We prospectively enrolled patients who were newly diagnosed with pathologically confirmed ovarian cancer between February 2018 and February 2020 at Severance Hospital (Seoul, Korea). After exclusion of patients who underwent neoadjuvant chemotherapy, our analysis included 65 patients. From each patient, we collected whole blood samples and separately collected fresh tissue samples of the primary site (ovary) and metastatic sites when possible. Metastatic site tissues samples were collected from the bladder, transverse colon, spleen, sigmoid colon, diaphragm, cul-de-sac, peritoneum, pelvis and omentum. table 1 summarizes the clinical characteristics of the enrolled patients.

From whole blood, we isolated peripheral blood mononuclear cells (PBMCs) by Ficoll-Paque density gradient centrifugation (GE Healthcare, Uppsala, Sweden). To obtain single-cell suspensions of tumor-infiltrating lymphocytes (TILs) from tumor tissues, fresh tumor tissues were extracted from primary sites or metastatic sites and were immediately cut into small pieces (2–4 mm) using dissection scissors. Next, ≤1 g of tumor tissue was transferred to a gentle MACS C-Tube (Milteny Biotec, Bergisch Gladbach, Germany) containing an enzyme mixture (200 µL enzyme H, 100 µL enzyme R and 25 µL enzyme A; Milteny Biotec, Bergisch Gladbach, Germany) premixed with 4.7 mL RPMI. These C-tubes were then placed on a gentle MACS Octo Dissociator (Milteny Biotec) to undergo mechanical and enzymatic digestion using a predefined program ‘37C_h_TDK_3’ for 61 min. After digestion, the cell suspensions were passed through 40 μm pore cell strainers, and then washed once. Next, the cells were resuspended in 10 mL media, and 20 µL of this suspension was stained with an equal volume of AOPI staining solution (Nexcelcom Bioscience, Lawrence, Massachusetts, USA). To determine the number and viability of the mononuclear cells, the stained cells were assessed using a Cellometer Auto 2000 (Nexcelcom). Finally, the cells were washed once and cryopreserved until further use.

We performed multimer staining to detect tumor-specific CD8 T cells and tumor-unrelated virus-specific CD8 T cells. To detect tumor-associated antigen (TAA)-specific CD8 T cells, we used two phycoerythrin-conjugated dextramers: one specific for the NY-ESO-1_157–165 epitope27 28 (SLLMWITQV, HLA-A*02:01), and one for the MAGE-A3_112–120 epitope29 30 (KVAELVHFL, HLA-A*02:01) (Immudex, Copenhagen, Denmark). To detect virus-specific CD8 T cells, we used two allophycocyanin-conjugated pentamers: one specific for the HCMVpp65_495–504 epitope (NLVPMVATV, HLA-A*02:01) and one for the EBV LMP-2_426–434 epitope (CLGGLLTMV, HLA-A*02:01) (Proimmune, Oxford, UK).

For multimer staining, 1–2×10⁶ cells were pretreated with 50 nmol/L of the protein kinase inhibitor dasatinib (Axon Medchem, Groningen, The Netherlands) for 30 min at 37°C to inhibit TCR internalization.31 32 Then these cells were stained with multimers for 20 min at 4°C, and washed twice with fluorescence-activated cell sorting (FACS) buffer (phosphate-buffered saline+5% fetal bovine serum+1% sodium azide). Next, the cells were stained using the LIVE/DEAD fixable dead cell stain kit (Invitrogen, Carlsbad, California, USA) for 5 min at 20°C. Cells that were not stained with multimers were first stained with the LIVE/DEAD fixable dead cell stain kit. Without washing, the LIVE/DEAD-stained cells were next stained with multiple fluorochrome-conjugated antibodies against surface markers for 20 min at 4°C, and then washed again with FACS buffer. For intracellular staining, the cells stained with LIVE/DEAD and with antibodies against surface makers were fixed and permeabilized using a forkhead box P3 staining buffer kit (eBioscience, San Diego, California, USA), and then further stained with antibodies against intracellular proteins. Flow cytometry was performed using either an LSR II instrument or an ARIA II cell sorter and FACSDiva software (BD Biosciences, San Jose, California, USA). Data were analyzed using FlowJo software (Treestar, San Carlos, California, USA).

Multimer-positive cells were clearly identified when compared with fluorescence minus one controls or irrelevant epitope (HCV NS3_1406-1415)-specific multimer-stained cells (online supplemental figure S1A,B). Next, we strictly gated multimer⁺ cells to remove signals caused by non-specific binding, and confirmed that the multimer⁺ cells were tumor-specific CD8 TILs by comparing tumor-reactive markers (PD-1^high and 4-1BB) and the exhaustion status marker Tcf-1 between multimer⁺ and multimer^– cells (online supplemental figure S1C). A multimer-stained sample was used for analysis only when ≥50 multimer⁺ cells were detected.

Online supplemental tables S1,S2 present all the antibodies and FACS panels that were used in this study.

RNA sequencing (RNA-seq) was performed using sorted 4-1BB^neg PD-1^high CD39⁺, 4-1BB^pos PD-1^high CD39⁺ CD8 TILs and PD-1⁺CD39⁻ CD8 TILs (reference population) (n=3). Cell sorting was performed in two steps: first with the magnetic-activated cell sorter (MACS, Miltenyi Biotec), and then with an ARIA II cell sorter. Total tumor-infiltrating cells were magnetically labeled using CD8 microbeads, and positively selected using the MACS cell sorter. Next, the sorted CD8 T cells were stained with antibodies against CD3, CD39, PD-1 and 4-1BB, and sorted based on surface marker expressions using an ARIA II cell sorter (online supplemental figure S2). Online supplemental table S3 presents the obtained cell counts for each population. Finally, the sorted cells were dissolved in TRIzol reagent (Invitrogen). For RNA extraction, chloroform was added and the samples were vigorously mixed by hand for 15 s and then incubated at room temperature (RT) for 2 min. Next, the incubated mixtures were centrifuged at 12 000 rpm for 15 min at 4°C. The aqueous phase was removed and mixed with 1 µL of 1:2 diluted glycogen (Roche 10901393001) and 1 µL isopropanol. These mixtures were incubated at RT for 10 min, to allow total RNA precipitation. Next, the samples were centrifuged at 12 000 rpm for 10 min at 4°C, washed with 75% ethanol, and then centrifuged again at 10 000 rpm for 5 min at 4°C. The resultant RNA pellets were dried for 10 min at RT, and then dissolved in RNase-free water. The purified RNA quality was measured using an Agilent 4200 tape station. Most samples showed RNA integrity number (RIN) values ranging from 3.6 to 9.1, with an average of 5.8.

RNA-seq libraries were constructed using the Ovation RNA-Seq System V2 kit (NuGen 7102-32). Briefly, Ovation RNA-Seq System V2 targets RNA through a DNA-universal RNA chimeric primer, and implements first and second complementary DNA (cDNA) synthesis. Subsequently, the cDNA is amplified through single primer isothermal amplification, with repeated hybridization and degradation of the single universal primer. To assess non-specific amplification, one sample was diluted 1:10 000 as a control for each experiment. The cDNA libraries were constructed using the Illumina indexed adapter attachment, and converted using the MGI library conversion kit (MGI Tech 1000004155). The MGI libraries were sequenced using 100 bp paired-end mode on an MGI DNBSEQ-G400 sequencer.

The sequencing reads were aligned to the human reference genome using STAR (V.2.7.2),33 and normalized for effective library size. Differentially expressed genes (DEGs) were analyzed using DESeq2.34 In the DESeq2 algorithm, first, size factors are estimated using the median-of-ratios method. Next, to accurately estimate the dispersion parameter for within-group variability, gene-wise dispersion estimates are generated, and are then shrunk toward the values predicted by the curve of location parameters of the distribution, to obtain final dispersion values. For significance testing, DESeq2 applies a Walt test: the shrunken estimate of the logarithmic fold-change divided by its SE. Finally, the Walt test p values are adjusted for multiple testing using the procedure of Benjamini and Hochberg.35 We selected the top DEGs based on the highest p values. To identify enriched gene ontology (GO) terms, we performed GO enrichment analysis with defined DEGs, and used a visualization tool (Gorilla tools) at http://cbl-gorilla.cs.technion.ac.il. Gene sets related to the activation and the exhaustion signatures of CD8 T cells36–39 were used for gene set enrichment analysis (GSEA) and gene set variation analysis (GSVA).40 The GEO accession number for our RNA-seq data is GSE160705.

For the analysis, we obtained gene expression profiles and clinical data of 309 patients with ovarian cancer in The Cancer Genome Atlas (TCGA) database from the GDC Data Portal (National Cancer Institute, Rockville, Maryland, USA). For survival analysis, we calculated the gene set enrichment score via GSVA using the 177 DEGs (with a p value of <0.05 and log₂fold-change >1) from our RNA-seq data. The cut-off value was determined using the maximal χ² method.

Statistical analyses were performed using Prism software V.8.4.3 (GraphPad, La Jolla, California, USA). We applied the Mann-Whitney U test and Wilcoxon matched-pairs signed-rank test for unpaired and paired non-parametric comparisons, respectively. The non-parametric Spearman’s rank correlation test was used to assess correlations between two parameters. A p value of <0.05 was considered to indicate statistical significance.

We first analyzed the immune checkpoint receptor expression on CD8 TILs compared with peripheral blood CD8 T cells—examining PD-1, CTLA-4 and T cell immunoglobulin and mucin domain-containing protein 3 (Tim-3). The frequencies of PD-1⁺, CTLA-4⁺ and Tim-3⁺ cells were significantly higher among CD8 TILs compared with peripheral blood CD8 T cells (figure 1A). Among these immune checkpoint receptors, CD8 TILs exhibited predominantly expression of PD-1 in terms of the percentage of positive cells. Recent publications suggest that tumor-unrelated bystander CD8 T cells are abundant in tumor microenvironments.41 42 Thus, we further examined tumor-specific CD8 TILs using two HLA-A*0201-restricted multimers specific to NY-ESO-1_157–165 and MAGE-A3_112-120, and examined tumor-unrelated virus-specific CD8 T cells using multimers specific to HCMVpp65_495–504 and EBV LMP-2_426–434. Among the 26 patients having HLA-A*0201, we detected 17 TAA-specific CD8 T-cell populations in 10 patients (38.5%), and 16 virus-specific CD8 T-cell populations in 12 patients (46.2%) (online supplemental table S4).

Next, we examined the expressions of PD-1, CD39, CD103 and CD69 as known surrogate markers for tumor-specific CD8 T cells42–44 (figure 1B and C). In contrast to prior findings, ovarian cancers exhibited no significant difference in PD-1 or CD69 expression between tumor-specific CD8 TILs and tumor-unrelated virus-specific CD8 TILs (ie, CMV-specific or EBV-specific) (figure 1C). However, the percentages of CD39⁺ cells and CD103⁺ cells were significantly higher among tumor-specific multimer⁺ CD8 TILs compared with among virus-specific CD8 TILs (figure 1C). Notably, CD39 expression was barely detected in virus-specific CD8 TILs (figure 1C), which was further confirmed by paired analysis directly comparing the percentage of CD39⁺ cells between tumor-specific CD8 TILs and virus-specific CD8 TILs (figure 1D). These results suggested that, in ovarian cancers, CD39 is predominantly expressed on tumor-specific CD8 TILs, and may thus be useful as a surrogate marker to exclude tumor-unrelated virus-specific bystander CD8 T cells in immune analysis using tumor tissues. In addition, there was no significant difference in the frequency of CD39⁺ cells among CD8 TILs depending on the ovarian cancer stages, tumor grades or tumor histology (figure 1E).

Based on PD-1 expression levels, we subdivided CD39⁺ CD8 TILs into three distinct populations: PD-1-negative (PD-1^neg), PD-1-intermediate (PD-1^int) and PD-1-high (PD-1^high) subpopulations (figure 2A). The frequency of PD-1^high cells among CD39⁺ CD8 TILs did not significantly differ according to ovarian cancer stages, tumor grades or tumor histology (online supplemental figure S3). We then examined differences in the characteristics of these CD39⁺ tumor-infiltrating CD8 T cell subpopulations distinguished by differential PD-1 expression in ovarian cancers. The frequencies of CTLA-4⁺ and Tim-3⁺ cells were significantly higher in PD-1^high subpopulations than in PD-1^int or PD-1^neg subpopulations (figure 2B). We also examined the expression of transcription factors—such as T-bet, Eomes and Tcf-1, which contribute to the heterogeneous status of exhausted CD8 T cells.45–47 We found that PD-1^high subpopulations exhibited the highest proportion of Eomes high expressing and T-bet low expressing (Eomes^high T-bet^low) cells, that is, a subset of terminally exhausted that is associated with poor reinvigoration on ICI treatment45 48 (figure 2C). On the other hand, the frequency of Tcf-1⁺ cells, which represent progenitor-like exhausted T cells associated with T-cell reinvigoration potential by ICIs,46 47 was lowest in the PD-1^high subpopulation and highest in the PD-1^neg subpopulation (figure 2D).

We next investigated parameters of tumor reactivity (ie, CD103) and T-cell activation (ie, CD38 and HLA-DR). The subpopulation of PD-1^high CD39⁺ CD8 TILs exhibited the highest percentages of CD103⁺ cells representing tumor reactivity,43 and of CD38⁺ HLA-DR⁺ cells representing activation (figure 2E and F). Additionally, the frequency of CD38⁺ HLA-DR⁺ CD39⁺ CD8 TILs was inversely correlated with the frequency of Tcf-1⁺ CD39⁺ CD8 TILs (figure 2G). These findings suggested that PD-1^high CD39⁺ CD8 TILs exhibit features of tumor-reactive but terminally exhausted T cells, indicating active antitumor engagement and resultant exhausted phenotypes.

The tumor-reactive but terminally exhausted features of PD-1^high CD39⁺ CD8 TILs suggested the need for additional therapeutic strategies in combination with ICI. One promising approach involves targeting co-stimulatory receptors.49–51 However, little is known about the expression patterns of co-stimulatory receptors on CD8 TILs, or about the immunological implications of co-stimulatory receptor expression in ovarian cancers. Therefore, we next examined the expression levels of various members of the tumor necrosis factor receptor superfamily (TNFRSF) on CD39⁺ CD8 TILs from patients with ovarian cancer, including 4-1BB (TNFRSF9, CD137), CD30 (TNFRSF8), CD40 ligand (CD40L), GITR (TNFRSF18) and OX40 (TNFRSF4, CD134). Notably, compared with other TNFRSF members, 4-1BB was more prominently expressed on CD8⁺ TILs (figure 3A). Additionally, 4-1BB was rarely expressed on PD-1⁺CD39⁻ CD8 TILs (online supplemental figure S4).

Comparing the 4-1BB expression levels according to differential PD-1 expression revealed that 4-1BB was predominantly expressed on PD-1^high CD39⁺ CD8 TILs (figure 3B). Indeed, terminally exhausted (PD-1^high CD39⁺) CD8 TILs exhibited the highest percentage of 4-1BB^pos cells compared with the percentages of other TNFRSF-expressing cells (figure 3C).

To characterize the 4-1BB^pos terminally exhausted CD8 TILs, we further examined markers of T-cell activation (CD38 and HLA-DR), proliferation (Ki-67) and exhaustion status (Tcf-1, Eomes and T-bet). Among PD-1^high CD39⁺ CD8 TILs, compared with 4-1BB^neg cells, the 4-1BB^pos cells displayed immunophenotypes indicating higher degrees of T-cell activation and proliferation, and less exhaustion (figure 3D).

Next, to further investigate the distinct features of 4-1BB^pos cells among highly exhausted PD-1^high CD39⁺ CD8 TILs in ovarian cancers, we performed RNA-seq using 4-1BB^pos and 4-1BB^neg cells sorted from PD-1^high CD39⁺ CD8 TILs. We also analyzed PD-1⁺CD39⁻ CD8 TILs as a reference population (online supplemental figure S2). Compared with this reference population, the transcriptome from PD-1^highCD39⁺ CD8 TILs was enriched for the genetic signature of T-cell exhaustion36 (figure 3E). Intriguingly, compared with PD-1⁺ CD39⁻ CD8 TILs, examination of DEGs in PD-1^high CD39⁺ CD8 TILs revealed that 4-1BB^pos and 4-1BB^neg cells shared 485 DEGs in common, while an additional 683 genes were upregulated or downregulated in only one subpopulation (120 genes in 4-1BB^neg and 563 in 4-1BB^pos) (figure 3F; online supplemental table S5). TCF7 was differentially expressed between PD-1^high CD39⁺ and PD-1⁺ CD39⁻ CD8 TILs (online supplemental figure S5A and online supplemental table S5). This is because the PD-1⁺ CD39⁻ CD8 TILs consist of mostly PD-1^int-expressing cells (78.9%, online supplemental figure S5B), and this finding is consistent with the flow cytometry results (figure 2D).

Direct comparison of the 4-1BB^pos and 4-1BB^neg subpopulations among PD-1^high CD39⁺ CD8 TILs yielded the identification of 520 differentially expressed genes, of which 369 genes were significantly upregulated and 151 genes were downregulated in 4-1BB^pos cells compared with 4-1BB^neg cells (figure 3G and H; online supplemental table S6). DEGs between 4-1BB^pos and 4-1BB^neg PD-1^high CD39⁺ CD8 TILs were subjected to GO term analysis, revealing that 4-1BB^pos cells showed enrichment of genes related to ‘cell cycle process’, ‘cell division’, ‘DNA replication’, ‘heterocycle metabolic process’ and ‘cellular response to stimulus’, whereas 4-1BB^neg cells showed enrichment of genes related to ‘regulation of cell death’, ‘regulation of apoptotic process’, ‘negative regulation of response to stimulus’, ‘negative regulation of cellular metabolic process’ and ‘negative regulation of immune effector process’ (figure 3I; online supplemental table S7). GSEA confirmed that, compared with 4-1BB^neg cells, 4-1BB^pos cells exhibited significant enrichment of the ‘activated CD8 TILs gene signature’38 (figure 3J). Additionally, GSVA analysis with the activation and exhaustion signatures of CD8 T cells revealed that 4-1BB^pos cells were highly enriched for gene signatures related to ‘activated CD8 TILs gene signature’38 and ‘effector CD8 T cells gene signature’,37 whereas 4-1BB^neg cells were highly enriched for ‘exhausted CD8 TILs gene signature’38 and ‘dysfunction module’39 (figure 3K). The 177 DEGs that were upregulated in 4-1BB^pos PD-1^high CD39⁺ CD8 TILs compared with in the 4-1BB^neg counterpart (p value of <0.05 and log₂fold-change >1) were defined as the ‘4-1BB signature’, and tumors with high 4-1BB signature were significantly correlated with better survival outcome in patients with ovarian cancer in TCGA cohort (figure 3L).

Overall, our results suggested that the 4-1BB^pos subpopulation displays features of distinctive T-cell activation status, and may retain higher proliferative and reinvigoration potential among highly exhausted tumor-specific CD8 TILs.

Considering that most patients with ovarian cancer receiving immunotherapy have metastatic sites, it is important to understand the distinct characteristics of exhausted CD8 T cells from the metastatic site compared with the primary site. We obtained TILs from various metastatic sites—including the bladder, transverse colon, spleen, sigmoid colon, diaphragm, cul-de-sac, peritoneum, pelvis and omentum (online supplemental table S8 presents detailed information about the metastatic sites). We confirmed that CD39 is predominantly expressed on tumor-specific CD8 TILs also in metastatic site, wherease CD39 expression was barely detected in tumor-unrelated virus-specific CD8 TILs (figure 4A, online supplemental figure S6). To investigate the immunological characteristics of CD39⁺ CD8 TILs in the metastatic sites, we first compared the expressions of immune checkpoint receptors such as PD-1, CTLA-4 and Tim-3 in CD39⁺ CD8 TILs between primary sites and the metastatic sites, and found that the frequencies of PD-1⁺ and CTLA-4⁺ cells in CD39⁺ CD8 TILs were significantly lower in metastatic sites than in primary sites (figure 4B). However, there was no significant differences in the frequency of CD103+, Eomes^high T-bet^low cells and Tcf-1⁺ cells among CD39⁺ CD8 TILs (online supplemental figure S7).

For the detailed characterization of CD39⁺ CD8 TILs from metastatic sites, CD39⁺ CD8 TILs were subdivided by PD-1 expression level (figure 4C). Compared with CD39⁺ CD8 TILs from primary sites, those from metastatic sites exhibited a slightly lower percentage of PD-1^high cells (41.7% vs 54.7%), and slightly higher percentage of PD-1^int cells (40.2% vs 30.4%) (figure 4D, online supplemental figure S8). The frequency of PD-1^neg cells did not significantly differ between CD39⁺ CD8 TILs from primary versus metastatic sites. These results indicated that although CD39⁺ CD8 TILs from primary and metastatic sites showed similar frequencies of PD-1⁺ cells, the PD-1 expression level was slightly higher at primary versus metastatic sites. As found at primary sites, higher PD-1 expression was associated with higher frequencies of CD103⁺, CD38⁺ HLA-DR⁺, Eomes^high T-bet^low and 4-1BB⁺ cells, and with a lower frequency of Tcf-1⁺ cells (figure 4E).

We then compared features of T-cell exhaustion and activation between primary and metastatic sites. As shown in figure 4F, PD-1^high CD39⁺ CD8 TILs from metastatic sites expressed similar levels of CD103, representing tumor reactivity; exhibited a similar frequency of CD38⁺ HLA-DR⁺ cells; similar frequencies of Eomes^high T-bet^low cells and Tcf-1⁺ cells and similar frequencies of 4-1BB⁺ cells (figure 4F). These findings suggest that PD-1^high CD39⁺ CD8 TILs showed similar characteristics in terms of T-cell exhaustion and activation between the primary and metastatic sites. As found at primary sites, among PD-1^high CD39⁺ CD8 TILs from metastatic sites, the 4-1BB^pos cells displayed immunophenotypes indicating higher degrees of T-cell activation and proliferation, and less exhaustion, compared with 4-1BB^neg cells (figure 4G).

We further investigated CD8 TILs from the original and metastatic sites to determine which subpopulation was most enriched with 4-1BB⁺ cells. Combined t-distributed stochastic neighbor embedding (t-SNE) analysis of paired CD8 TILs from primary and metastatic sites (22 samples) revealed 16 different clusters, which were further grouped into 10 distinctive subpopulations based on differential expression of PD-1, CD39 and CD103 (figure 5A–B). Among these 10 distinct subpopulations, we found that the PD-1^high CD39⁺ CD103⁺ subpopulation exhibited the highest proportion of 4-1BB⁺ cells (figure 5C), as well as the highest proportions of CTLA-4⁺ cells and Tim-3⁺ cells (figure 5D). We further confirmed that the highest proportion of 4-1BB⁺ cells was observed in PD-1^high CD39⁺ CD103⁺ subpopulation from both primary sites and metastatic sites (figure 5E). Analysis of PBMCs revealed that most CD8 T cells were PD-1^int CD39^– CD103^– or PD-1^neg CD39^– CD103^– (online supplemental figure S9).

Although the PD-1^high subpopulation of CD39⁺ CD8 TILs showed the features of terminally exhausted T cells, a significant portion expressed 4-1BB—suggesting that they may retain higher proliferative and reinvigoration potential. Therefore, we investigated whether anti-4-1BB agonistic antibodies could enhance the function of exhausted CD8 TILs in ovarian cancers. The ex vivo T-cell functional restoration assay revealed that 4-1BB co-stimulation with agonistic antibodies significantly enhanced CD8 TIL proliferation indicated by the SI (figure 6A), as CD8 TIL production of IFN-γ and TNF-α (figure 6B). Interestingly, anti-4-1BB treatment also significantly enhanced the function of CD8 TILs from metastatic sites (figure 6C), suggesting that anti-4-1BB treatment may be effective at both primary and metastatic sites.

Recent clinical trials of anti-PD-1 therapy in patients with ovarian cancer suggest that the therapeutic efficacy must be improved.14 Therefore, we lastly examined whether 4-1BB co-stimulation with agonistic antibodies could further enhance anti-PD-1-mediated T-cell reinvigoration. Compared with single PD-1 blockade, the combination of PD-1 blockade plus 4-1BB co-stimulation with agonistic antibodies further enhanced the proliferative response of CD8 TILs (figure 7A), as well as the capacity of CD8 TILs to produce IFN-γ and TNF-α (figure 7B). Notably, 4-1BB co-stimulation further enhanced the anti-PD-1-mediated reinvigoration of exhausted CD8 TILs from both metastatic sites and primary sites (figure 7B and C).