Movement of exhausted CD8+ T cells linked to immune structures involved in response to PD-1 therapy in esophageal cancer

We hypothesized that the phenotypes of CD8⁺ T cells might differ within the cohort of CD39⁺PD-1⁺ Tex cells. From 11 independent ROIs, each containing more than 500 CD8⁺ T cells, we applied unsupervised clustering to CD8⁺ T cells using FlowSOM (Fig. S2A) and visualized these clusters in UMAP plots (Figs. S2B, S2C). We identified five CD39^hiPD-1^hi Tex clusters (C1–5) characterized by high TOX and PD-1 expression (Fig. S2A). When mapping these clusters to the representative ROIs of tumors lacking TLSs (Fig. S2D) and tumors containing TLSs (Fig. S2E), clusters C1 and C2, which exhibited high levels of TCF-1 expression and correspond to Tpex cells due to their exhaustion profile, were primarily localized to the stroma and TLSs. The localization characteristics of these TCF-1⁺ populations to TLSs were consistent with previously reported findings^29,30. In contrast, clusters C3–5, primarily localized within the tumor parenchyma, corresponded to differentiated Tex (dTex) cells lacking TCF-1 expression (Fig. S2D, S2E). The presence of CD39⁺ Tpex (TCF-1⁺CD39⁺PD-1⁺CD8⁺ T) cells, previously reported in several studies^26,33–35, was also confirmed through co-staining in the tumor ROIs (Figs. 3a, S11B). To characterize this unique population within the TME, we categorized CD39⁺PD-1⁺ Tex cells from 27 cases as CD39⁺ Tpex and CD39⁺ dTex (TCF-1⁻CD39⁺PD-1⁺CD8⁺ T) cells. Both subsets exhibited high exhaustion marker levels (Fig. S1A, S1B), suggesting that they are potentially tumor-reactive T cells under continuous antigen stimulation in the TME^14–17.

To obtain further substantiation of their existence, we analyzed a dataset³³ from a cohort of head and neck squamous cell carcinoma (HNSCC) patients treated with atezolizumab (anti-PD-L1), where tumor-infiltrating lymphocytes (TILs) were examined using mass flow cytometry (MC) (Fig. S3A). We identified populations of CD39^hiPD-1^hi clusters (Clusters; C1, C2, C3, C4) through unsupervised clustering of CD8⁺ T cells with FlowSOM (Fig. S3B) and visualized them using UMAP plots (Fig. S3C). Notably, clusters C3 and C4, which correspond to CD39⁺ Tpex cells, were characterized not only by their expression of TCF-1, but also their relatively higher expression of CCR7, CD127, and TIGIT (Figs. S3B, S3D), indicating that they are phenotypically distinct Tpex cells, unlike CD39⁺ dTex cells^24,26,31,33. These results collectively indicate that CD39⁺PD-1⁺CD8⁺ T cells constitute a heterogeneous group that includes not only dTex cells but also Tpex cells.

Previous studies have highlighted the correlation between intratumoral TCF-1⁺CD8⁺ T cells and clinical benefit from ICB therapy in melanoma^21,36 and breast cancer³⁷. Given these findings, we explored this correlation in ESCC by analyzing the association between clinical outcomes and the abundance of CD39⁺ Tpex cells. The relative abundance of these cells among CD39⁺PD-1⁺ Tex cells in the tumor and stroma area (excluding TLS areas) showed no clear impact on clinical outcomes (Fig. S1F), which is consistent with previous melanoma findings²¹. However, a significant positive correlation was observed between the densities of CD39⁺ Tpex cells and CD39⁺PD-1⁺ Tex cells predictive of ICB therapy benefit (n = 27 patients, r = 0.795, P < 0.001; Fig. 3b). A similar correlation was also observed using the ICB cohort dataset³³ (n = 8 patients, r = 0.831, P = 0.0106; Fig. S3E). Accordingly, the absolute density of CD39⁺ Tpex cells in the tumor and stroma area was significantly higher in ICB therapy responders than non-responders (n = 22 patients, P = 0.0082; Fig. 3c). These results suggest that CD39⁺ Tpex cells, as potential precursors of CD39⁺ dTex cells, constitute a significant component of the immune responses against the tumor.

To examine hierarchical patterns, we focused on phenotypic and functional differences between unsupervised Tex clusters (C1 and C4) in blood (Fig. 7d, S4A). C4 expressed higher TCF-1, TOX, PD-1 and CTLA4 levels (Fig. S4A), representing proliferative CD39⁺ Tpex cells. Manual detection by MC confirmed both proliferative CD39⁺ Tpex and CD39⁺ dTex populations with a phenotypically exhausted profile expressing high levels of TOX, Eomes, PD-1, CD39, and CTLA4 (Fig. 7g, h). CD39⁺ Tpex cells in blood expressed relatively higher levels of CCR7 and CD127 (Fig. 7h) and lower levels of TIM-3 (Fig. S6A), consistent with previous reports^24,26,33. Despite the lower frequency of cluster C4 compared to the manually gated groups from both FC and MC, a strong correlation was observed between the frequency of C4 and each manually gated group (Fig. S4D). Notably, a significant increase in C4 was observed in one particular case (Patient ID KU37), with similar results also reported in manually gated groups (Fig. 7e, S4D). Additionally, a greater proportion of CD39⁺ dTex cells than CD39⁺ Tpex cells produced IFN-γ after TCR-independent in vitro stimulation (Fig. S6C), confirming the greater cytotoxic activity of these cells^21,33. In contrast, CD39⁺Ki67⁺ Tex cells circulating in the blood lacked CD69 expression (Fig. S6B) and resembled Tex progenitor2 or intermediately differentiated Tex cells of the four previously reported developmental stages^26,33. These observations suggest that CD39⁺Ki67⁺ Tex populations in the blood consist of proliferative CD39⁺ Tpex cells, exhibiting phenotypic and functional features distinct from those of CD39⁺ dTex cells.

We also investigated the association between CD39⁺Ki67⁺ Tex cells in the blood and the clinical response to ICB therapy. Following ICB therapy, the frequency of CD39⁺ dTex cells increased significantly, regardless of clinical response (Fig. 8c). However, the frequency of CD39⁺ Tpex cells showed a significant increase only in responder group (n = 10 patients, P = 0.0039; Fig. 8c). Moreover, the abundance of proliferative CD39⁺ Tpex cells at 2 weeks after ICB therapy was significantly associated with clinical response (P = 0.0065; Fig. 8d). The abundance (> 10%) of CD39⁺ Tpex cells after ICB therapy was also significantly associated with longer progression-free survival (PFS) (n = 28 patients, P = 0.0012; Fig. 8e). These findings indicate that the increase in proliferating CD39⁺ Tpex cells in the blood following ICB therapy is correlated with clinical benefit and reflects the robust immune response driving their expansion and facilitating their subsequent differentiation into dTex cells.

Circulating tumor-specific Tpex cells are reportedly recruited from lymph nodes (LNs) into the TME^33,43,44. Consequently, we hypothesized that CD8⁺ T cells with a CD39⁺PD-1⁺ Tex phenotype are present within SLOs. Therefore, we performed unsupervised clustering analysis of CD8⁺ T cells manually extracted from the ROIs of individually resected SLOs from three patients (Fig. S7B) and visualized these clusters using UMAP analysis (Fig. S7C). We identified clusters of CD39^hiPD-1^hi Tex cells (C1–5) with high levels of TOX expression. The presence of this CD39⁺PD-1⁺ Tex subset was also confirmed in a representative SLO image (Fig. 8b). Although only a single case of HNSCC in the ICB cohort had non-metastatic LN samples available post-ICB therapy for MC analysis, we found a remarkably abundant population of CD39⁺PD-1⁺ and CD39⁺ Tpex cells within the CD8⁺ T cells (Fig. S3I). Among these clusters, C5 exhibited relatively low TCF-1 expression, possibly reflecting CD39⁺ dTex cells (Fig. S7B). However, most importantly, C1 and C2 which exhibited high levels of Ki67 expression also expressed high levels of TCF-1 (Fig. S7B), corresponding to CD39⁺ Tpex cells. This characteristic of high Ki67 expression was similar to the findings in the dataset of resected LNs from the HNSCC that did not receive ICB treatment, where CD39⁺ Tpex cells had higher Ki67 expression rates than CD39⁺ dTex cells (Fig. S3J).

To elucidate the spatial characteristics of these clusters within LNs, we mapped each CD8⁺ T cell cluster as shown in Figure. S7A and S7D, and analyzed their proximity to major immune cells manually extracted from each ROI. We found that all CD8⁺ T cell clusters were generally in close proximity to CD4⁺ T cell subsets (Fig. S7E), compared to CD20⁺ B cells and antigen-presenting cells (CD11c⁺HLA-DR⁺). Notably in the representative SLO, CD39⁺ Tpex clusters (C1-4) were localized in significantly closer proximity to PD-1⁺ Treg cells compared to other non-exhausted CD8⁺ T cells (Fig. S7F). This proximity characteristic was consistently observed in all three cases (Fig. S7G) and suggests possible immunosuppressive conditions surrounding CD39⁺PD-1⁺ Tex cells in LNs^33,43,44.

Recently, it was reported that ICB induces the conversion of most Tpex cells in SLOs into more-differentiated Tex cells that then transit into the blood³³. Our findings revealed that the proportion of circulating CD39⁺ Tpex cells within CD39⁺Ki67⁺ Tex cells decreased 2 weeks after ICB therapy compared to before treatment (Figs. S6D, S6E) and was further reduced by the time of disease progression (PD) (Fig. S6E). Furthermore, the increased frequency of circulating CD39⁺ dTex cells (Figs. 7e, 8c) supports the potential scenario where ICB triggers a conversion of highly proliferative CD39⁺ Tpex cells within the SLOs.

ICB therapy preferentially increases the frequency of circulating CD39⁺Ki67⁺ Tex cells, suggesting that the CD39⁺PD-1⁺ Tex population in the TME after ICB consists primarily of reinvigorated Tex cells recruited from the periphery^12,20. We therefore analyzed ROIs from biopsied tumors of three cases at PD during ICB therapy (Fig. 8f). Although the sample size is small (n = 3 independent tumor-ROIs), the density of T cells, including CD39⁺ Tpex and CD39⁺ dTex cells, within the tumor area did not increase markedly (Fig. 8g), similar to the lack of increase in their frequency in PBMCs at PD (Fig. S6F) and consistent with previously reported findings in a study of MSI-high rectal cancer⁴⁵. However, CD39⁺ Tpex cells exhibited consistently high Ki67 expression in the second biopsy, a pattern also observed in our IMC results prior to ICB therapy (Fig. S3B). This result is further supported by the characteristic high Ki67 expression of intratumoral CD39⁺ Tpex cells in the HNSCC dataset, which includes 6 out of 8 samples post-ICB therapy (Figs. S3B, S3C, S3D). Manual gating also confirmed that CD39⁺ Tpex cells exhibited high Ki67 expression comparable to CD39⁺ dTex cells (Figs. S3F, S3G). Furthermore, TCF-1⁺Ki67⁺CD8⁺ TILs, a key predictor of ICB efficacy in breast cancer³⁷, showed a significantly higher CD39⁺PD-1⁺ positivity rate compared to TCF-1⁺Ki67^–CD8⁺ TILs in the HNSCC dataset (Fig. S3H), indicating that many highly proliferative TCF-1⁺CD8⁺ T cells are reflective of CD39⁺ Tpex cells. This characteristic aligns with the properties of CD39⁺ Tpex cells within blood (Fig. 7f) and the SLOs (Fig. S7B). Collectively, the results support the hypothesis that CD39⁺ Tpex cells in the tumor, blood, and lymphoid tissues, characterized by high proliferative activity, are closely associated with ICB-mediated anti-tumor responses and play a critical role in therapeutic efficacy.

From March 2020 to July 2021, a total of 31 patients with unresectable or advanced ESCC (median age 72 years, range 40–85) were prospectively enrolled for nivolumab treatment (240 mg/patient every 2 weeks until disease progression, death, or unacceptable toxicity) (Fig. 1a). Inclusion criteria were as follows: histologically confirmed ESCC, a baseline Eastern Cooperative Oncology Group performance status of 0–1, resistance or intolerance to fluoropyrimidine-based and platinum-based chemotherapy, age over 20 years, and no prior immunotherapy (Fig. 1a, b). One patient, without the necessary chemotherapy history, was excluded from the analysis. Patient characteristics are outlined in Table S1. Written informed consent was obtained from each patient for participation in this study, and blood samples from healthy donors were collected from volunteers who also provided written informed consent. There was no additional compensation for patients. Ethical approval was centrally granted by the Ethics Committee of Kyushu University Hospital (Approval No. 2023-17) and by the respective Ethics Committees of the following institutions: Japan Community Healthcare Organization Kyushu Hospital, NHO Kyushu Medical Center, Hamanomachi Hospital, Fukuoka Wajiro Hospital, and National Kyushu Cancer Center. This study was conducted in accordance with the Declaration of Helsinki. Sex or gender were not considered in the study design as it was not relevant to our research objectives.

Blood was collected in tubes containing acid citrate dextrose solution before initial and second nivolumab administrations and at the point of PD. PBMCs were obtained pre-treatment from all 31 patients, at 2 weeks post-ICB from 28 patients, and at PD from 19 patients (Fig. 1b). Post-ICB samples of 30 eligible patients could not be obtained from 2 patients due to patient refusal and COVID-19 restrictions. PBMCs were isolated via Ficoll gradient centrifugation and stored at −80 °C. FFPE tissue blocks were obtained from archival pathology specimens. Pre-treatment tumor samples were obtained from 27 patients, post-treatment tumor samples from 3 patients, and pre-treatment regional LNs from 3 of the 30 eligible patients (Fig. 1b). Pre-treatment FFPE samples for 3 patients were unavailable because two were held by other hospitals and one was non-tumorous. Each sample’s viable tumor area was identified and mapped to its H&E slide by two pathologists (D.K. and H.Y.).

Tumor response to anti–PD-1 therapy was assessed based on best response using RECIST, version 1.1, including cases without measurable lesions. Clinical benefit from ICB therapy was defined as either Non–complete response (CR)/NonPD or SD with PFS ≥ 6 months (Fig. 1c). As per prior reports^32,49, the clinical response was modified as follows: 1) patients exhibiting PR or CR were deemed clinical responders; 2) patients exhibiting NonCR/NonPD or SD with PFS ≧ 6 months were deemed responders; and 3) all other patients were classified as non-responders. Cases not evaluable (NE) included 3 patients changing treatment, including radiotherapy prior to the initial treatment evaluation, and 2 patients dying before treatment evaluation could be conducted. NE cases were excluded from the analysis based on treatment efficacy analysis.

Antibodies were sourced from Standard Biotools and validated on tonsil and ESCC tissues using IMC. Antibodies were validated on PBMCs from healthy donors and ESCC patients using MC to ensure strong staining intensity and optimal signal-to-noise ratio. Unlabeled antibodies underwent metal conjugation using a Maxpar X8 Multimetal Labeling Kit (Standard Biotools). Titration and specificity were visually confirmed for new conjugates using tonsil images for IMC and PBMCs for MC. CD45-152Sm (CD45-2B11) and CD28-173Yb (EPR22076↗) in tumor-ROIs exhibited an extremely low signal-to-noise ratio and was therefore not used for analysis. Further antibody and critical reagents are described in Tables S2, S3, S4 and S5.

Serial sections (4 μm) from ESCC patient FFPE blocks were prepared and stained per the Standard Biotools IMC protocol. The tumor samples were de-waxed by baking at 60 °C for 2 h, followed by deparaffinization in fresh xylene (twice, 10 min each) and rehydration through a graded alcohol series (5 min each at 100%, 95%, 80%, 70%). Antigen retrieval was performed using a decloaking chamber with Tris-EDTA buffer (Antigen Unmasking Solution, pH 9.0, Vector Laboratories) at 110 °C for 20 min, followed by cooling to 70 °C at room temperature (RT). Slides were blocked with 3% BSA in Maxpar PBS at RT for 60 min. Concurrently, the antibody cocktail, prepared in 0.5% BSA buffer, was applied post-blocking and incubated overnight at 4 °C. Following incubation, slides were washed and treated with Cell-ID Intercalator-Ir (Standard Biotools) for DNA staining at RT for 30 min. Finally, slides were rinsed and air-dried for at least 20 min before IMC acquisition.

IHC staining was performed on serial tumor sections using the following antibodies: CD3 (clone F7.2.38, Agilent Technologies), CD20 (clone L26, Agilent Technologies), and anti-PNA (MECA79R, Novus). The slides were first deparaffinized and rehydrated, followed by antigen retrieval using EDTA buffer (Antigen Retrieval Buffer, EDTA buffer, pH 8.0, Abcam) through a 20-min microwave treatment. Endogenous peroxidase activity was blocked by incubating the slides in 3% hydrogen peroxidase for 5 min. Protein blocking was performed using Protein Block Serum-Free (Agilent Technologies). After applying the primary antibody, a DAKO Envision Detection System (Agilent Technologies) was used for primary antibody detection.

Whole immunohistochemical-stained slides from all cases were scanned to identify areas rich in CD3⁺ cells and CD3⁺/CD20⁺ TLSs within the tumor. For tumors lacking TLS, areas with a high density of CD3⁺ cells were randomly mapped from consecutive IHC sections stained for CD3. One ROI per mapped area was then acquired from IMC-stained samples (Fig. S8A). Before the acquisition of ROIs within TLSs, TLSs were initially screened in tumor-adjacent or -inherent areas using consecutive IHC sections stained for CD3 and CD20. Subsequently, randomly selected TLSs were then mapped to their corresponding locations on IMC-stained samples, ensuring that at least one tumor site per TLS was included. The accuracy of acquiring designated ROIs was experimentally confirmed twice using the representative sample (patient ID KU37) (Fig. 2b). Upon verification, for each sample in which TLSs were detectable, ROIs corresponding to TLSs and associated tumor areas were captured (Fig. S8C). Optionally, additional tumor ROI without TLSs, previously mapped based on IHC-stained serial sections, were also acquired. Given the predefined conditions described above, the area for ROI acquisition was predetermined, and ROIs were acquired within a range of up to 2.25 mm² (0.39–2.25 mm²).

IMC images were acquired using a Hyperion Imaging System (Standard Biotools) linked to a Helios mass cytometer (Standard Biotools) (Fig.). All images were collected according to the manufacturer’s guidelines. Following daily calibration, each tissue slide was subjected to 200-Hz laser ablation. The resulting IMC data were saved as MCD files. The integrity and quality of staining across all samples in each channel were manually verified by reviewing the visualized MCD files in MCD Viewer (ver. 1.0.560.6, Standard Biotools). Figureandshow the expression of all markers in SLOs or tumors. S10A S9A S9B

Raw data from 27 eligible patients were converted to 32-bit TIFF via MCD Viewer, followed by processing using two different imaging methods (Fig. S10B, S10C). Briefly, while both methods were applied to tumor samples, “IMC image processing B”, using HALO image analysis software (ver. 3.5.3577.265, Indica Labs), was predominantly employed for these samples, as per the manufacturer’s protocol. “IMC image processing A”, based on a previously reported algorithm⁵⁰, was utilized to generate spatial information for CD8⁺ T cell clustering within tumors. For SLOs, “IMC image processing A” was primarily utilized due to its superior small lymphocyte nuclei identification capability compared to image processing B.

We noted the presence of hot pixels and shot noise in the IMC images obtained for several markers, which could potentially affect downstream analysis. To clearly visualize protein expression, particularly in channels with low signal-to-noise ratios, set precise positive thresholds, and improve cell clustering, we utilized the IMC-Denoise software package⁵¹. This software was used to analyze markers including CD3, CD4, CD8, CD20, PD-1, CD39, TOX, Ki67, CD45RA, CD45RO, HLA-DR, CD11c, FoxP3, TCF-1, LAG-3, OX-40, TIM-3, CD69, and CD25.

IMC-Denoise employs two main steps: a differential intensity map-based restoration (DIMR) for hot pixel removal and a self-supervised deep learning algorithm for shot noise image filtering (DeepSNiF) for shot noise suppression. DeepSNiF requires model training prior to implementation. We trained a separate model for each marker, resulting in the creation of a separate dataset for each marker from our IMC images after DIMR hot pixel removal. Each dataset comprised ~30,000 64 × 64 patches after data augmentation. Subsequently, all models were trained on a single NVIDIA RTX 3000 GPU over 200 epochs, using the default parameter settings of the software. After training, the DIMR-processed IMC images were denoised by DeepSNiF using their corresponding trained models. The denoised IMC images were then evaluated by a pathologist (T.Iw.) (Fig.). The quality of these images was confirmed to be comparable in quality to immunofluorescence staining for CD39, PD-1, and CD8 (Figs.B,). S11A S11 S11C

Staining and data acquisition were carried out in accordance with the manufacturer’s guidelines. Notably, Cd-CD45 antibody barcoding mixtures were created according to the barcode key (Fig. 7c) and then added to the 12 corresponding sample tubes, each containing 2 × 10⁶ cells. In addition, PD-1 expression on CD8⁺ T cells in peripheral blood was analyzed using 175Lu anti-human IgG4 (clone HP6025, Southern Biotec) following saturation with nivolumab (human IgG4, Sellek Chemicals). Figure S4F presents a comparison of the direct and indirect methods for detecting PD-1 expression by MC.

A solution was prepared using RPMI 1640 medium supplemented with 10% FBS, 5 mM MgCl₂, and 15 units/ml of DNAase I and incubated at 37 °C in a 5% CO₂ environment. Cryopreserved PBMCs were thawed in a 37 °C water bath, transferred to a 5-ml tube containing the aforementioned solution, and centrifuged at 300 g for 10 min. After removing the supernatant by aspiration, an additional 3–4 ml of the same solution was added, followed by incubation at 37 °C for 30 min. The sample was then centrifuged at 300 g and the supernatant was removed by aspiration. The pellet was washed twice using Maxpar Cell Stain Buffer (CSB) (Standard Biotools) supplemented with 1 mM EDTA. Fc-blocker was added, followed by a 10-min incubation at RT, and the cells were then washed with CSB plus 1 mM EDTA.

After the supernatant was removed by aspiration, the cells were incubated for 30 min at RT with a solution containing Nivolumab at 25 μg/ml and 103Rh DNA intercalator (1:500) (Standard Biotools). After washing, the cells were incubated at RT for 30 min in CSB with 175Lu anti-IgG4. After two more washes, the samples were incubated for 30 min at RT with live-cell barcoding antibody mix against CD45. After four washes with CSB plus 1 mM EDTA, all 12 samples were mixed and then incubated for 30 min at RT with a surface antibody cocktail. The cells were then washed once with CSB plus 1 mM EDTA and incubated for 30 min at RT with nuclear antigen staining buffer concentrate (4×): nuclear antigen staining buffer diluent mixed at a 1:3 ratio (Standard Biotools). After washing twice with 2 ml of Maxpar antigen staining Perm (Standard Biotools), intracellular antibody mix was added, and the solution was incubated for another 30 min at RT. After washing once with 2 ml of Maxpar antigen staining Perm, the cells were fixed with a 1.6% FA solution for 15 min. Following a wash with Maxpar antigen staining Perm, the cells were incubated for 2 h at 4 °C in Maxpar Fix and Perm buffer solution containing Cell-ID Intercalator-Ir (1:1000; Standard Biotools). After two washes with CSB, the cells were washed twice with Maxpar Cell Acquisition Solution (CAS) (Standard Biotools). The supernatant was then removed by aspiration, and the cells were resuspended in a solution of EQ Four Element Calibration Beads (Standard Biotools) to CAS at a 1:9 ratio, followed by analysis using a Helios mass cytometer (Standard Biotools).

Prior to analysis, each barcoded MC sample was adjusted to ~1 × 10⁶ cells and exposed to CAS. The samples were then analyzed at an event rate of 300–400 events/s.

The original FCS files were normalized and subsequently de-barcoded using CyTOF software (ver. 7.0.8493, Standard Biotools) according to the manufacturer’s instructions. Each of the 12 FCS files, extracted as normalized and debarcoded CD45⁺ cells, was saved and prepared for further analysis using Cytobank.

MC clustering was performed based on a previously described method using Cytobank⁵⁸. CD8⁺ T cells were manually gated as CD3⁺CD8⁺CD4⁻ cells. Manual gating for PD-1 was also performed, as shown in Figure S4F. UMAP analysis was then applied to manually gated PD-1⁺CD8⁺ T cells from each of the 12 samples, involving a total of 17,616 cells equally sampled across all cases. The following markers were used to generate the UMAP plots: CCR6, CD45RA, CD4, Ki67, OX40, CD28, TIM3, TOX, CXCR3, CD27, CD39, Tbet, CD69, CD45RO, LAG3, CCR7, CD127, TCF1, CTLA4, CXCR5, CX3CR1, CXCR4, Eomes, and PD-1. Parameters were set to 30 neighbors and a minimum distance of 0.01. The resulting UMAP plots were fed into FlowSOM clustering with algorithm settings of 100 clusters and 10 metaclusters.

FlowSOM and UMAP analyzes were performed on eight available tumor samples, obtained through biopsy or resection, from the ICB-treated cohort of HNSCC patients³³ using Cytobank. CD8⁺ T cells were manually gated, as described in Figure S3I. FlowSOM analysis employed equal sampling for consistency across samples. For each SOM, 225 clusters and 9 metaclusters were identified for CD8⁺ T cells. The markers used for this analysis included PD-1, CD39, TCF-1, Ki67, TIM-3, CD127, CD69, CD103, GranzymeB, Tbet, CD39, CD38, CD25, CD45RA, CCR7, CD27, TCF1, TIGIT, and HLA-DR. The resulting clusters were then analyzed using UMAP, employing the same parameters as in FlowSOM with the algorithm adjusted for 15 neighbors and a minimum distance of 0.01.

Quantification by manual gating was performed on 8 tumor samples, 1 non-metastatic lymph node (LN) sample from the dataset of the ICB cohort of HNSCC patients, and 9 non-metastatic LN samples from the standard of care cohort³³. The expression rates of TCF-1, PD-1, CD39, and Ki67 in CD8⁺ T cells were determined based on the gating strategy shown in Figure S3I.

PBMCs collected pre-treatment and during nivolumab treatment were thawed and stained using a Zombie Aqua Fixable Viability kit (BioLegend). Cells were then stained with fluorochrome-conjugated antibodies against surface proteins of interest. For intracellular staining, the cells were fixed and permeabilized using Fixation/Permeabilization concentrate and Fixation/Permeabilization diluent (Thermo Fisher Scientific) and stained with fluorochrome-conjugated antibodies against intracellular proteins. PD-1 was detected using anti-human IgG4-Alexa Fluor 647 (HP6025, Southern Biotec) after saturation with nivolumab (Fig.). The following markers were used with an Attune flow cytometer (Thermo Fisher Scientific): Ki-67 (BioLegend, Ki67), CD4 (BioLegend, SK3), CD39 (BioLegend, A1), Eomes (eBioscience, WD1928), Tox (eBioscience, TXRX10), Tbet (BioLegend, 4B10), CD14 (BioLegend, M5E2), CD8 (BioLegend, SK1), TIM-3 (BD, 7D3), CD69 (BioLegend, FN50), TCF-1 (BioLegend, 7F11A10), OX-40 (BioLegend, ACT35), IFN-γ (BD, 4S.B3), and TNF (BD, Mab11). Gating for all markers except lineage markers was determined based on isotype controls, with a representative gating strategy shown in Figure. Further antibody details are described in Table. S4E S6G S5

Thawed cells were stimulated with phorbol 12-myristate 13-acetate (Sigma) at 0.25 μg/ml and ionomycin (Sigma) at 2.5 μg/ml for 3 h in a 37 °C incubator and then stained. Cytokine production was analyzed by intracellular staining using antibodies against IFN-γ (BD, 4S.B3) and TNF (BD, Mab11). A sample from an ineligible patient was used in this analysis.

PD-Ll staining of samples from each eligible patient was performed using FFPE tumor samples and an anti–PD-L1 antibody (1:100, 28-8, Abcam). The CPS was calculated by a pathologist (K.D.).

Fluorescence multiplex immunostaining was performed on FFPE samples from an enrolled patient to assess the staining quality of denoised images. Antigen retrieval was performed using Antigen Retrieval Solution pH 9 (Nichirei, Tokyo, Japan) and microwaving for 20 min. The primary antibodies used were as follows: PD-1 (Clone NAT105, ab52587, Abcam), CD39 (Clone ERP20627, ab223842, Abcam), and CD8 (Clone 4B11, PA0183, Leica Biosystems). All primary antibodies were incubated at RT for 90 min. The following secondary antibodies were used: anti-mouse IgG2b CF Dye 647 (Biotium), anti-mouse IgG1 CF Dye 488 (Biotium), and anti-rabbit IgG CF Dye 568 (Biotium). All secondary antibodies were incubated at RT for 40 min. DAPI was used as the nuclear counterstain. The images were captured using a Leica DMi8 THUNDER imaging system (Leica, Weztlar, Germany).

Statistical analyzes for group comparisons and correlations were conducted using the Prism 9.5.1 and JMP pro 16 software packages. Data normalization and denoising of IMC images were performed in Python 3.6. The treatment effect difference between the two TLS-based groups was compared using a two-sided Fisher’s exact test. The relationships of continuous variables between two groups were analyzed using non-parametric Mann–Whitney U-tests or Wilcoxon matched-pairs signed rank tests for two-sided unpaired and paired analyzes, respectively. For multiple groups, repeated measures ANOVA with post-hoc tests using Bonferroni correction (including one model using a mixed-effect model) or a Kruskal-Wallis test with post-hoc Dunn comparisons were used. Correlations between continuous variables were determined by Pearson’s r coefficient using a two-sided test. For all tests, P < 0.05 was considered statistically significant. PFS was defined as the time from the initiation of ICB therapy to the date of confirmed PD or death from any cause. The cutoff value for the frequency of CD39⁺ Tpex cells was determined using ROC curve analysis. The optimal threshold in which the sum of sensitivity and specificity was the maximum was detected.

Further information on research design is available in thelinked to this article. Nature Portfolio Reporting Summary