase 2 (PGAM2)– or alcohol dehydrogenase 7 (ADH7)–overexpressing tumor cells and treated with either αCTLA-4 or αPD-1 as mentioned above. Metformin (50 mg/kg; every other day) and 2-DG (500 mg/kg; daily) were given intraperitoneally beginning one day following tumor challenge. For metformin water cohorts, drinking water containing 1 g/L metformin was provided posttumor implantation. LDH-A inhibitor GSK2837808A (4 mg/kg) was prepared in 100 μL polyethylene glycol base and given through oral gavage everyday postimplantation. Tumors were measured using calipers every other day and tumor volume was calculated as length × width × height. Death was scored as tumor volume of 1,000 mm³.

Tumors were harvested and sorted using a BD FACSAria cell sorter and BD FACSDiva Software on tdtomato fluorescence in tumor versus tumor microenvironment (TME). Total RNA was extracted with the RNeasy Mini Kit (Qiagen, catalog no. 74106), MD). For Adh7 and Pgam2 expression, cells were lysed and RNA was extracted using the RNeasy Mini Kit. cDNA was generated using the Invitrogen SuperScript IV reverse transcriptase Kit (Thermo Fisher Scientific, catalog no. 18090050). TaqMan real-time PCR was performed on a Via 7 RT-PCR System (Applied Biosystems) as described previously (26). For patient biopsies, the presence of tumor was confirmed by a pathologist, and total RNA was extracted from the tumor tissue using the RNeasy Mini Kit (Qiagen).

Tumor cells and nontumor cells of the TME were sorted by flow cytometry and RNA was isolated from both as above. Microarray analysis was done on both tumor cells and TME from RNA samples (n = 2) from parental tumors and from 3I-F4 tumors (n = 4). Each RNA sample was isolated from tumors pooled from three mice. Microarray analysis was also done on RNA from patients' tumor biopsies. MouseRef-8 and HumanHT-2 bead chip arrays (Illumina) were used. For resistant B16 melanoma analysis, MouseRef-8 microarray analysis was performed on RNA from a 15 cm plate from each generational line at the earliest passage freeze available. Microarray data are deposited into Gene Expression Omnibus (GEO) with the record number GSE122222.

RNA was reverse transcribed using the SuperScript IV First-strand Synthesis Kit (Thermo Fisher Scientific, catalog no. 18091050). RNA concentration was quantified using a NanoDrop 8000 Spectrophotometer (Thermo Fisher Scientific, catalog no. ND-8000-GL). A total of 1 μg of RNA template was used as input for cDNA synthesis. Samples were run in triplicates on a ViiA 7 Real-Time PCR System and gene expression was normalized to Hprt (Thermo Fisher Scientific, Assay ID: Mm03024075_m1). Fold change in gene expression of Adh7 (Thermo Fisher Scientific, Assay ID: Mm03121387_m1) or Pgam2 (Thermo Fisher Scientific, Assay ID: Mm01187768_m1) was calculated using the 2^−ΔΔC_T method.

Microarray data were normalized per manufacturer's instructions and processed in R. Low intensity probes that were not significantly expressed above background level (detection P value≥0.05 in at least one of the samples) were excluded. Differential expression between resistant and parental for tumor and TME were determined by fold-change in absolute value ≥1.1 and P value from the moderated t-statistic from LIMMA package ≤0.05. To support visual data exploration, we employed R to generate volcano plots, as well as heatmaps using the heatmap.2 function of gplots library.

For The Cancer Genome Atlas (TCGA) analysis, the data were obtained from TCGA portal and lymphocyte score (LS score) table was obtained from publisher's website (Supplemental Table S1D: Patient Centric Table). The LS scores were used to categorize patients into two groups: LS-high and LS-low patients. The LS-high group included the patients whose LS score were between 3 and 6, and LS-low group contained patients of LS scores between 0 and 3. The Student t test was applied to compare gene expression between LS-high and LS-low groups.

Gene set enrichment analysis (GSEA) and Ingenuity Pathway Analysis (IPA) were applied to the datasets in an unbiased fashion to compare resistant tumors with parental tumors and responding patients with nonresponders.

Parental and resistant cells were plated at a density of 5,000 cells per well on 96-well plates in 10 replicates. Photomicrographs were taken every hour using an Incucyte cell imager (Essen Biosciences) and confluency was determined using Incucyte software. Cells were grown in the HypOxystation H35, (Hypoxygen) under 1% oxygen and 85% humidity. Parental and resistant cells were plated as above and cell viability was measured by Cell Titer 96 Aqueous One Solution Cell Proliferation Assay (MTS; Promega, catalog no. G3582).

Resistant and parental cell lines were seeded at a density of 25,000 cells per well 24 hours prior to the assay. Oxygen consumption rate (OCR; Seahorse XF Cell Mito Stress Test Kit, Agilent Technologies, Part #103015-100) and extracellular acidification rate (ECAR; Seahorse XF Glycolysis Stress Test Kit, Agilent Technologies, Part # 103020-100) were measured as per the manufacturer's protocols on an XF96 Analyzer (Agilent Technologies).

Mice were injected intravenously with Pimonidazole (Hypoxyprobe, catalog no. HP-500 mg) 30′ prior to euthanasia so that hypoxia could be imaged in tumor sections by immunofluorescence staining with a pimonidazole adduct FITC antibody (Hypoxyprobe). Mouse tissues were collected and embedded in Tissue-Tek OCT Compound (Sakura). Embedded tissues were frozen in liquid nitrogen and sectioned at the MD Anderson Histology Core. Sectioned tissue was fixed with acetone for 10′, permeabilized with the FoxP3 Staining Kit (eBioscience, catalog no. 00-5521-00) for 10′, and blocked with Superblock (Thermo Fisher Scientific, catalog no. 37515) for 15′. The samples were stained with antibodies in 2% BSA, 0.2% Triton-X100 in PBS for 30′ and, after being washed in PBS, mounted with Prolong Gold anti-fade reagent (Invitrogen, catalog no. P10144). Fluorescence microscopy was performed using a TCS SP8 confocal microscope equipped with lasers for 405, 458, 488, 514, 568, and 642 nm wavelengths (Leica Microsystems, Inc.). The ImageJ software was used to analyze the images.

Cells were trypsinized and washed twice with PBS and frozen in liquid nitrogen. Tumors from mice with and without immunotherapy treatment were collected on days 12 to 16 postimplantation and frozen in liquid nitrogen. Cells were counted, and tumor tissues were weighed before extraction of metabolites. Cells and tumor tissues were homogenized and added with 2:1 methanol and ceramic beads. The tissues/cells were then vortexed for 40 to 60 seconds followed by freezing in liquid nitrogen and thawing on ice. Water soluble proteins and other biopolymers were precipitated in methanol solvent leaving the small molecular weight metabolites in the solution which were then extracted using ultracentrifugation. Remaining residual solvent was removed by overnight lyophilization.

The lyophilized sample was dissolved in 800 μL of 2H₂O and centrifuged at 10,000 rpm. The 600 μL of sample was added with 40 μL of 8 mmol/L 4,4-dimethyl-4-silapentane-1-sulfonic acid (DSS) before acquisition on nuclear magnetic resonance (NMR). NMR data were collected on an Avance Bruker spectrometer operating at 500 MHz proton (¹H) resonance frequency, equipped with cryogenically cooled triple resonance (¹H, ¹³C, ¹⁵N) TXI probe. All one-dimensional (1D) ¹H NMR spectra were acquired with suppressed solvent signal achieved by presaturation during longitudinal relaxation time. The interscan delay of 6 seconds is used to rule out the longitudinal relaxation-related signal attenuation. The 900 radiofrequency pulse of 12 μs, spectral width of 8,000 Hz, and 256 transients were used to acquire the 1D ¹H NMR. All spectra were processed in topspin 3.1 and metabolites were assigned with the help of Chenomx and Human Metabolomics Database. The intensities of metabolites were taken with respect to NMR reference compound of 0.5 mmol/L 4,4-dimethyl-4-silapentane-1-sulfonic acid (DSS) appearing at 0 ppm. Intensities (AUC) of the metabolites were normalized to the cell numbers and tumor mass. The normalized intensities were used to calculate the Z score expressing relative expression of metabolite in resistant tumors/cell lines compared were parental tumors/cell line.

The mixture of 20 μL 1–¹³C, 10 μL of 15 mmol/L trityl radical OX63, and 0.4 μL Gd2⁺ was hyperpolarized for an hour with microwave irradiation at 94 GHz at low temperature 1.5 K in Oxford Hypersense instrument. The hyperpolarized pyruvate was dissolved at high temperature in 4 mL of TRIS/EDTA buffer at physiologic pH 7.8 to a final concentration of 80 mmol/L of pyruvate. A total of 200 μL of the solution was injected into the mice via tail vein injection in a horizontal bore 7 T Bruker MR Scanner (27).

The anatomic proton image and ¹³C magnetic resonance spectroscopy (MRS) were acquired using a surface transceiver ¹³C-¹H coil (Doty Scientifics). Anatomic images of coronal, axial, and sagittal were acquired with T2 weighted Rapid Imaging with Refocused Echo sequence to determine the size and location of tumors. The ¹³C enriched urea phantom was used as spectroscopic reference and to locate the tumor. The single pulse Fast Low Angle Shot was used to acquire 1D ¹³C MRS with repetition time of 2 seconds, flip angle 200, image size 2048 × 90, and single slice of thickness 5–10 mm and acquired over a period of 180 seconds (27).

Following gradient separation, samples were fixed using the Foxp3 Buffer Set (eBioscience) and then incubated with anti-mouse CD15/CD32 Ab (clone 2.4G2, BioXcell, catalog no. BE0307) for 15 minutes before being stained with up to 18 antibodies from Abcam, Biolegend, BD Biosciences, eBioscience, or Life Technologies. Antibodies were used against the following mouse proteins in different combinations: CD45 (30-F11), CD45.2 (104), CD3 (17A2), CD8a (53-6.7), CD4 (GK1.5), CD11b (M1/70), F4/80 (BM8), IDO (mIDO-48), GLUT1 (EPR3915), PD-1 (RMP1-30), Granzyme B (NGZB), 4-1BB (17B5), IL-2 (JES6-5H4), TNFα (MP6-XT22), IFNγ (XMG1.2), FoxP3 (FJK-16s), Gr-1 (RB6-8C5), Arginase 1 (A1exF5). Flow cytometry data were collected on a 5-laser BD LSR II cytometer and analyzed using FlowJo (Treestar; refs. 26, 28).

For metabolic characterization of lymphocytes, fluorescently labeled glucose (2-NBDG; Cayman Chemical Company, catalog no. 11046) was injected intravenously 30′ prior to sacrificing mice for tumor harvest. For characterization of T-cell effector function, mice were injected subcutaneously with a single-cell suspension of 250 K parental or resistant tumor cells in 100 μL of PBS. Mice were treated on days 3, 6, 9 postimplantation with FVAX and triple immune checkpoint blockade as above. On day 11, mice were euthanized, and viable immune cells from tumor digests were enriched through density gradient separation over Histopaque 1119 (Sigma-Aldrich, catalog no. 111-91-100ML). CD8⁺ T cells were isolated from immune cells via negative selection through magnetic beads using a MACS CD8⁺ T cells Isolation Kit (Miltenyi Biotec, catalog no. 130-104-075). CD8⁺ T cells were activated for 6 hours using eBioscience Cell Stimulation Cocktail (plus protein transport inhibitors; catalog no. 00-4975-93) and analyzed by flow cytometry.

Murine PGAM2 and ADH7 cDNAs were cloned into the pMG-rtNGFr retroviral vector. This vector resembles pGC-IRES except that for a truncated form of rat p75 nerve growth factor receptor (rtNGFr) is used for selection (29). Recombinant virus production and infection were performed as described previously (30) to generate B16-BL6-td or Panc02 cells overexpressing either ADH7 or PGAM2.

PGAM2 (sgPGAM2), ADH7 (sgADH7), or control (sgScramble) KO clones were generated using the all-in-one CRISPR/Cas9 system (Genecopoeia, pCRISPR-CG04). Vectors encoding a scramble guide (sg) RNA sequence or guides targeting exon 1 or 2 of each gene were transfected into 3I-F4 cells. GFP⁺ cells were sorted 24 hours later, plated, and loss of either ADH7 or PGAM2 expression was validated by qRT-PCR. To generate PGAM2 knockout, ADH7 knockdown cells (sgPGAM2/shADH7), a lentiviral short hairpin (shRNA, LVRH1GH) construct was introduced into sgPGAM2 cells. Control sgScramble/shScramble and sgPGAM2/shScramble were generated by transduction of a plasmid containing a scrambled shRNA. GFP⁺ cells were sorted, and knockdown of ADH7 was validated by qRT-PCR as described above.

All statistics were calculated using Graphpad Prism. Statistical significance was determined using a two-tailed Student t test applying Welch correction for unequal variance. Graphs show mean ± SD unless otherwise indicated. P values less than 0.05 were considered significant.

We generated the immunotherapy-resistant melanoma cell line (3I-F4) by in vivo passaging a B16/BL6-tdTomato melanoma in the presence of the combination of a B16-Fms–like tyrosine kinase 3 ligand vaccine (FVAX) and antibody blockade of CTLA-4, PD-1 and PD-L1, a therapy with initial survival benefit in ≥75% of animals (Fig. 1A and B; ref. 23).

To ensure that the lack of immunotherapy response in the resistant B16/3I-F4 clone did not result from accelerated proliferation, we compared in vitro and in vivo proliferation of B16/3I-F4 and B16/BL6 (Parental). Using the IncuCyte confluency assay, we found no significant difference in proliferation between parental and resistant B16 (Fig. 1C). We also compared in vivo tumor growth and survival of mice with parental and resistant tumors in normal C57BL6 and immune-deficient B6.Rag^−/− mice. Untreated parental and resistant tumors showed no significant difference in tumor growth rate or host survival in C57BL6 (Fig. 1D; Supplementary Fig. S1A) or B6.Rag^−/− mice (Fig. 1E; Supplementary Fig. S1B). In the presence of triple checkpoint blockade; however, C57BL6 mice with parental tumors showed reduced tumor growth and significant survival benefit (Fig. 1D; Supplementary Fig. S1A). In contrast, in B6.Rag^−/− mice, both parental and resistant melanomas grew at the same rate, even in the presence of immunotherapy, demonstrating that resistance depends on differential insensitivity to adaptive immunity and does not result from amplified proliferation.

We next probed the underlying acquired genetic changes within immunotherapy-resistant tumors responsible for their resistance. We harvested resistant 3I-F4 tumors, sorted tumor cells from nontumor cells based on tdTomato expression as shown in the histogram (Fig. 2A), and performed gene expression profiling on tumor cells. Expression of numerous genes was significantly altered during acquisition of triple checkpoint resistance; however, top candidate genes generally clustered into metabolic pathways, in particular, glycolysis, OxPhos, oxidative stress, and hypoxia (Fig. 2B–D).

GSEA and an IPA were performed and showed that immunotherapy-resistant tumors augmented biological pathways involving mitochondrial OxPhos, oxidoreductase activity, hypoxia response, and glycolysis (Fig. 2C). Resistant tumors increased OxPhos as suggested by the “MOOTHA_VOXPHOS” gene set and these increases in OxPhos appeared to be associated with deepening hypoxia as suggested by “MANALO_HYPOXIA_DN” (Fig. 2D). The gene set “NFE2L2.V2,” representing genes that are critical for oxidative stress responses, was also positively enriched in 3I-F4. This suggested that resistant tumor cells altered their oxidative damage response to become better adapted to the cellular stress caused by hypermetabolism and its associated worsening of hypoxia (Fig. 2D). Upregulation of the key metabolic regulatory genes, Adh7 and Pgam2, associated with acquired checkpoint blockade resistance from the microarray was validated by Taqman qRT-PCR analysis comparing resistant with parental lines (Fig. 2E). Gene expression analysis of microarray data from cell lines generated at each passage (i.e., 3I-F1–3I-F4) shows progressive genetic evolution of each line as they move toward more complete immunotherapy resistance (Supplementary Fig. S2). Taken together, these data suggested that resistant tumors deplete oxygen and other nutrients in the TME and nucleate a state of hypoxia in which they can flourish in their metabolically adapted state, while lymphocytes face a harsh milieu in which they are metabolically unfit to thrive.

To validate the metabolic adaptations of resistant tumors, we assessed their glycolytic metabolism by measuring their ECAR, and their rate of OxPhos by measuring their OCR using a Seahorse XF analyzer (Agilent). The immunotherapy-resistant 3I-F4 line demonstrated higher basal ECAR and OCR relative to parental B16 with elevated maximum glycolytic capacity and mitochondrial respiration (Fig. 3A and B). This enhancement of both glycolysis and OxPhos is a departure from the expected Warburg effect, in which tumor cells rely primarily on glycolysis for ATP production in the oxygen-depleted TME (31). To further validate the hypermetabolic phenotype of immunotherapy-resistant tumors, we analyzed their cellular metabolites using NMR spectroscopy. The resistant cells showed relative increases in lactate and other tricarboxylic acid (TCA) cycle metabolites compared with the parental line (Supplementary Fig. S3A). We also compared metabolites extracted from whole tumor lysates of resistant versus parental tumors with and without treatment. Consistent with the cell line data, ex vivo resistant tumors also showed increased relative expression of lactate and other TCA cycle metabolites with and without therapy (Fig. 3C). The observed increase in these metabolites was more profound in the presence of immunotherapy treatment, suggesting that treatment itself directly or indirectly triggers these metabolic changes in resistant tumors.

On the basis of our in vitro and ex vivo metabolic analyses, we hypothesized that the increase in lactate production in resistant tumors could serve as a marker to separate immunotherapy-sensitive and immunetherapy-resistant tumors by visualizing conversion of hyperpolarized pyruvate into lactate utilizing noninvasive MRI. Using this approach, we showed that the rate of pyruvate to lactate conversion was significantly higher in immunotherapy-resistant tumors (Fig. 3D and E). This demonstrated the potential to segment immunotherapy-sensitive B16 melanoma away from checkpoint-resistant 3I-F4 tumors in live, untreated animals. Together, these data suggest that checkpoint blockade immunotherapy–resistant tumors acquired a hypermetabolic state where they upregulated both glycolysis and OxPhos to evade the host immune response.

We further investigated the role of hypoxia in mediating resistance to checkpoint blockade immunotherapy. Using confocal microscopy and the hypoxia-specific dye Pimonidazole, we observed how resistant and parental tumors (shown in red) interact with hypoxic zones in the TME. No difference was observed in the extent of hypoxia in untreated resistant and parental tumors (Supplementary Fig. S3B); however, following treatment, hypoxia appeared more prevalent in resistant compared with parental tumors (Supplementary Fig. S3C). Also, tdTomato⁺ cancer cells in resistant tumors grew in equal or greater density within hypoxic zones compared with parental. Quantitative analysis across multiple parental and resistant tumors confirmed a significant increase in the number of discrete zones of hypoxia within treated, resistant tumors relative to both untreated resistant and treated parental melanomas (Fig. 3F). An in vitro survival assay comparing resistant and parental tumors in a hypoxic chamber (pO2 = 1%) showed increased accumulation of the resistant 3I-F4 cell line, further illustrating that these cells can thrive under adverse metabolic conditions (Fig. 3G). Thus, immunotherapy-resistant 3I-F4 cells have acquired a hypermetabolic phenotype and help to propagate a metabolically hostile TME in which they have adapted to flourish.

We next assayed the impact of the hypermetabolic adaptations of the tumor cells themselves on the gene signature of the surrounding TME. In this case, we separated the TME (tdTomato⁻) away from the tumor (tdTomato⁺) by FACS and analyzed gene expression of the tdTomato⁻ fraction (Fig. 4A). In contrast to our prior observation for the tumor itself (Fig. 2D), the GSEA analysis of the resistant TME showed gene signatures indicative of a failure to adapt to hypoxia (Fig. 4B and C). Perhaps because of this inability to adapt to increasing hypoxic stress, TME gene expression data suggested diminished antitumor immune function as indicated by negative enrichment of gene sets encompassing T-cell effector function, myeloid (DC and microphages) cell activation, and DC maturation (Fig. 4D and E).

Next, we investigated the effects of metabolic adaptation by resistant tumor cells on the composition and phenotype of immune cells in the TME. We established both parental and resistant tumors with or without treatment and then isolated them on day 14 and performed 18-parameter flow cytometry analysis to characterize immune infiltration (Supplementary Fig. S4). While baseline CD8⁺ T-cell infiltration of all untreated B16 melanoma tumors was poor, vaccination and triple checkpoint blockade elicited profound expansion of the tumor-infiltrating CD8⁺ compartment in the parental tumors (Fig. 5A). In contrast, CD8⁺ T-cell density failed to significantly increase in response to the same therapy in resistant 3I-F4 melanomas. The treatment expanded CD4⁺ effector cells and improved the CD8⁺ T cell to Treg ratios but failed to do so in resistant tumors (Supplementary Fig. S5A and S5B). The baseline proliferation of CD8⁺ T cells in the TME of resistant tumors was significantly depressed; however, treatment elicited similar Ki67 expression in both melanomas suggesting that reduced CD8⁺ T-cell density in the resistant setting likely resulted from impaired T-cell persistence and/or emigration (Fig. 5B; Supplementary Fig. S4). Granzyme B, a marker for cytolytic potential, and Glut1, a glucose transporter and maker of glycolytic potential (32), showed attenuated therapeutic induction in the resistant melanomas relative to parental (Fig. 5C and D; Supplementary Fig. S4).

In the CD4⁺ compartment, effectors had lower baseline Glut1 in 3I-F4 but could still respond to immunotherapy, whereas Treg cells lost Glut1 responsiveness to treatment (Supplementary Fig. S5C and S5D). CD8⁺ T cells in resistant melanoma showed significantly increased expression of the exhaustion marker PD-1 suggesting reduced fitness in the presence of hypermetabolic melanoma, although PD-1 remained unchanged on CD4 T cells perhaps reflecting differing metabolic demands on each subset (Fig. 5E; Supplementary Fig. S5E and S5F).

Effector capacity of cytotoxic T cells is tightly linked to their metabolic fitness, particularly to their glycolytic capacity. To test the effect of metabolic adaptation of resistant tumors on CD8⁺ T-cell function, we measured glucose uptake using the fluorescently labeled glucose analog 2-NBDG (Thermo Fisher Scientific) and mitochondrial membrane potential using MitoTracker Deep Red FM (Thermo Fisher Scientific) in tumor-infiltrating T cells. CD8⁺ T cells demonstrated reduced glucose uptake and showed high Mito FM staining in resistant compared with parental tumors (Fig. 5F and G). This diminished glycolytic metabolism coupled with a compensatory increase in OxPhos is emblematic of CD8⁺ T cells which lack significant effector function and are metabolically focused on survival. Of note, splenic CD8⁺ T-cell metabolism was unchanged between parental and resistant melanomas showing that diminished glycolysis in the resistant-tumor resident CD8⁺ T was a result of localized conditioning by 3I-F4 tumors. Consistent with our earlier observations of resistant tumor hypermetabolism under therapeutic pressure, we observed both enhanced glycolytic potential and OxPhos in treated 3I-F4 melanomas relative to parental (Fig. 5G).

Next, we examined a population enriched for tumor antigen–specific CD8⁺ T cells which express the costimulatory molecule 4-1BB (33). CD8⁺ T cells extracted from parental tumors that received combination immunotherapy produced significantly increased levels of the cytokines TNFα and IL2, and expressed more CD107a, a marker of T-cell degranulation. In contrast, 4-1BB⁺ CD8⁺ T cells extracted from treated resistant tumors failed to receive benefit from immunotherapy treatment across CD107a, TNFα, IL2, and IFNγ (Fig. 5H; Supplementary Fig. S4).

We also investigated the effects of tumor hypermetabolism on the frequency and phenotype of tumor-infiltrating myeloid-derived suppressor cells (MDSC). While the overall frequency of MDSC did not increase in resistant versus parental tumors (Supplementary Fig. S6A), the expression of the T-cell suppressive enzymes IDO and Arginase both significantly increased in 3I-F4 tumor MDSC in response to treatment (Supplementary Fig. S6B and S6C). Together, these data suggested that metabolic adaptation of immunotherapy-resistant tumors created a hostile TME where antitumor CD8⁺ T cells failed to accumulate or develop enhanced effector capacity in response to checkpoint blockade, whereas the immunosuppressive potential of resident of MDSC increased.

Within the metabolic pathways induced in the resistant 3I-F4 melanoma cells, gene expression analysis revealed certain key nodes which were among the most highly and significantly induced (Fig. 2B–E). Among these, we investigated PGAM2, a key enzyme in both glycolysis and the synthesis of nucleotide and amino acid precursors, and ADH7, which can decrease oxidative stress (oxidoreductase pathway) by reducing NAD to NADH and can also contribute to retinoic acid synthesis, as a potential contributor to 3I-F4 resistance (34, 35). To test the capacity of PGAM2 and ADH7 to confer checkpoint blockade resistance, we retrovirally overexpressed each, or a control vector, in the parental B16 melanoma cell line. Next, we implanted these vector control and PGAM2- or ADH7-overexpressing melanomas and followed tumor growth and survival in the recipient mice with and without immunotherapy. In the absence of treatment, both PGAM2- and ADH7-overexpressing tumors did not show significant differences in tumor growth or survival relative to parental (Fig. 6A–D; Supplementary Fig. S4). When treated with triple T-cell checkpoint blockade; however, PGAM2- and ADH7-overexpressing tumors both conferred significant immunotherapy resistance relative to the vector-transduced parental tumor. Similarly, we found that both Panc02-PGAM2 and Panc02-ADH7 also showed significant resistance to PD-1 blockade and a trend toward resistance to CTLA-4 blockade relative to the vector transduced Panc02 line, further validating the capacity of these genes to confer checkpoint blockade resistance (Supplementary Fig. S7A–S7H; Supplementary Fig. S4).

Using CRISPR/Cas9 targeting (Genecopeia) to knockout PGAM2 and ADH7 in resistant 3I-F4 melanoma, in turn, produced tumors that showed increased checkpoint sensitivity (Fig. 6E–H; Supplementary Fig. S8A). We also attempted to knock out both PGAM2 and ADH7 using CRISPR/CAS9 but were unable to obtain viable double knockout cells. We were able to create an ADH7 knockdown using shRNA in the 3I-F4 PGAM2 knockout cell line; however, these cells showed no greater sensitivity to triple checkpoint blockade than PGAM2 knockout alone (Supplementary Fig. S8A and S8B).

Although specific inhibitors of these key genes are not available, we speculated that general inhibition of glycolysis or OxPhos might help restore sensitivity to immunotherapy in 3I-F4. We therefore treated resistant tumors and control parental tumors with 2-Deoxy-D-glucose (2DG), a structural analogue of glucose that inhibits glycolysis, and with a selective lactate dehydrogenase-A inhibitor (GSK2837808A) that both can inhibit glycolytic activity (Supplementary Fig. S9A; refs. 32, 36, 37). Unexpectedly, both drugs failed to provide any therapeutic advantage to resistant tumors when given in combination with immunotherapy. Metformin diminishes OxPhos and decreases hypoxia in a manner which can complement immunotherapy (29). In this case, however, neither oral (drinking water) nor injected Metformin was able to sensitize resistant tumors to immunotherapy (Supplementary Fig. S9B).

We sought to validate the role of metabolic adaptation in modulating the response to checkpoint blockade in patients with melanoma. We therefore performed gene expression analysis on mRNA samples from a cohort consisting of patients with metastatic melanoma who progressed on CTLA-4 blockade and then were treated with αPD-1 (13). Patients were biopsied prior to αPD-1 therapy and responses were assessed with serial CT scan after initiation of therapy. In this cohort, there were 4 patients who responded and 5 who did not respond to therapy (Fig. 7A). Detailed information on this cohort and the samples profiled is available in the GEO database as record GSE122222. IPA showed that, compared with responders, nonresponders enriched similar metabolic pathways to those identified in our resistant mouse models, including elevated OxPhos, glycolysis, and buffering of oxidative stress (Fig. 7B). Similarly, GSEA analysis showed enhanced hypoxia gene set induction, higher OxPhos, related alterations associated with enhanced mitochondrial respiration, and similar signatures of CD8⁺ T-cell exhaustion (Fig. 7C). Our confidence in this cohort was increased on the basis of finding alterations in PI3K/AKT and mTOR expression, as well as VEGF and IL8, as major nonmetabolic pathways associated with checkpoint resistance, as the significant role of these changes in mediating PD-1 resistance in patients with melanoma has been described previously (15, 38).

To determine the potential breadth of this observation, we divided the melanoma TCGA cohort into high T-cell versus low T-cell infiltration fractions based on lymphocyte score (LS score), which summarizes the lymphocyte distribution and density based on pathologic review (39). GSEA analysis shows a similar pathway of enhanced metabolic activity in melanoma tumors with poor immune infiltration (Fig. 7D). Aside from a modest signal in IPA, we did not find prominent induction of gene sets associated with enhanced glycolysis compared with those for OxPhos in either patient cohort. In contrast to our murine system, patient tumor RNA is isolated from whole tumor including the TME. In this setting, enhanced tumor glycolytic activity is likely balanced by reduced glycolysis in the surrounding TME resulting from depletion of critical nutrients. Melanoma tumor cell lines from patients who were resistant versus sensitive to adoptive transfer of ex vivo amplified tumor-infiltrating lymphocytes (TIL therapy) showed that enhanced glycolytic metabolism mediates resistance to TIL immunotherapy (40). In an in vitro screen, knockout of PGAM2 was one of two most potent hits in augmenting sensitivity to TIL lysis in patient melanoma cell lines, and in TIL-resistant patient-derived cell lines, PGAM4, an ortholog of mouse PGAM2, was identified as a significant contributor to resistance (40). Overall, these findings suggest that the murine model we generated to study checkpoint immunotherapy resistance provides key insights with direct relevance to human disease.

Despite the success of T-cell checkpoint blockade across a wide range of human cancers, our understanding of the factors driving both innate and acquired resistance to these therapies remains limited. Using B16 melanoma, we performed an unbiased investigation of acquired resistance to multicheckpoint blockade immunotherapy. To our knowledge, we are the first to use a system in which evolved changes in gene expression associated with increasing immunotherapy resistance could be analyzed separately in the tumor cells versus compensatory alterations in the surrounding TME. In this case, through augmenting glycolysis, oxidoreductase, and OxPhos, resistant melanoma was able to escape the initially curative immunotherapeutic pressure of cellular vaccination coupled with blockade of CTLA-4, PD-1, and PD-L1.

This immunotherapy-resistant melanoma defied Warburg theory, which states that tumor cells rely primarily on glycolysis for generation of ATP and downregulate mitochondrial OxPhos (41). Resistant 3I-F4 tumors increased both glycolysis and OxPhos concordantly, which we define as a hypermetabolic state. PGAM2, a glycolytic enzyme, was highly upregulated in immunotherapy-resistant tumor cells compared with parental cells. PGAM2 converts 2-phosphoglycerate to 3-phosphoglycerate, which is an essential step in glycolysis as well as contributing to anabolism (biosynthesis) of amino acids and nucleotides (42). Induction of PGAM2 activity may be a result of the oxidative stress response, which was enriched in both our resistant tumor cell line and patients with melanoma that do not respond to immunotherapy. Previous reports suggest that reactive oxygen species decrease PGAM2 acetylation at the lysine 100 (K100) site, which promotes its interaction with the cytosolic protein deacetylase sirtuin 2 (SIRT2) and activates PGAM2 (42). Whether there are epigenetic, transcriptional, or other posttranscriptional mechanisms involved in engaging metabolic pathways in immunotherapy-resistant tumors remains to be studied.

The overactive glycolysis pathway in resistant tumor cells can induce oxidative stress, which may be counterbalanced by upregulation of oxidoreductase pathways. ADH7, a gene in the oxidoreductase family, is an NAD(P)⁺/NAD(P)H coupling agent (43). Highly upregulated ADH7 in resistant tumor cells offers several advantages to highly glycolytic, immune-resistant tumors (43, 44). It reduces oxidative stress, generates reduced glutathione, a known scavenger of reactive oxygen species, and NAD(P)H, a substrate in mitochondrial OxPhos. We propose that upregulation of these glycolytic nodes and oxidoreductase pathways provides metabolic advantages to tumor cells, allowing them to increase mitochondrial OxPhos and foster a hypoxic microenvironment in which they have adapted to flourish.

Baseline CD8⁺ T-cell infiltration of B16 melanoma is poor (23); however, in response to triple checkpoint blockade therapy, cytotoxic effector density increased substantially. Resistant tumors blunted this capacity of immunotherapy to increase CD8⁺ T-cell density, likely due to the increased metabolic hostility of the TME. These hypermetabolic melanoma tumor cells can deplete nutrients in the TME, increase tumor-derived lactate and help promote a state of hypoxia. In this hostile microenvironment, cytotoxic CD8⁺ T cells, which require glycolytic metabolism for peak effector function, lose their metabolic fitness and associated effector function (45–48). We have also seen an increase in the suppressive capacity of Treg and MDSC in resistant tumors, which could also be a consequence of low glucose and the presence of tumor-derived lactate as these conditions expand immunosuppressive Treg and MDSC (37, 49).

In our resistant tumor line, we did not see evidence of substantial increased expression of alternative checkpoint pathways, nor did we see any changes in the IFNγ and JAK1 pathways; however, B16 melanoma, which is poorly immunogenic, may have a baseline insensitivity to IFNs (14, 50). We also did not observe downregulation of MHC class I or II complexes on the surface of resistant tumors (14, 51, 52). In the resistant tumors, we instead found increased class I and II MHC expression at both genetic and protein levels, perhaps reflecting loss of environmental immune pressure. Whereas metabolic adaptation appeared dominant in our system, we cannot deny that other biological processes may contribute to resistance to immunotherapy such as mutational load (13–15), neoantigen load (13), and copy-number loss (18, 19). Future studies may analyze the role of the mutational landscape in our resistant tumor model, although this was not the focus of this study.

We therapeutically targeted metabolic adaptation of resistant tumors with inhibitors of glycolysis (2DG and an LDH-A inhibitor) and mitochondrial complex 1 (metformin) but failed to reverse resistance to therapy. As we found that genes representing key nodes in both elevated glycolysis and OxPhos could individually confer resistance when introduced into the parental B16 melanoma, it is not surprising that blocking single pathways with relatively weak inhibitors was incapable of restoring checkpoint sensitivity. Whereas tumor cells rely on glycolysis and mitochondrial OxPhos, both metabolic pathways are equally important to the function of antitumor lymphocytes. Ideal therapeutic targets in this setting will likely be enzymes or pathway elements selectively induced by the tumor, rather than broadly active repressors of glycolytic or OxPhos metabolism which will also cripple T cells.

Finally, in a small cohort of patients with melanoma who failed to respond to both CTLA-4 and PD-1 blockade, we found upregulation of similar pathways associated with OxPhos and oxidoreductase activity. While we did not find strong signs of enhanced glycolysis in these resistant tumors, data from T-cell therapy–resistant melanoma cell lines suggests that enhanced glycolytic activity generally, including PGAM family upregulation specifically, is associated with acquisition of immunotherapy resistance (40). Given this, it is likely that analyzing whole tumor samples including the TME in our patient cohort caused us to lose glycolytic signal originating from the tumor due to compensatory dampening of glycolysis in the nutrient-starved stroma. Overall, the suggestive data from our patient cohort coupled with the published study of TIL therapy resistance (40) and the presented TCGA infiltrate data all suggested that the hypermetabolic state we observed to confer checkpoint resistance in our murine melanoma system was also a relevant pathway of immune escape in human melanoma patients.

A.R. Jaiswal reports current affiliation with MedImmune/AstraZeneca (the study was conducted at MD Anderson prior to A.R. Jaiswal joining MedImmune/AstraZeneca). Z.A. Cooper reports other from AstraZeneca (current employee and stock) outside the submitted work. M.A. Davies reports personal fees from Bristol-Myers Squibb (consultant), Novartis (consultant), Apexigen (consultant), and Array (consultant); grants and personal fees from Roche/Genentech (consultant; principal investigator of grant to institution) and GlaxoSmithKline (consultant; principal investigator of grant to institution); grants from AstraZeneca (principal investigator of grant to institution); and personal fees from Vaccinex (consultant) outside the submitted work. J.A. Wargo reports grants from GlaxoSmithKline, Roche/Genentech, Bristol-Myers Squibb, grants and personal fees from AstraZeneca and Merck, and personal fees from PeerView and Physician Education Resource outside the submitted work, as well as a patent for PCT/US17/53.717 pending to MD Anderson. D.S. Hong reports for the past 36 months research/grant funding from AbbVie, Adaptimmune, Aldi-Norte, Amgen, AstraZeneca, Bayer, Bristol-Myers Squibb, Daiichi Sankyo, Eisai, Fate Therapeutics, Genentech, Genmab, Ignyta, Infinity, Kite, Kyowa, Lilly, LOXO, Merck, MedImmune, Mirati, miRNA, Molecular Templates, Mologen, NCI-Cancer Therapy Evaluation Program (CTEP), Novartis, Pfizer, Seattle Genetics, Takeda, and Turning Point Therapeutics; travel, accommodations, and expenses from Bayer, LOXO, miRNA, Genmab, AACR, American Society of Clinical Oncology (ASCO), and Society for Immunotherapy of Cancer (SITC); consulting/advisory roles with Alpha Insights, Acuta, Amgen, Axiom, Adaptimmune, Baxter, Bayer, COG, Ecor1, Genentech, GLG, Group H, Guidepoint, Infinity, Janssen, Merrimack, Medscape, Numab, Pfizer, Prime Oncology, Seattle Genetics, Takeda, Trieza Therapeutics, and WebMD; and other ownership interests in Molecular Match (adviser), OncoResponse (founder), and Presagia (founder and adviser). M.A. Curran reports personal fees from ImmunoGenesis, Inc., Alligator Bioscience, Inc., ImmunOs, Inc., ImmunoMet, Inc., Oncoresponse, Inc., Pieris, Inc., Nurix, Inc., Aptevo, Inc., Merck, Inc., Oncomed, Inc., Kineta, Inc., Servier, Inc., Salarius, Inc., Xencor, Inc., and Agenus, Inc. outside the submitted work, as well as a patent for “Human PD-L1 Antibodies and Methods of Use Therefor” pending and licensed to ImmunoGenesis, Inc. and a patent for “Dual Specificity Antibodies Which Bind Both PD-L1 and PD-L2 and Prevent Their Binding to PD-1” pending and licensed to ImmunoGenesis, Inc. No potential conflicts of interest were disclosed by the other authors.

A.R. Jaiswal: Conceptualization, data curation, formal analysis, investigation, methodology, writing–original draft, project administration, writing–review and editing. A.J. Liu: Conceptualization, formal analysis, investigation, methodology, writing–original draft, writing–review and editing. S. Pudakalakatti: Formal analysis, investigation, writing–review and editing. P. Dutta: Formal analysis, investigation. P. Jayaprakash: Formal analysis, investigation. T. Bartkowiak: Formal analysis, investigation. C.R. Ager: Formal analysis, investigation. Z.-Q. Wang: Data curation, formal analysis. A. Reuben: Resources, data curation. Z.A. Cooper: Resources, data curation. C. Ivan: Data curation, formal analysis. Z. Ju: Data curation, formal analysis. F. Nwajei: Formal analysis. J. Wang: Data curation, formal analysis, supervision. M.A. Davies: Resources, supervision. R.E. Davis: Data curation, formal analysis, supervision, writing–review and editing. J.A. Wargo: Resources, supervision. P.K. Bhattacharya: Formal analysis, supervision. D.S. Hong: Supervision, writing–review and editing. M.A. Curran: Conceptualization, formal analysis, funding acquisition, investigation, writing–original draft, project administration, writing–review and editing.

The authors thank The University of Texas MD Anderson Melanoma Moonshot Program and the Williams family for providing funding for these studies. A.R. Jaiswal was supported by a Cancer Prevention & Research Institute of Texas (CPRIT) Research Training Award (RP170067). A.J. Liu is supported by a Marilyn and Frederick R. Lummis, Jr., MD, Fellowship in Biomedical Sciences. C.R. Ager was supported by NIH TL1 fellowships (TL1TR000369 and TL1TR000371).

The authors thank Scott Woodman, Midan Ai, Spencer Wei, Sangeeta Goswami, and Naveen Sharma for their input and consultation. This study used the South Campus Flow Cytometry & Cell Sorting Core, supported by NCIP30CA016672, and the Research Histology, Pathology, and Imaging Core, supported by P30 CA16672 DHHS/NCI Cancer Center Support Grant (CCSG).

The costs of publication of this article were defrayed in part by the payment of page charges. This article must therefore be hereby marked advertisement in accordance with 18 U.S.C. Section 1734 solely to indicate this fact.