C57BL/6 J mice were purchased from The Jackson Laboratory. All animal experiments were approved by the Animal Experiments Committee of LUMC and were executed according to the animal experimentation guidelines of the LUMC in compliance with the guidelines of Dutch and European committees.

Metal conjugated antibodies were purchased from Fluidigm or conjugated to unlabeled antibodies in-house. All non-platinum conjugations were performed using X8 polymer as per manufacturer’s protocol (Fluidigm) and were performed at 100 μg scale. Conjugation with 208 Bismuth was performed using a protocol adapted from M. Spitzer [15]. All in-house conjugated antibodies were diluted to 0.5 mg/ml in antibody stabilizer supplemented with 0.05% sodium azide (Candor Biosciences). Appropriate antibody dilution was determined by serial dilution staining to minimize background and optimize detection of positively expressing populations.

CyTOF data were acquired and analyzed on-the-fly, using dual-count mode and noise-reduction on. All other settings were either default settings or optimized with tuning solution, as instructed by Fluidigm Sciences. After data acquisition, the mass bead signal was used to normalize the short-term signal fluctuations with the reference EQ passport P13H2302 during the course of each experiment and the bead events were removed [16].

To isolate immune cells from the tumor, solid tumors were excised after a flushing step to remove the blood from TME. Exclusion criteria were ulceration of tumors, incomplete or unsuccessful flushing (determined by an unexpected high numbers of B cells in the TME). Single-cell suspensions were then prepared by mechanical and enzymatic (collagenase D and DNase, Sigma-Aldrich) dissociation, followed by density gradient centrifugation on an 100% / 70% / 40% / 30% Percoll (GE Healthcare) gradient.

After staining cells according to van Unen et al. [17], we analyzed live immune cells from the TME. We set our gating strategy to live single cells, positive for CD45, and excluded reference beads. For further analysis, live CD45⁺ gated files were sample-tagged, their marker expression arcsinh5 transformed and subjected to dimensionality reduction analyzes in Cytosplore [18]. All markers were taken in account to process the clustering analysis except PD-L1, which is a marker used only as a quality control to check the efficacy of PD-L1 blocking antibodies. The PD-L1 blocking antibody we used (clone MIH5, rat-anti-mouse, IgG2a subtype) binds to FcyRIIb and FcyRIII but not to FcyRI and FcyRIV, and is not able to mediate specific killing or depletion [19]. By staining with the same antibody clone, PD-L1 downmodulation was determined to show the effectiveness of the provided therapeutic antibodies to block PD-L1 binding.

Pooled samples from control and PD-L1 treated groups were analyzed by hierarchical stochastic neighbourhood embedding (HSNE) [20] based on approximated t-distributed stochastic neighbourhood embedding (A-tSNE) [21]. The default perplexity and iterations of the HSNE analysis were 30 and 1.000, respectively. If some clusters showed a similar phenotype, they were manually merged in Cytosplore. For further data exploration, CD4⁺ T cell, CD8⁺ T cell, CD19⁺ B cell, CD11b⁺ myeloid cell lineages were analyzed separately. Downstream analysis was performed by Cytofast [14] and Cytofworkflow [22].

Diffusion map was generated with R using the cytofkit package [23] by displaying only CD3⁺ metaclusters identified by PhenoGraph [24] as a confirmation method of the HSNE clustering.

Reference standard samples were compared with each other by calculating the similarity between their respective t-SNE maps. We used the Jensen-Shannon (JS) divergence to quantify the similarity between t-SNE maps. After converting t-SNE maps into two-dimensional probability density functions, the similarity between two maps is quantified as the JS divergence between their corresponding probability density functions. We used the base 2 logarithm in the JS divergence computation, which results in a continuous range of JS divergence values between 0 (for identical distributions) and 1 (for fully disjoint distributions), the algorithm being provided by E.D. Amir [25]. The average overlap frequency (AOF) is determined as described by E.D. Amir [26]

MC-38 colon adenocarcinoma cells were injected at a dose of 0.3 × 10⁶ cells subcutaneously (s.c.) in the right flank. Antibodies blocking LAG-3 and PD-L1 were injected intraperitoneally and agonistic anti-ICOS antibodies were given subcutaneously, next to the tumor. Tumor diameter was measured every 2 to 3 days with a calliper and reported as volume using the formula (w × h × l) x (π/6).

Statistical analyses were performed using R software or Prism (GraphPad). Unpaired two-tailed t-tests were used for subset abundance comparisons.

To examine the effect of PD-L1 blocking therapy, we used the colorectal adenocarcinoma mouse model MC-38. Mice were inoculated with MC-38 tumor cells, and when tumors were established after 10 days (tumor volume of 30–40 mm³), mice were treated with PD-L1 blockade therapy or left untreated (control group) (Fig. 1A ). To identify biomarkers that respond to immunotherapy with PD-L1 blockade, we set up a CyTOF mass cytometry panel for in-depth phenotypic characterization of tumor-infiltrated lymphocytes (TILs) in preclinical tumor models, which allows kinetic dissection of anti-tumor immune responses. The panel consisted of 38 cell surface markers and was designed to identify the major lymphoid and myeloid subsets and to ascertain the differentiation and activation status of these subsets (Additional file 1: Figure S1). We isolated immune cells from the tumor 8 days after start of immunotherapy and stained the single cell suspensions followed by mass cytometry acquisition of 3.5 million cells in total. In parallel, tumor growth was measured to assess the therapeutic benefit of the PD-L1 blockade treatment. Treated animals displayed a significant delay in tumor progression or even had complete tumor eradication (Fig. 1B). To determine the effectiveness of the provided therapeutic antibodies to block PD-L1 binding, the cell surface expression of PD-L1 in the TME was assessed by staining with the same antibody clone (i.e. MIH5). Indeed, the PD-L1 expression on CD45⁺ tumor-infiltrated immune cells from the treated group was significantly decreased compared to control animals (Fig. 1C).Fig1Fig. 1

PD-L1-blocking treatment induces delay of MC-38 tumor growth. (a) Schematic of CyTOF mass cytometry experiment investigating the effect of PD-L1 antibody treatment on the TME. Mice were challenged with MC-38 tumor cells and subsequently tumor-bearing mice were treated either with PBS (n = 16 mice) or PD-L1 blocking antibodies (n = 16 mice). Tumors were isolated and analyzed by mass cytometry (CyTOF). Cluster identification was performed with HSNE and downstream analysis was done with Cytofast. (b) Tumor growth curves of individual mice in the control group (mock injected with PBS, blue lines) and PD-L1-treated group (red lines). (c) Frequency of CD45⁺ PD-L1⁺ cells in the TME 8 days after therapy starts displayed on a per-mouse basis with mean ± SEM. (d) Overview of the immune cell composition in the TME shown in percentage of cells on a per-mouse basis with mean ± SEM (n = 16 mice per group)

Because treatment with anti-PD-L1 has major effects on the expansion of the CD8⁺ T-cell compartment, we analyzed in detail the CD8⁺ TIL subset at this timepoint and identified 48 different CD8⁺ T-cell subsets (Fig. 2A). t-SNE clustering allowed distinction between naïve (e.g. cluster C28 expressing CD62L, CD27), effector (e.g. cluster C13 and C14 expressing CD54, CD38, CD27, CD44) and central-memory subsets (e.g. cluster C34 expressing CD54, CD62L, CD44, CD27). Remarkably, one cluster (cluster C4) displayed both activating (ICOS, CD69, CD43) and inhibitory receptors (PD-1, LAG-3, NKG2A). To visualize the distribution of each identified cluster, we displayed the abundance of each subset per treatment group (Fig. 2B). The t-SNE map overlaid with the expression of specific markers showed that the cluster C4 subset could be defined by the inhibitory molecule LAG-3 and the costimulatory receptor ICOS. Essentially, co-expression of ICOS and LAG-3 was highly specific to the PD-L1 blockade treated group (Fig. 2C, D). Further characterization of this subset also demonstrated up-regulation of the ectonucleotidase CD39, the early activation marker CD69, the inhibitory NKG2A receptor, and the activation/exhaustion cell surface marker PD-1. The CD8⁺ T-cell subset expressing both the activating and inhibitory molecules, referred hereafter as T_AI cells, represented approximately 17% of all the CD8⁺ T cells across individual mice in the PD-L1 blockade group compared to 7% in the control group (Fig. 2E). Next, we validated the presence of CD8⁺ T_AI cells by flow cytometry. We isolated TILs from the TME and stained for the markers ICOS, LAG-3, CD69, CD39 and PD-1. The CD8⁺ T_AI subset (CD8⁺, LAG-3⁺, CD39⁺, PD-1⁺, ICOS⁺) population could indeed be identified, and was more abundant following PD-L1 blockade therapy (mean = 22%, sd = 16%, n = 6) than in the untreated group (mean = 9%, sd = 8%, n = 6; p-value = 0.03 by Student’s t-test). In addition, we confirmed our findings in the MCA205 sarcoma model. We identified the CD8⁺ T_AI cells by flow cytometry and observed that PD-L1 treatment increased this subset as compared to the control untreated group (Additional file 1: Figure S3A).Fig2Fig. 2

Identification of CD8⁺ T-cell clusters in Tumor Infiltrating T-cell populations (a) Heatmap of all CD8⁺ T-cell clusters identified at day 8 after the start of the PD-L1 treatment. Data shown is based on t-SNE plots, and is pooled from the control and PD-L1 treated group. Level of ArcSinh5-transformed expression marker is displayed by a rainbow scale. Dendrogram on the top represents the hierarchical similarity between the identified clusters. (b) Average and SEM in percentage of each CD8⁺ T-cell cluster among the CD8⁺ T-cell population of control (blue bars) and PD-L1 group (red bars). (c) t-SNE plot of respectively 0.32 × 10⁶ and 0.35 × 10⁶ CD8⁺ T cells from control (blue) and PD-L1 treated (red) group. (d) Same t-SNE plots as above now showing the level of expression marker by a rainbow scale. The arrow identifies the cluster of interest C4 (having a shared CD8⁺ LAG3⁺ ICOS⁺ phenotype). (e) Bar graph showing the mean frequency of cluster 4 (± SEM, unpaired t-test). Individual mice belonging to the control (blue) and PD-L1 treated (red) group are indicated

We next analyzed whether PD-L1 blockade therapy-specific subsets were also apparent in the CD4⁺ T-cell compartment. The t-SNE algorithm identified 45 CD4⁺ T-cell subsets revealing the heterogeneous profile of the CD4⁺ T cells (Fig. 3A, B). Notably, as for the CD8⁺ T cells, one subset was identified that correlated with PD-L1 treatment (cluster C12) and displayed the activating molecule ICOS and the inhibitory molecule LAG-3. In addition, these CD4⁺ T_AI cells expressed CD27, CD39, CD43, CD44, CD54, KLRG1 and PD-1. The t-SNE map overlaid with the expression of specific markers showed that these subsets could also be defined by LAG-3, ICOS and CD39 and the co-expression of those markers was highly specific to the PD-L1 treated group (Fig. 3C, D). The T_AI subset of CD4⁺ T cells was also significantly more abundant, representing about 17% of the total CD4⁺ T-cell population within the tumor infiltrated immune cells of the treated group compared to 8% in the control group (Fig. 3E). Also, in the MCA205 tumor model, the CD4⁺ T_AI cells were identified and were increased by PD-L1 treatment (Additional file 1: Figure S3B).Fig3Fig. 3

Identification of CD4⁺ T-cell clusters in Tumor Infiltrating T-cell populations (a) Heatmap of all CD4⁺ T-cell clusters identified at day 8 after the start of the PD-L1 treatment. Data shown is based on t-SNE plots, and is pooled from the control and PD-L1 treated group. Level of ArcSinh5-transformed expression marker is displayed by a rainbow scale. Dendrogram on the top represents the hierarchical similarity between the identified clusters. (b) Average and SEM in percentage of each CD4⁺ T-cell cluster among the CD4⁺ T-cell population of control (blue bars) and PD-L1 group (red bars). (c) t-SNE plot of respectively 0.23 × 10⁶ and 0.25 × 10⁶ CD4⁺ T cells from control (blue) and PD-L1 treated (red) group. (d) Same t-SNE plots as above now showing the level of expression marker by a rainbow scale. The arrow identifies the cluster of interest 12 (having a shared CD4⁺ LAG3⁺ ICOS⁺ phenotype)

(e) Bar graph showing the mean frequency of cluster 12 (± SEM, unpaired t-test). Individual mice belonging to the control (blue) and PD-L1 treated (red) group are indicated.

To corroborate the results obtained from the previous t-SNE analysis regarding the PD-L1 treatment-associated T-cell subsets, we used the PhenoGraph algorithm to identify cell clusters and their differentiation status [24]. Similar T-cell metaclusters as those depicted by t-SNE earlier were indeed identified (Fig. 4A ). The CD4 and CD8 T-cell lineages could be distinguished into a resting phenotype (called CD44 ^low ), an activated intermediate phenotype without inhibitory marker expression (called CD44 ^int ), and the T_AI cells expressing both inhibitory and activation molecules (called T _AI ). To investigate the relationship between those metaclusters identified by PhenoGraph, we used the diffusion map algorithm [28].Fig4Fig. 4

Diffusion maps of the identified CD4⁺ and CD8⁺ subsets in the control and treated group. (a) Two-dimensional diffusion map of the CD4⁺ and CD8⁺ T cells present in the tumor at day 8 after the first PD-L1 treatment. Three different CD4⁺ and CD8⁺ T cell metaclusters have been identified by PhenoGraph. Continuity of the pattern reveals relationship between the different represented metaclusters (n = 5 mice per group). (b) Diffusion map of the CD4⁺ and CD8⁺ T_AI cell displayed by group origin (PBS in blue and PD-L1 in red). (c) Diffusion map of the CD4⁺ and CD8⁺ T_AI cell displayed by marker expression ICOS, LAG-3, CD39, PD-1 and CD69. (d) Expression levels of CD44 and CD62L on the metaclustered CD4⁺ and CD8⁺T cell populations. (e) Expression levels of ICOS, LAG-3, CD39, PD-1 and CD69 on the metaclustered CD4⁺ and CD8⁺T cell populations

PD-L1 blocking treatment enhanced CD4⁺ and CD8⁺ T_AI cell subsets in the TME 8 days post therapy. To determine if the expansion of these compartments occurred already early after treatment, we analyzed the TME at day 3 post-treatment (i.e. 13 days after tumor inoculation). The expansion of the CD4⁺ T_AI cells started at an earlier stage, 3 days post-therapy, and continued over time. The presence of the CD8⁺ T_AI cells could also be observed 3 days after the start of the treatment, but these cells significantly increased over time (Fig. 5A). Essentially, the vast majority of the CD39⁺ PD1⁺ CD8⁺ T cells that are present in the TME produce copious amounts of granzyme B, revealing their cytotoxic potential (Fig. 5B).Fig5Fig. 5

Quantification and cytotoxic capacity of the T_AI cells in the TME (a) Average percentage (and SEM) of T_AI cells within the CD8⁺ (left panel) and CD4⁺ (right panel) T-cell compartment at day 3 and 8 post PD-L1 blockade therapy in MC-38 tumor challenged mice. (b) Granzyme B expression of CD8⁺ T cell subsets at day 8 post PD-L1 treatment in MC-38 tumor bearing mice. The grey shaded histograms represent the CD39⁻ PD-1^+/− CD8⁺ T cells and the red shaded histograms depict the CD39⁺ PD-1⁺ CD8⁺ T_AI cells of individual mice. (c) The percentage of granzyme B⁺ cells among the CD39⁺ PD-1⁺ CD8⁺ T_AI cells after 8 days of PD-L1 treatment in the MC-38 tumor model compared to the CD39⁻ PD-1^+/− CD8 T cells

The data above indicate that the activity of anti-PD-L1 treatment could be mediated via the expansion of CD4⁺ and CD8⁺ T_AI cells that express activating receptors and inhibitory receptors. We assessed if we could further enhance the functionality of the T_AI cells by combining the PD-L1 blockade treatment with antibodies targeting inhibitory and stimulatory molecules. For the proof-of-principle, we performed co-treatment studies with blocking antibodies to the inhibitory receptor LAG-3 and with agonistic antibodies to ICOS during PD-L1 blockade (Fig. 6A).Fig6Fig. 6

Correlation between the presence of T_AI cells in the TME and tumor growth. (a) Regimen scheme of (combinatorial) antibody treatment after tumor injection. (b) Comparison of tumor growth between control group (n = 9), PD-L1 antibody treated group (n = 9), PD-L1 and ICOS antibody treated group (n = 11), PD-L1 and LAG-3 antibody treated group (n = 10). (c) Survival curves for each treatment mentioned above. (d) Study of the tumor-microenvironment after control (n = 6), single therapy (PD-L1, n = 8) or combinatorial therapy (PD-L1 and ICOS, n = 6) of the CD8⁺ T_AI-like cells (left panel) and CD4⁺ T_AI-like cells (right panel) at day 8 (unpaired t-test) displayed on a per-mouse basis with mean ± SEM

To extrapolate our findings in preclinical models to clinical settings, we questioned whether the T_AI cells were present within tumor-infiltrated immune cell populations in human tumors. We investigated the phenotype of the TILs in colorectal tumors of five patients, who have not undergone any immunotherapy. To reflect the immunogenicity of the MC-38 model, we selected MMRd colorectal cancer patients [29]. We designed our flow cytometry panel to characterize putative T_AI subsets within the tumor infiltrated CD8⁺ and CD4⁺ T cells. Hence, we included the activating receptors ICOS and CD69, also the inhibitory receptors like LAG-3 and CD39. We depicted the CD8⁺ T cells phenotypic diversity by gating on CD45⁺ CD8⁺ CD4⁻ cells and showed that a subset (cluster 8) with a similar phenotype (CD69⁺ ICOS⁺ and LAG-3⁺) as identified in mice tumors could be found in human tumors (Fig. 7A). The CD4⁺ T cell pool in human tumors contained a substantial fraction of cells with a CD69⁺ PD1⁺ phenotype, and within this population a CD39⁺ ICOS⁺ subset could be identified (Fig. 7B). Together, these results established that in tumors of mice and humans, CD4⁺ and CD8⁺ T_AI cell subsets are present.Fig7Fig. 7

Identification of the T_AI cell subset in humans. (a) Heatmap of CD8⁺ T-cell phenotypes (pre-gated on CD45⁺ CD3⁺ CD4⁻ cells) in tumors of 5 human colorectal cancer (MMRd) patients. Dendrogram above shows the hierarchical similarity between the identified clusters. Right panel shows the frequency of CD8⁺ LAG3⁺ ICOS⁺ cells (cluster 8) among total CD8⁺ T cells across the 5 patients. (b) Gating strategy to identify the CD4⁺ CD69⁺ PD1⁺ CD39⁺ ICOS⁺ population in human colorectal cancer. The abundance proportional to circle area is shown