Although understanding of T-cell exhaustion is widely based on mouse models, its analysis in patients with cancer could provide clues indicating tumor sensitivity to immune checkpoint blockade (ICB). Data suggest a role for costimulatory pathways, particularly CD28, in exhausted T-cell responsiveness to PD-1/PD-L1 blockade. Here, we used single-cell transcriptomic, phenotypic, and functional approaches to dissect the relation between CD8+ T-cell exhaustion, CD28 costimulation, and tumor specificity in head and neck, cervical, and ovarian cancers. We found that memory tumor–specific CD8+ T cells, but not bystander cells, sequentially express immune checkpoints once they infiltrate tumors, leading, in situ, to a functionally exhausted population. Exhausted T cells were nonetheless endowed with effector and tumor residency potential but exhibited loss of the costimulatory receptor CD28 in comparison with their circulating memory counterparts. Accordingly, PD-1 inhibition improved proliferation of circulating tumor–specific CD8+ T cells and reversed functional exhaustion of specific T cells at tumor sites. In agreement with their tumor specificity, high infiltration of tumors by exhausted cells was predictive of response to therapy and survival in ICB-treated patients with head and neck cancer. Our results showed that PD-1 blockade–mediated proliferation/reinvigoration of circulating memory T cells and local reversion of exhaustion occur concurrently to control tumors.
CD8+ T cells targeting cancer antigens infiltrate tumors, and the extent of the CD8+ T-cell infiltrate correlates with better prognosis (1). Nonetheless, expression of inhibitory immune checkpoints (IC), such as PD-1, CTLA-4, TIGIT, and TIM-3, in tumor-infiltrating CD8+ T cells contributes to their functional exhaustion (2, 3). Therapeutic blockade of ICs is effective in the treatment of cancers of different histologic types. However, even in responsive malignancies, not all treated patients experience meaningful clinical responses (2). Understanding mechanisms of exhaustion and, consequently, those of response to immune checkpoint blockade (ICB) is necessary for the clinical development of next-generation immunotherapies and the identification of biomarkers of response.
Through analysis of transcriptomic and epigenetic profiles in murine models of chronic viral infection or cancer, a picture has emerged whereby exhausted T cells could represent a CD8+ T-cell lineage distinct from effector or memory cells (4–6). Tissue-resident memory T (Trm) cells represent, likewise, a distinct lineage of cells emerging during immune responses to infection (7–9). These cells reside in tissues, are maintained regardless of antigen persistence, and ensure protection upon local reencounter with infectious agents. In patients with cancer, CD8+ T cells expressing Trm markers have been identified (10, 11) and seem to overlap with exhausted T cells at tumor sites. It remains unclear, however, whether these cells represent, as in murine models of infection, CD8+ T-cell lineages or a functional state acquired at tumor sites.
ICs can limit T-cell priming and expansion by interfering with CD28-mediated costimulation and can control effector functions through the impairment of T-cell receptor (TCR) complex signaling (2, 12). PD-1 engagement leads to dephosphorylation of molecules in the TCR complex signaling pathway (13, 14), and studies have challenged this dogma and propose CD28 as the main target of PD-1 (15–17). These findings may seem at odds with the distribution of PD-1 ligands, in particular PD-L1 expressed at effector sites (18), and with the correlation of PD-L1 expression at tumor sites to response to PD-1/PD-L1 blockade (19). Precise assessment of PD-1 and CD28 expression in tumor-specific CD8+ T cells could contribute to the clarification of the relative contribution of ICB-mediated PD-1 inhibitory signal relief on expansion and effector functions of tumor-specific T cells.
Here, in ovarian, cervical, and head and neck cancers, three epithelial malignancies exhibiting resistance to ICB, we characterized exhaustion in tumor antigen–specific CD8+ T cells. We showed that along with chronic stimulation of tumor-specific T cells, IC expression was sequentially acquired, leading to a population expressing the four checkpoints PD-1, TIGIT, CTLA-4, and TIM-3, which we called quadruple-positive (QP). Incremental checkpoint expression was accompanied by increased expression of Trm markers (8, 9), ectonucleotidase CD39 (20), and transcription factor TOX (21), and QP cells exhibited significant loss of CD28. QP cells were endowed with high cytotoxic potential. Circulating cancer antigen–specific T cells had a memory phenotype and expressed PD-1 and TIGIT only. At the tumor site, specific CD8+ T cells acquired an exhausted Trm-like phenotype. Circulating specific PD-1intCD28+ T cells responded to anti–PD-1 by enhancing their proliferation in response to antigen stimulation, and specific PD-1hiCD28+/− tumor-infiltrating lymphocytes (TIL) exhibited reversal of their functional exhaustion. QP cells were predictive of response to therapy and overall survival (OS) in patients with head and neck cancer treated by PD-1/PD-L1 blockade therapy. Our results show that the combination of proliferation/reinvigoration of circulating memory T cells that could replenish the tumor site combined with reversal of exhaustion at the tumor site contributes to PD-1/PD-L1 blockade–mediated tumor control.
Materials and Methods
Patient and healthy donor samples
Peripheral blood and tumor samples were collected from patients with head and neck, ovarian, and cervical cancer at the time of surgery for primary disease (54 patients) or for recurrence (1 patient) at the Institut Universitaire du Cancer de Toulouse – Oncopole (IUCT-O, Toulouse, France) in accordance to the Declaration of Helsinki, upon approval by the institutional review board (n°DC-2016-2656) and written informed consent. The study included: patients with stage FIGO IB1–IIIB HPV 16+ and/or 18+ cervical cancer with squamous and adenocarcinoma histology. Patients with stage FIGO IIIB–IVA high-grade serous or clear cell (1 patient) ovarian cancer and locally advanced, recurrent or metastatic head and neck squamous carcinoma. All patients had histologically documented tumors, were ≥18 years old at the time of study entry, were followed within a standard-of-care procedure, and had an Eastern Cooperative Oncology Group performance status of 0–2. Exclusion criteria included: known history of a positive test for hepatitis B, hepatitis C, human immunodeficiency, or hantaviruses; any condition contraindicated with blood sampling procedures; pregnancy or breast feeding; and active, suspected, or prior documented autoimmune disease or use of immunosuppressive medication. Patients did not receive any therapy during the 3 months prior to study entry. Blood samples from 14 healthy donors were obtained from the Etablissement Français du Sang.
Plasma was harvested after heparinized whole-blood centrifugation at room temperature. Peripheral blood mononuclear cells (PBMC) were isolated by density-gradient sedimentation using Ficoll-Hypaque (Sigma-Aldrich). Tumor samples were rapidly transported to the research facility on ice. On arrival, samples were rinsed with PBS (Sigma-Aldrich), subsequently minced on ice to smaller pieces (between 2–4 mm), and dissociated using C-Tubes (Miltenyi Biotec) and the gentleMACS Octo Dissociator (Program MultiC01_01; Miltenyi Biotec) in Iscove's Modified Dulbecco's Medium (IMDM, Sigma-Aldrich). PBMCs and tumor single-cell suspensions were cryopreserved in FBS (Gibco) containing 10% DMSO (Sigma-Aldrich).
Pretreatment tumor biopsies used for IHC analyses were obtained from a second cohort of patients with head and neck squamous cell carcinoma receiving ICB therapy with PD-1/PD-L1–blocking agents (nivolumab, n = 21; pembrolizumab, n = 1; durvalumab, n = 7; and avelumab, n = 1) to treat locally advanced or metastatic disease. Biopsies were either obtained up to 2 months prior to the first ICB dosing (5 patients) or retrieved from archival samples (25 patients: <1 year, 8 patients; ≥1 year and <2 years, 9 patients; ≥2 years and <3 years, 4 patients; and ≥3 years and <5 years, 4 patients). Samples were handled by the Biopathological Support Platform for Clinical Studies, IUCT-O. Response to therapy was evaluated by iRECIST criteria. Progressive disease (PD) was defined as the increase of >20% of target lesions or appearance and increase in size of new lesions in at least two CT scan evaluations performed at least 4 weeks apart. Partial response (PR) was defined as a decrease of >30% in target lesions and complete response (CR) as disappearance of target and nontarget lesions, both in at least two CT scans performed at least 4 weeks apart. Any response other than PD or PR/CR was considered as stable disease (SD).
Cell purification and phenotypic assessment
CD8+ T cells were enriched from PBMCs or tumor single-cell suspensions by positive magnetic selection (CD8 MicroBeads, human, Miltenyi Biotec) using OctoMACS Separator and MS Columns (Miltenyi Biotec). Cells (0.5 to 1 × 106 CD8+ T cells for PBMCs and 0.1 to 0.5 × 106 CD8+ T cells for TILs) were assessed phenotypically by staining with mAbs specific for CD3, CD4, CD8, CD45RA, CCR7, PD-1, TIGIT, TIM-3, CD28, CD103, CD69, CD49a, and CD39, as indicated, in PBS containing 5% FBS, for 15 minutes at 4°C. For intracellular and intranuclear staining, cells were fixed and permeabilized with fixation/permeabilization buffer for 45 minutes at 4°C, and stained in permeabilization buffer for 45 minutes at 4°C with mAbs specific for CTLA-4, granzyme B, perforin, TCF-1, and TOX using the Foxp3/Transcription Factor Staining Buffer Set (eBioscience). Antibodies used are listed in Supplementary Table S1. Where indicated, CD8+ T cells were stained with HLA class I dextramers containing NY-ESO-1 Peptides (Immudex) or pentamers containing CMV or EBV Peptides (ProImmune) for 30 minutes at room temperature prior to phenotypic assessment (see Supplementary Table S2 for the complete listing of multimers). Cells were analyzed using BD LSRFortessa X20, and data were analyzed using DIVA Software (BD Biosciences) or the vi-SNE Algorithm (Cytobank, Inc.).
In vitro differentiation of naïve CD8+ T cells
Naïve CD8+ T cells were enriched from PBMCs either by magnetic selection (Naïve CD8+ T Cell Isolation Kit, human, Miltenyi Biotec) using QuadroMACS Separator and LS Columns (Miltenyi Biotec) or by total CD8+ T-cell–positive magnetic selection (CD8 MicroBeads, human, QuadroMACS Separator, LS Columns) followed by staining with mAbs specific for CD8, CD45RA, CCR7, and CD28, and naïve (CD45RA+CCR7+CD28+) CD8+ T cells were sorted via FACS (BD FACSAria Fusion). Antibodies are listed in Supplementary Table S1. Cells were then labeled with CFSE (5 μmol/L; eBioscience), stimulated with anti-CD3/28 beads (Miltenyi Biotec) at a bead-to-cell ratio of 1:1, and maintained in culture in the presence of recombinant human (rh) IL2 (50 IU/mL; Miltenyi Biotec) and in the presence or absence of TGFβ (50 ng/mL; PeproTech) in IMDM (Sigma-Aldrich) supplemented with 1% penicillin–streptomycin Solution (Sigma-Aldrich), 1% MEM Nonessential Amino Acids Solution (Invitrogen), l-Glutamine (2 mmol/L, Invitrogen), and 10% human serum (Institut de Biotechnologies Jacques Boy). On day 7, cells were stained with CD3-, CD8-, CD45RA-, CCR7-, and CD28-specific mAbs (Supplementary Table S1) and analyzed by flow cytometry, as described above. In some experiments CD28− cells were sorted from day 7 TGFβ cultures via FACS (BD FACSMelody), restimulated with anti-CD3/28 beads (bead-to-cell ratio of 1:1) in the presence of rhIL2 (50 IU/mL) and in absence or presence of blocking PD-1 mAbs (10 μg/mL; Bristol-Myers Squibb), and assessed 7 days later for CD28 expression by flow cytometry.
Droplet-based single-cell RNA-sequencing and single-cell gene expression analysis
Tumors, minced as detailed above, were transferred into digestion medium (Tumor Dissociation Kit, human, Miltenyi Biotec) and dissociated using C-tubes and gentleMACS Octo Dissociator (Program h_TDK3, 37°C, Miltenyi Biotec). Samples were filtered using a 40-μm Nylon Mesh (BD Biosciences). Cell suspensions were then centrifuged at 300 × g at 4°C for 7 minutes, the supernatant was discarded. Cell pellets were resuspended in 1 mL Red Blood Cell Lysis Buffer (Miltenyi Biotec), incubated for 10 minutes at 4°C, centrifuged, and the pellets were resuspended in PBS containing 0.04% BSA (Euromedex). CD45+ cells were enriched by positive magnetic selection from single-cell suspensions (CD45 MicroBeads, human, OctoMACS Separator, MS Columns, Miltenyi Biotec), and cells were counted to determine the proportion of live cells. Only samples containing >90% live cells were used for single-cell RNA-sequencing (scRNA-seq) experiments. CD45+ cells (2 to 4 ×105 cells) were stained with barcoded TotalSeq-A mAb (BioLegend).
Single-cell libraries (3′ gene expression and antibody-derived tag fractions) were generated using the Chromium Controller Instrument and Chromium Single Cell 3′ Library and Gel Bead Kit v3 according to the manufacturer's protocol (10× Genomics) with some modifications as described previously (22). To detect barcoded TotalSeq-A antibodies, an ADT library was constructed as previously described for CITE-seq (22). Single-cell library size and quality were confirmed on Fragment Analyzer (Agilent). KAPA Quantification Kit for Illumina platforms (Kapa Biosystems, Roche) was used to quantify library. Samples were pooled in equimolar fashion with desired proportions for the two library types (cDNA library fraction at 90% and ADT library at 10%). The libraries were sequenced on a NextSeq 550 (Illumina) in pair-end sequencing 28 bp (read1) × 91 bp (read2) and a single index 8 bp in length. Raw data (FastQ files) for expression and antibody detection were computed with CellRanger 3.0 and the GRCh38 transcriptome as reference (https://support.10xgenomics.com/single-cell-gene-expression/software/pipelines/latest/using/count). Data were then loaded in an R session with the Seurat 3.0 toolkit package involving the normalization and variance stabilization package sctransform (23). Samples were individually filtered using unique molecular identifier (<30′000) and percentage of mitochondrial genes (<0.25%) criteria. Using Seurat, datasets were reduced by principal component analysis using the 11 first principal components to reduce dimensionality by t-distributed stochastic neighbor embedding (t-SNE). A resolution parameter set the granularity at 1.2 for the clustering by the K-nearest neighbor graph-based clustering approach of Seurat's FindClusters function. CD8+ T lymphocytes were selected using Single-Cell Virtual Cytometer software (https://sites.google.com/site/fredsoftwares/products/single-cell-virtual-cytometer) using the sum of CD3D, CD3E, and CD3G gene expression and CD8B gene expression. Wilcoxon P values for differentially expressed genes were adjusted with Benjamini–Hochberg and illustrated in a volcano plot. The maturation trajectories of the selected T cells were computed with FateID as described previously (24), specifying CD8+ T cells in cluster 5 (Fig. 3A) as an endpoint. The gene set enrichment analysis of the coexpression nodes, here, taken as lists of genes, were computed with Autocompare-ZE (https://sites.google.com/site/fredsoftwares/products/autocompare-ze) for enrichment for gene sets from the C5 collection of gene ontology (GO) database (GSEA/MSigDB database, v6.3; refs. 25–27). ScRNA-seq data have been deposited in NCBI's Gene Expression Omnibus (GEO; ref. 28) and are accessible through GEO series accession number GSE148162.
T-cell functional assessment
CD8+ T cells isolated from tumor single-cell suspensions were stimulated with phorbol 12-myristate 13-acetate (PMA, 100 ng/mL; Sigma-Aldrich) and ionomycin (1 μg/mL; Sigma-Aldrich) or plate-bound anti-CD3 (1 μg/mL; eBioscience) in the presence of anti-CD107a (Supplementary Table S1) for 6 hours in IMDM supplemented as described above. Brefeldin-A (10 μg/mL; Sigma-Aldrich) was added 1 hour after the beginning of the incubation. Cells were then stained with mAbs, as described above, specific for CD3, CD4, CD8, and PD-1, and cytokine production was assessed by intracellular staining using IFNγ- and TNFα-specific mAbs (Supplementary Table S1). The cells were analyzed by flow cytometry (BD LSRFortessa X20).
For the assessment of NY-ESO-1–specific cells, CD8+ T cells were magnetically sorted from PBMCs and tumor single-cell suspensions, as already described, from patients with ovarian cancer exhibiting antibody responses to NY-ESO-1. Antibody responses to NY-ESO-1 were assessed in patient plasma by ELISA, as described previously (29) and detailed hereafter. Eighteen overlapping 20- to 24-amino acid long peptides encompassing the full-length NY-ESO-1 sequence (25 μg/mL each in PBS; Peptide 2.0 Inc) were coated on Nunc MaxiSorp Flat-Bottom ELISA Plates (Thermo Fisher Scientific) overnight at 4°C. Plates were then washed with PBS, 0.005% Tween (Sigma-Aldrich), incubated for 2 hours at 37°C with blocking buffer (PBS, 2% BSA), and washed with PBS, 0.005% Tween. Plasma (diluted 1/100 in blocking buffer) was incubated for 2 hours at room temperature and washed using the same buffer. Plates were then incubated for 1 hour at room temperature with goat anti-human IgG secondary antibodies (Sigma-Aldrich; 1 μg/mL in blocking buffer) and washed. ELISA was developed using the Alkaline Phosphatase Yellow (pNPP) Liquid Substrate System for ELISA (Sigma-Aldrich). Isolated CD8+ T cells (20,000 to 150,000 CD8+ T cells from TILs and 2 to 5 × 105 CD8+ T cells from PBMCs, per well) were stimulated with NY-ESO-1 peptides (1 μmol/L; Peptide 2.0 Inc) in the presence of autologous CD14+ cells sorted from PBMCs by positive magnetic selection (CD14 MicroBeads, human, OctoMACS Separator and MS Columns, Miltenyi Biotec) at a 1.5-to-1 CD14+-to-CD8+ cell ratio, and rhIL2 (50 IU/mL) in the absence or presence of PD-1–blocking mAb (10 μg/mL) in IMDM supplemented as above. Day 5–7 cultures were either stained with HLA class I dextramers containing NY-ESO-1 peptides and with anti-CD8 or stimulated or not with NY-ESO-1 peptides (1 μmol/L) and assessed for IFNγ and TNFα production by intracellular staining, as detailed above.
Quantitative multiplex IHC
Sequential chromogenic IHC was performed on 4-μm-thick tumor tissue sections based on the protocol described by Glass and colleagues with some modifications (30) using mAbs listed in Supplementary Table S3. Briefly, slides were dewaxed, subjected to heat-mediated antigen retrieval (Target Retrieval Buffer pH9 K800021-2, Dako Agilent), and immunostained using 3-amino-9-ethylcarbazole (AEC, Dako Agilent) as chromogen according to the manufacturer's recommendations. To preserve antigens, slides were then mounted with an Aqueous Mounting Medium (Dako Agilent) containing 60% glycerol, 40% distilled H2O, and 10% (w/w) saturated sucrose-H2O (584 mmol/L; ref. 30) and scanned using Pannoramic 250 Flash III Scanner (3DHISTECH Ltd.). After scanning, AEC and antibodies were removed. Slides were unmounted and washed with distilled H2O before being discolored with increasing gradients of absolute ethanol (70%–90%, Sigma-Aldrich) according to the SIMPLE1 method (30), and rinsed with distilled H2O. Antibody stripping was performed by incubation of slides for 30 minutes at 56°C in a freshly prepared stripping buffer [1% SDS (w/v, Sigma-Aldrich), 0.2 mol/L Tris HCl pH 6.8 (CliniScience), and 0.1 mol/L β-Mercaptoethanol (Sigma-Aldrich) in distilled H2O; ref. 31], followed by rinsing with distilled H2O for 45 minutes. The entire sequence was repeated for each antigen assessed.
For each case, at least one region of interest (ROI) of 5,000 × 5,000 pixels (1,200 μm²) was selected and extracted from each of the virtual slides using ImageScope (Aperio Leica Biosystems). Extracted images were registered using ImageJ “Register Virtual Stack Slices” plugin and color-deconvoluted using FIJI (1.52n). Nuclei segmentation was performed with ILASTIK software. Cell identification, staining evaluation, and results export in csv format was performed using CellProfiler software. Pathologists defined fluorescence value for each cell and each marker in the ROI and the positivity thresholds for each marker. These thresholds were used for the determination of CD3+CD8+ (CD8+ T cells) and CD3+CD8+TIM-3+ (TIM-3+CD8+ T cells) cell proportions among all nuclei. The median of CD3+CD8+ and CD3+CD8+TIM-3+ cell proportions in all patients was used as cutoff to define high (≥median) and low (<median) infiltration for each population.
In vivo experiments
Female C57BL/6 mice, 7 weeks old, were purchased from Janvier Labs. Experimental protocols were approved by regional Ethic Committee of Toulouse Biological Research Federation (C2EA-01, FRBT) and by the French Ministry for Higher Education and Research. The European directive 2010/63/EU was followed for guidelines on animal welfare. TC1 cells expressing the HPV 16-E6 and E7 proteins were developed in the laboratory of T.C. Wu (Department of Pathology, School of Medicine, Johns Hopkins University, Baltimore, MD). Cells were cultured in complete RPMI1640 Medium (Life Technologies) for 2–6 passages and tested negative for Mycoplasma.
When 8 weeks old, mice were anesthetized by intraperitoneal injection of anesthetic mix [10 μL/kg; made up of ketamine (100 mg/kg) and xylazine (10 mg/kg); Centravet], and implanted in the oral cavity (intracheek) with 3 × 104 TC1 cells suspended in a final volume of 10 μL of PBS using a Hamilton syringe and 30G needle. At day 10 and 18, mice were sacrificed by intraperitoneal injection of anesthetic mix (10 μL/kg), followed by vertebral dislocation. Oral tumors were harvested, cut in pieces, and enzymatically digested using the Mouse Tumor Dissociation Kit (Miltenyi Biotec), C-tubes, and gentleMACS Octo Dissociator (program 37C_m_TDK_1). Digestion was followed by filtration through a 70-μm cell strainer. Cells were resuspended in PBS containing 2% FBS, anti-CD16/CD32 (Supplementary Table S1), and 1:1,000 Fixable Viability Dye Stain 700 (BD Biosciences). Staining with mAbs specific for cell surface markers (Supplementary Table S1) and flow cytometry analyses were performed as described above for human samples.
Normality was assessed using the Shapiro–Wilk test. For normally distributed values, the t test was used for paired or unpaired data. When the values were not normally distributed, the comparison of variables was performed with Wilcoxon or Mann–Whitney test for paired and unpaired data, respectively. Lines and error bars in scatter plots represent mean ± SD. For correlations, Pearson test was used to compare variables. For OS analyses, ICB-treated patients were subdivided into two groups according to high (≥median) and low (<median) tumor infiltration by CD8+ T cells or TIM-3+CD8+ T cells, Kaplan–Meier curves were plotted, and P values determined by log-rank test. Results of statistical analyses are annotated as follows in figures: ns, not significant (P ≥ 0.05); *, P < 0.05; **, P < 0.01; ***, P < 0.001; ****, P < 0.0001. All analyses were performed with GraphPad Prism 7 software.
IC expression defines six tumor-infiltrating CD8+ T-cell populations
We assessed ex vivo CD8+ ovarian, cervical, and head and neck cancer (HNC) TILs for PD-1, TIGIT, CTLA-4, and TIM-3. Each of the four checkpoints was expressed in the three tumor types (Fig. 1A; Supplementary Fig. S1A). PD-1 was on average expressed in a higher proportion of cells, followed by TIGIT, CTLA-4, and TIM-3. Although TIGIT could be expressed both independently and alongside PD-1 (Supplementary Fig. S1A, left contour plots), CTLA-4 and TIM-3 expression appeared systematically associated to that of PD-1 (Supplementary Fig. S1A, center and right contour plots, respectively). This was supported by correlation analyses, which revealed significant positive correlation between the proportion of PD-1–expressing cells and those expressing CTLA-4 and TIM-3 but not TIGIT (Supplementary Fig. S1B). We then analyzed CTLA-4 and TIM-3 expression in cells expressing or not PD-1 and TIGIT (Fig. 1B). The majority of PD-1−TIGIT− cells did not express CTLA-4 and TIM-3, which we named “quadruple-negative” (QN). PD-1+TIGIT− and PD-1−TIGIT+ cells did not express CTLA-4 and TIM-3. We named these cells “PD-1 single” (PD-1s) and “TIGIT single” (TIGITs), respectively. PD-1+TIGIT+ cells could be divided into three populations according to CTLA-4 and TIM-3 expression: (i) expressed neither CTLA-4 nor TIM-3; (ii) expressed CTLA-4 only; and (iii) expressed both, called “double-positive” (DP), “triple-positive” (TP), and “quadruple-positive” (QP), respectively (Fig. 1B). Among the 16 possible coexpression profiles (Fig. 1B, bottom), the six populations described above were the most frequently and consistently detected in the three tumor types. The QP population showed the highest interindividual variation (range 0.7%–62%, median 20.69%). Similar analyses from circulating memory/effector CD8+ T cells from healthy individuals and patients with cancer showed that only QN, PD-1s, TIGITs, and DP populations were detectable, implying that acquisition of CTLA-4 and TIM-3 expression takes place in situ (Supplementary Fig. S2).
In CD8+ TILs, we found that cells expressing both PD-1 and TIGIT exhibited higher staining intensity for both markers compared with cells expressing each alone (Fig. 1B; Supplementary Fig. S1A, left contour plots). Analysis of mean fluorescence intensity (MFI) showed that PD-1 MFI incrementally and significantly increased in parallel to the number of expressed ICs in each PD-1+ population (Fig. 1C, left). A similar pattern was observed for TIGIT, although the MFI did not vary significantly between the TIGITs and DP populations (Fig. 1C, right).
Altogether, these analyses revealed a conserved pattern of IC expression in the three tumor types, which defined six CD8+ T-cell populations. Analysis of PD-1 and TIGIT suggested an increase in their expression alongside a possible sequential acquisition of CTLA-4 followed by TIM-3. In an orthotopic mouse model of HNC, PD-1 MFI at the tumor site was, likewise, increased in PD-1+CD8+ T cells coexpressing TIM-3 compared with those expressing PD-1 only (Supplementary Fig. S3). In further support of the sequential acquisition model, the proportion of PD-1+TIM-3+ cells was higher at day 18 than day 10 tumors (Supplementary Fig. S3A).
IC acquisition is paralleled by Trm markers and CD39 expression, and CD28 loss
We assessed the expression of Trm markers, CD103, CD69, and CD49a (8), and CD39, also expressed in CD103+CD8+ T cells at tumor sites (20), in the six CD8+ T-cell populations identified (Fig. 1D; Supplementary Fig. S4). Trm markers and CD39 were absent or expressed in low proportions in QN and TIGITs populations. Their proportions were significantly higher among PD-1s cells versus QN cells. Whereas no difference in Trm marker expression was detected between PD-1s and DP cells, CD39 expression was significantly higher in the DP population. Expression of all markers significantly increased within the TP population relative to DP cells and was further increased in the QP population. These data further supported the serial acquisition pattern of ICs by CD8+ T cells at the tumor site.
The sequential acquisition of ICs is indicative of sustained T-cell stimulation and could imply disparities in the differentiation stage between cells expressing few or several ICs. Assessment of the differentiation stage of tumor-infiltrating CD8+ T cells showed that the population was homogenous and mostly composed of effector memory cells (CD45RA−CCR7−; Fig. 2A). Effector memory cells are composed of cells expressing CD28 and cells in a more advanced differentiation stage that are CD28− (32). We found that effector memory TILs contained a significantly lower proportion of CD28+ cells versus circulating effector memory cells (Fig. 2B; Supplementary Fig. S5A). The CD28− population at the tumor site was enriched in cells expressing the four ICs (Fig. 2C), suggesting loss of CD28 alongside IC acquisition. Analysis of CD28 expression according to IC coexpression profiles defined above showed significant loss in CD28 expression in the QP population (Fig. 2D; Supplementary Fig. S5B). Unsupervised integration of flow cytometry data using vi-SNE (33) identified a QP population expressing Trm markers and CD39 while being CD28− (Fig. 2E).
We sought to identify tumor-related factors underlying CD28 loss. The QP population was characterized by CD103 and high PD-1 expression, two features that depend on TGFβ signaling (9, 34). We found that in vitro stimulation of naïve CD8+ T cells in the presence of TGFβ led to an enhanced loss of CD28 (Fig. 2F; Supplementary Fig. S5C). This was not simply due to progression in the differentiation stage as TGFβ led to delayed differentiation (Fig. 2G). Loss of CD28 could not be recovered by PD-1 blockade of stimulated CD28− cells, isolated from TGFβ cultures (Supplementary Fig. S5D).
ScRNA-seq reveals active immune pathway signatures in IC-positive CD8+ T cells
To characterize the QP population, we performed scRNA-seq of total tumor-infiltrating CD45+ cells isolated from 2 patients with HNC. A total of 17.9% were identified as CD8+ T cells and their clustering identified 11 clusters (Fig. 3A, top left). Clusters 0, 1, 4, 5, and 9 showed the highest HAVCR2 (TIM-3) expression (Fig. 3A). CTLA4 was found in the same clusters as well as in clusters 7 and 3. Expression of TIGIT was found in most clusters, although it was more frequent in clusters expressing HAVCR2. PDCD1 (PD-1) was also detected in most clusters, although the frequency of positive cells was lower than expected, presumably owing to low levels or low detection of this specific mRNA. Clusters 1, 4, 5, and 9 contained the highest proportion of cells exhibiting a QP profile and also contained cells expressing ITGAE (CD103), ITGA1 (CD49a), and ENTPD1 (CD39; Fig. 3A). We then performed differential mRNA expression analysis between regrouped clusters 1, 4, 5, and 9 and all other clusters. We identified 761 genes with higher expression in the QP clusters and 94 with higher expression in the other clusters (Fig. 3B). Genes shown in green in the volcano plot (Fig. 3B) correspond to markers we had used to characterize these populations (Fig. 2E), and their expression (Fig. 3C) confirmed that clusters on the right side of the t-SNE plot were most representative of the QP population.
Consistent with IC expression, QP clusters expressed TOX, which encodes a DNA-binding protein involved in T-cell exhaustion (Fig. 3B and C; ref. 21). Differential analysis identified a clear cytotoxicity signature in QP clusters, including granzymes A, B, and H (GZMA, GZMB, and GZMH, respectively), perforin (PRF1), and granulysin (GNLY; Fig. 3B and C). Twenty-one transcription factors were previously identified as being key for memory CD8+ T-cell development and function (35). Of those, seven were differentially expressed in our analysis with SOX-4 (SOX4), Class E bHLHe40 (BHLHE40), PRDM1, and RUNX2 being more expressed in the QP clusters and KLF2, T-cell factor 1 (TCF-1, TCF7), and BACH2 in the other clusters (Fig. 3B and D). Expression of SOX-4 is promoted by TGFβ (36) and induces CXCL13 production by CD4 T cells (37). CXCL13 was more expressed in QP clusters (Fig. 3B and D). Among transcription factors that were less expressed in QP clusters, KLF2 has been shown to be downregulated in Trm cells, leading to downmodulation of S1PR1 expression (7). Differential analysis showed lower S1PR1 expression in QP clusters (Fig. 3B and D). TCF-1 (TCF7) was also less expressed in QP clusters. In addition to TCF7 and CD28, other early memory T-cell markers were also less expressed in QP clusters, including CCR7 and IL7R (Fig. 3B and D). Differential analysis also revealed higher expression of the chemokine receptor gene CXCR6 in QP clusters (Fig. 3B and D).
We then computed a pseudotime maturation trajectory anchored by cluster 5 cells as an endpoint (Fig. 3A). Visualizing gene expression in cells along this trajectory showed TCF7, IL7R, and CCR7 expression in cells clustered at early stages of the maturation trajectory, whereas cells expressing ICs, Trm and cytotoxicity markers, as well as CXCL13 and CXCR6, clustered at later stages of the trajectory (Fig. 3E). Respective expression profiles of chronologically ordered cells were then Z-score transformed to reveal putative temporally restricted gene sets. This identified 21 gene expression nodes, which were then analyzed for enrichment of signatures from the C5 collection of GO database. This unveiled the prominence of translational signatures during the earliest maturation stages (nodes 16–20), followed in the latest stages by signatures of immune responses (nodes 1–3; Supplementary Fig. S6). These results revealed that exhausted Trm-like QP cells are the most immunologically active CD8+ T cells at the tumor site.
IC QP CD8+ T cells are tumor-specific and respond to PD-1 inhibition
We next assessed QP cell effector function, antigen specificity, and response to PD-1 inhibition. Tumor-infiltrating CD8+ T cells showed progression in TOX expression along IC acquisition (Fig. 4A and B; Supplementary Fig. S7A and S7B). An opposite pattern was observed for the expression of TCF-1 (Supplementary Fig. S8A). Absence of TCF-1 in late-stage populations at the tumor site was in agreement with its progressive loss along differentiation in circulating CD8+ T cells (Supplementary Fig. S8B and S8C). Expression of granzyme B and perforin, inferred from scRNA-seq analyses, was also confirmed at the protein level. The proportion of cells expressing both proteins was higher in the QP population than in other CD8+ TILs (Fig. 4A and B). This was consistent with their CD28− phenotype indicating advanced differentiation (32).
To validate the effector potential of QP cells, we stimulated tumor-infiltrating CD8+ T cells with PMA/ionomycin (i.e., bypassed proximal TCR signaling and IC inhibition). PMA/ionomycin stimulation perturbed, in some samples, TIGIT, CTLA-4, and TIM-3 expression but not that of PD-1. We, therefore, analyzed cytokine production and CD107a surface expression in CD8+ populations defined by PD-1 expression (Fig. 4C). PD-1− cells encompassed the QN and TIGITs populations, PD-1int corresponded to PD-1s and DP populations, and PD-1hi represented the TP and QP populations (Fig. 1B and C). Our data showed that PD-1hi cells were able to release cytotoxic granules and produce cytokines (Fig. 4D and E). Compared with the two other populations, PD-1hi cells encompassed fewer polyfunctional cells, that is, able to produce both IFNγ and TNFα, and more IFNγ+TNFα− cells (Fig. 4E), a feature that has been associated with exhaustion (38). Accordingly, upon anti-CD3 instead of PMA/ionomycin stimulation, PD-1hi cells comprised fewer cytokine-producing cells than less exhausted, PD-1− and PD-1int, populations (Fig. 4F).
NY-ESO-1, a cancer testis antigen expressed in ovarian cancer, induces both antibody and T-cell responses (29). Several NY-ESO-1–derived CD8+ T-cell epitopes are presented by frequently expressed HLA class I alleles (39). Within the ovarian cancer cohort, we identified patients with serologic responses to the antigen, indicative of an ongoing T-cell response (29). We used HLA class I/NY-ESO-1 fluorescent multimers to detect antigen-specific T cells in these patients. The proportion of antigen-specific T cells was higher in CD8+ cells isolated from TILs than from PBMCs, suggesting that they were preferentially attracted, proliferated, and/or were retained at tumor sites (Fig. 5A and B; Supplementary Fig. S9A). Similar proportions of CD8+ T cells specific for chronic viruses were detected in PBMCs and TILs (Fig. 5A and B; Supplementary Fig. S9A), further confirming the specific in situ accumulation of tumor-specific T cells. NY-ESO-1–specific T cells in the periphery expressed PD-1 and TIGIT, both with intermediate MFIs, whereas they had a QP phenotype at the tumor site (Fig. 5A and B; Supplementary Fig. S9). In contrast, bystander virus-specific T cells never acquired the QP phenotype at the tumor site (Fig. 5A and B; Supplementary Fig. S9). NY-ESO-1–specific CD8+ TILs were CD103+ and CD39+ (Supplementary Fig. S10A). Altogether, these data demonstrated that acquisition of the QP phenotype results from T-cell stimulation in the tumor microenvironment and inferred that the QP population is solely composed of tumor-specific T cells. This is consistent with the high expression of MKI67, encoding Ki-67, in QP clusters (Fig. 3B).
Significantly fewer CD28+ cells among NY-ESO-1–specific T cells from TILs was seen versus PBMCs (Supplementary Fig. S10B). Antigen-specific T cells were CD28+PD-1int in PBMCs and CD28+/−PD-1hi at the tumor site. To examine the functional consequences of these opposite phenotypes, we stimulated circulating and tumor-infiltrating CD8+ T cells from NY-ESO-1 seropositive patients with NY-ESO-1 peptides in the absence or presence of blocking PD-1 mAb. Whereas PD-1 blockade led to enhanced growth of antigen-specific T cells from PBMCs, no effect was observed on the proliferation of antigen-specific T cells from TILs (Fig. 5C). PD-1 inhibition was not, however, ineffective at the tumor site, as it led to increased production of effector cytokines by antigen-specific T cells (Fig. 5D and E).
IC-positive CD8+ T cells predict clinical response to PD-1/PD-L1 blockade
Using quantitative multiplex IHC (30, 31), we assessed QP cells in pretreatment tumor samples from a cohort of 30 patients with metastatic or recurrent head and neck squamous cell carcinoma treated with ICB targeting the PD-1/PD-L1 axis. The proportion of CD8+ or QP TILs was then correlated to response to therapy and OS. Because TIM-3 was the final checkpoint acquired at the tumor site, we used it as a surrogate marker of the QP population. Examples of staining obtained in samples exhibiting a high or low TIM-3+CD8+ T-cell infiltrate are shown in Fig. 6A (top and bottom, respectively). Quantification of QP (0.03%–15.98% of all nuclei, median = 1.9) and total CD8+ (0.15%–23.49% of all nuclei, median = 5.57) TILs showed that their proportion was variable among patients. A correlation was observed between the CD8+ infiltrate and response to therapy, although it did not reach statistical significance (Fig. 6B). Infiltration by QP cells significantly correlated with response to therapy and OS (Fig. 6B and C), whereas the total CD8+ infiltrate did not significantly correlate with OS (Fig. 6C). The median survival of patients in the high- and low-CD8+ T-cell groups was 12.7 and 8.8 months, respectively. Median survival of patients with low QP infiltrate was 5.8 months, whereas that of patients with high QP infiltrate was not reached. These results corroborate the major contribution of QP cells to antitumor immunity.
Our results put forward T-cell exhaustion as an indicator of spontaneous adaptive immune response to tumors and, as such, as a biomarker of response to ICB. We showed that exhaustion was acquired at the tumor site through the sequential expression of ICs solely in tumor antigen–specific T cells but not in bystander T cells. Specific cells additionally acquire Trm markers and CXCR6, allowing for their in situ residency. We also showed that exhausted Trm-like cells were endowed with high cytotoxic and functional potential but have lost, for a large part, CD28 and TCF-1 expression. PD-1 blockade could revert functional exhaustion of tumor antigen–specific CD8+ TILs in vitro, while it enhanced proliferation of their circulating CD28+ counterparts. Their tumor specificity, high functional potential, and sensitivity to PD-1 blockade supported their contribution to responsiveness to ICB in vivo, which was inferred by the correlation between the proportion of exhausted CD8+ TILs and response to therapy and survival following ICB.
On the basis of murine models of chronic infection, exhausted T cells are proposed as a lineage of CD8+ T cells distinct from memory T cells (4, 5). Our results showed that circulating tumor antigen–specific CD8+ T cells were memory CD28+ cells. These cells, like nonexhausted memory circulating PD-1+CD8+ T cells from healthy individuals (40), expressed lower PD-1 compared with their tumor-infiltrating counterparts. Accordingly, circulating tumor-specific T cells, which did not express TIM-3 or CTLA-4, could substantially expand in vitro upon antigen stimulation, which was enhanced in the presence of anti–PD-1. Murine models of exhaustion in cancer, largely relying on implantable syngeneic tumor cell lines, have quicker kinetics than human cancers. In one such model, transcription factors PRDM1 and c-MAF are proposed to coordinate a coinhibitory gene module leading to the concomitant coexpression of several ICs (41). In agreement with these data, we showed that PRDM1 was more expressed in exhausted Trm-like cells. However, our results argue in favor of a progressive acquisition of exhaustion with sequential gain of ICs accompanied by that of Trm markers. This is in agreement with the model that established degrees of exhaustion severity (21), where the transcription factor TOX is associated to exhaustion. Our data showed expression of TOX in QP cells and sequential gain of TOX expression at the tumor site. A study has shown, in a murine model of chimeric antigen receptor T-cell transfer, that transferred T cells acquire an exhausted state at the tumor site through TOX- and TOX2-dependent mechanisms (42).
Seminal work shows that CD8+ TILs express Trm markers and that CD103 is involved both in T-cell retention at the tumor site and in tumor recognition (10, 11). Another study shows that the ectonucleotidase CD39 is expressed in a subpopulation of CD103+CD8+ T cells enriched for tumor specificity (20). In agreement with these studies, we showed that Trm markers and CD39 expression is gained along the sequential acquisition of ICs and that they had high expression in antigen-specific T cells. We also showed that exhausted cells exhibited lower S1PR1 expression, consistent with their in situ retention, and expressed the chemokine receptor CXCR6. CXCR6 is expressed in liver Trm cells (43), and its ligand, CXCL16, which can be produced in soluble and membrane-bound forms (44), is expressed by tumor cells of different histologic types and is associated with increased tumor T-cell infiltration (45). We propose that CXCR6 expression in exhausted Trm-like cells facilitates their migration toward and their interaction with tumor cells.
We showed that TGFβ, involved in the induction of CD103 (9) as well as in high PD-1 expression (34), could also be implicated in CD28 loss. To our knowledge, this is not a function that has been attributed to TGFβ, although it is consistent with its immune-suppressive role. Expression of SOX-4 and CXCL13 in the QP population further supported the involvement of TGFβ in its in situ development. Indeed, TGFβ induces SOX-4 (36, 37), which mediates CXCL13 production by CD4+ T cells (37). CXCL13 has been involved in tertiary lymphoid structure (TLS) induction in inflammatory diseases (37), and its expression by tumor-infiltrating CD8+ T cells associates with high B-cell recruitment and TLSs (46, 47).
PD-L1 expression in peripheral tissues and the ability of PD-1–recruited SHP2 to dephosphorylate proteins involved in proximal TCR signaling implies a central role of PD-1 in blocking functions of effector T cells (13, 14). Studies show that PD-1 signaling inhibits CD28 more efficiently compared with TCR proximal signaling molecules (15), and, accordingly, early proliferation of circulating PD-1+CD28+CD8+ T cells could be associated to response to ICB in patients with non–small cell lung cancer (16, 17). We showed significant CD28 loss, specifically in exhausted tumor-specific CD8+ TILs. However, circulating antigen-specific T cells were mostly CD28+. Our functional analyses showed that in vitro PD-1 inhibition differentially modulated proliferation and effector functions in circulating and tumor-infiltrating antigen-specific T cells. These results are in agreement with previous studies (15–17) and could imply that low PD-1 expression in circulating antigen-specific T cells will preferentially affect CD28 signaling, whereas overexpression of PD-1 in exhausted Trm-like cells allows for inhibition of TCR signaling in addition to CD28. It is noteworthy that, in a reconstituted membrane in vitro setting, PD-1–mediated dephosphorylation of CD28 occurs at low PD-1 densities, whereas TCR signaling molecule dephosphorylation requires a higher density of PD-1 molecules (15), which could correspond to PD-1 densities found in PD-1hi exhausted T cells.
Finally, in agreement with their tumor specificity and high cytotoxic and effector potential, exhausted Trm-like T cells were predictive of response to ICB and survival in a cohort of patients with HNC. These results may contradict studies suggesting that TCF-1+CD8+ TILs, termed progenitor exhausted cells, are potential players in PD-1/PD-L1 blockade in vivo efficacy (48–51). However, these studies largely rely on murine models whereby the degree of exhaustion attained in few weeks is not necessarily comparable with that found in patients. One study shows that TCF-1 expression in tumor antigen–specific T cells in patients is detected in only a low proportion (∼1%) of circulating antigen-specific cells (51). Another study shows that the proportion of TCF-1+ cells does not correlate with response to therapy but does correlate with survival in responder patients only (48). Another report found that the ratio between TCF-1+ and TCF-1− cells, rather than the proportion of positive cells, correlates to clinical responses (49). In our study, TCF7 mRNA and TCF-1 protein were only detected in non-QP cells at the tumor site. Our results showed significant correlation between TIM-3+CD8+ T cells, which did not express TCF-1, and both response to therapy and OS following ICB. These results do not exclude a potentially important role of TCF-1+ tumor-specific CD8+ T cells in mediating response to therapy. We believe that the presence of TIM-3+CD8+ T cells is a direct indicator, that is, biomarker, of a spontaneous adaptive response to the tumor that can be mobilized by ICB. Altogether, our results imply that the combination of proliferation/reinvigoration of circulating memory T cells, that will replenish the tumor site, combined to reversion of exhaustion, even temporary (6), at the tumor site contribute to PD-1/PD-L1–mediated tumor control.
Disclosure of Potential Conflicts of Interest
C. Gomez-Roca reports receiving a commercial research grant from Bristol-Myers Squibb; reports receiving speakers bureau honoraria from Bristol-Myers Squibb, Hoffmann-La Roche; and Pierre Fabre; and is a consultant/advisory board member for Bristol-Myers Squibb. S. Motton is a consultant for Intuitive Surgical. J.-P. Delord reports receiving speakers bureau honoraria from Roche, MSD, Bristol-Myers Squibb, and AstraZeneca. P. Rochaix reports receiving other commercial research support from Roche Diagnostic/Ventana and MSD. M. Ayyoub reports receiving a commercial research grant from Roche/Genentech (imCORE), reports receiving speakers bureau honoraria from AstraZeneca and Bristol-Myers Squibb, and is a consultant/advisory board member for AstraZeneca and Pierre Fabre. No potential conflicts of interest were disclosed by the other authors.
Conception and design: C.-C. Balança, C.-M. Scarlata, M. Michelas, C. Devaud, V. Sarradin, J.-P. Delord, M. Ayyoub
Development of methodology: V. Sarradin, C. Martinez Gomez, M. Tosolini, F. Pont
Acquisition of data (provided animals, acquired and managed patients, provided facilities, etc.): C.-C. Balança, C.-M. Scarlata, M. Michelas, C. Devaud, V. Sarradin, C. Franchet, C. Martinez Gomez, C. Gomez-Roca, D. Heaugwane, F. Lauzéral-Vizcaino, L. Mir-Mesnier, V. Féliu, C. Valle, G. Ferron, L. Gladieff, S. Motton, Y. Tanguy Le Gac, A. Dupret-Bories, J. Sarini, B. Vairel, C. Illac, A. Siegfried-Vergnon, E. Mery, S. Vergez, J.-P. Delord, P. Rochaix
Analysis and interpretation of data (e.g., statistical analysis, biostatistics, computational analysis): C.-C. Balança, C.-M. Scarlata, M. Michelas, C. Devaud, V. Sarradin, C. Franchet, M. Tosolini, D. Heaugwane, F. Lauzéral-Vizcaino, L. Mir-Mesnier, V. Féliu, F. Pont, J.-J. Fournié, M. Ayyoub
Writing, review, and/or revision of the manuscript: C.-C. Balança, C.-M. Scarlata, M. Michelas, V. Sarradin, C. Franchet, C. Martinez Gomez, G. Ferron, L. Gladieff, J.-P. Delord, A. Martinez, M. Ayyoub
Administrative, technical, or material support (i.e., reporting or organizing data, constructing databases): C. Martinez Gomez
Study supervision: A. Martinez, M. Ayyoub
The study was supported by the Cancer Research Institute, by the Ludwig Institute for Cancer Research, and by MSDAVENIR. The authors are sincerely grateful to patients for their participation in the study. The authors thank Prof. B. Ségui, Dr. C. Colacios, Dr. L. Martinet, and Dr. S. Valitutti, CRCT, for fruitful discussions; Dr. T. Filleron, IUCT-O, for his recommendations regarding statistical analyses; Dr. F.-X. Frenois and IUCT-O ImagIN platform for sequential slides digitalization; IUCT nurses and support staff for their help in clinical research; and Mrs. M.-H. Lalaux, CRCT, for logistic support. This work was granted access to the HPC resources of CALMIP supercomputing center under the allocation 2019-T19001 and P19043. The authors also thank the Genotoul Bioinformatics Platform for providing computing resources.
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.