High-Dimensional Mass Cytometric Analysis Reveals an Increase in Effector Regulatory T Cells as a Distinguishing Feature of Colorectal Tumors

Immune cells play an important role in colorectal cancer (CRC) progression (1). Both CD3⁺ T cell and CD8⁺ T cell infiltration have been associated with improved patient prognosis in numerous patient cohorts in the development of the Immunoscore (2), a method for staging CRC tumors based on CD3⁺ and CD8⁺ T cell infiltration at the invasive margin and the center of the tumor. However, the tumor microenvironment (TME) contains highly diverse populations of T cells that are capable of both promoting and hindering tumor progression (3–6). For example, 50% of patients with an intermediate Immunoscore had disease recurrence in a New Zealand cohort of 32 American Joint Committee on Cancer stage II patients (7); these data highlight the need to further improve outcome prediction.

Regulatory T cells (Tregs) are a heterogeneous population of T cells with suppressive function. The phenotype and suppressive function of Tregs are plastic and can change in response to environmental stimuli (8). In many human cancers, a high infiltration of Tregs is associated with poor patient prognosis (9). However, in CRC, high infiltrates of Tregs have been associated with both positive and negative patient outcomes, despite having a mainly negative role in many other cancer types (reviewed in Ref. 10). Multiple Treg populations with activated phenotypes have been identified in human CRC (10). The term “effector Tregs” (eTregs) has been given to a population of suppressive T cells that simultaneously express markers of both regulatory and activated T cells (11). In this study, eTregs were defined as CD3⁺FOXP3⁺BLIMP-1⁺ T cells in accordance with the literature describing murine eTregs (12).

In mice, expression of the transcription factor BLIMP-1 in Tregs is restricted to the eTreg population and is essential for the production of IL-10 (12). BLIMP-1 (gene: PRDM1) is a transcription factor that has important roles in T cells, B cells, and innate cell biology and function. The primary role of BLIMP-1 is to control expression of IL-2 in T cells (13). In mice, BLIMP-1⁺FOXP3⁺ T cells expressed elevated levels of ICOS and CTLA-4 and produced more IL-10 than BLIMP-1⁻FOXP3⁺ T cells (14). In humans, activated Treg subsets (CD45RA⁻FOXP3^hi) have been identified that produce IL-10, but BLIMP-1 expression in these subsets has not yet been investigated (15, 16). T cells that coexpress BLIMP-1 and FOXP3 were identified in ex vivo human CRC tissue (17). Infiltrates of BLIMP-1⁺FOXP3⁺ T cells were associated with longer disease-free survival (7), indicating they may play a direct role in the antitumor response to CRC. This recent stratification of eTregs from Tregs may explain the conflicting findings related to the role of Tregs in cancer.

Mass cytometry allows for detailed phenotyping of cell populations in conjunction with functional readouts. The high-parameter data generated by mass cytometry cannot be effectively analyzed using traditional cytometry analysis practices (18). Instead, population-based cluster analyses are more useful (18–21). These analyses group cells into visual “clusters” based on the similarity between every pair of cells. This allows for an understandable interpretation of the cell populations present without losing the detail of individual cell expression. Cluster analyses were recently used to define CD4⁺ T cell populations in human peripheral blood assessed by mass cytometry (22). This not only highlighted the considerable diversity within T cell populations but also demonstrated the relationships between the subsets.

In this study, we used mass cytometry, validated with flow cytometry, to define eTregs as a distinguishing feature of the TME in human CRC. These eTregs were FOXP3⁺CD25^hiCD127^loBLIMP-1⁺CD45RO⁺ICOS⁺, and T cells expressing BLIMP-1 and FOXP3 have previously been shown to have a positive prognostic role in CRC (7). We also observed a significant shift toward CD4⁺ T cells over CD8⁺ T cells in the tumor compared with nontumor bowel (NTB). Our data highlight the discovery benefit of high-dimensional acquisition and analysis approaches. Further, we used flow cytometry to show that eTregs in CRC tissue produce IL-10. Finally, BLIMP-1 expression in Tregs was upregulated upon stimulation with IL-2, potentially revealing a new function of IL-2 on Treg populations. Therefore, CRC tumors contain an abundance of eTregs that are protective, despite having anti-inflammatory properties.

Tissue samples for both ex vivo and in vitro analyses were obtained from patients undergoing elective surgery for CRC at Dunedin Hospital. The study was approved by the Health and Disability Ethics Committee (no. 14/NTA/33) and all patients gave written informed consent prior to inclusion in the study in accordance with the Treaty of Helsinki. Specimens were dissected by a pathologist. Sample numbers for two cohorts are provided in the figure legends; the patient characteristics for each cohort are described in Table I (for mass cytometry studies) and Table II (flow cytometry studies). PBMCs were obtained from healthy donors, and all participants gave written consent (University of Otago 12/036).

Samples were maintained in sterile phosphate buffered saline (PBS; Sigma-Aldrich, St. Louis, MO) at 4°C for no longer than 3 h until transportation on ice. Samples were washed in sterile PBS and suspended in RPMI 1640 (Invitrogen, Carlsbad, CA) with 0.5 mg/ml collagenase (Invitrogen) and incubated at 37°C, 5% CO₂, for 1 h. The tissue was then mechanically dissociated with a sterile scalpel. The tissue suspension was removed from the well and filtered with a 70-μm cell strainer (BD Falcon, Franklin Lakes, NJ) into a 50-ml Falcon tube. Live immune cells were enriched using a three-layer Ficoll–Paque Plus (Ficoll; GE Healthcare, Piscataway, NJ) gradient. Five milliliters of 75% Ficoll, 25% RPMI 1640 supplemented with 100 μg/ml penicillin, 100 μg/ml streptomycin, and 55 μM 2-ME (all from Invitrogen) were layered on top of 5 ml 100% Ficoll. A 2.5-ml cell suspension in RPMI 1640 was layered above this. Ficoll gradients were then centrifuged at 800 × g for 20 min with no brake. The buffy coat layer containing live immune cells was then carefully removed using a Pasteur pipette. Cells were either used fresh (for flow cytometry or in vitro experiments) or frozen (for mass cytometry experiments) in freezing media (90% FCS [PAA Laboratories, Morningside, QLD, Australia], 10% DMSO [Sigma-Aldrich]) in liquid nitrogen for storage.

Samples were transported to Sydney on dry ice for a maximum of 12 h and stored at −80°C prior to staining. Tissue samples from CRC patients were labeled for mass cytometry using established methods (23). All reagents were provided by the Ramaciotti Centre for Human Systems Biology, at the University of Sydney.

Cells were incubated in 50 μl of 5 μM cisplatin (5 min, room temperature; Fluidigm, South San Francisco, CA) and quenched with 5× volume of FACS buffer (0.02% sodium azide, 0.5% BSA, and 2 mM EDTA in PBS) to identify viability. An Ab master mix was prepared to apply to all paired sample tubes (the list of Abs can be found in Table III). CD45 barcoding was used so that all samples from one patient could be analyzed in the same tube (CD45-¹⁰⁴Pd for NTB and CD45-¹⁰⁸Pd for tumor) and discriminated during analysis. Cells were incubated with surface Abs (30 min, 4°C). Following washes with FACS buffer, cells were fixed and permeabilized with a FOXP3 buffer kit (eBioscience) according to the manufacturer’s recommendations and stained with an intracellular Ab mixture for 30 min at 4°C. Cells were washed and fixed overnight in 4% paraformaldehyde containing Cell-ID Intercalator-Ir (0.125 μM ^191/193Ir; Fluidigm). After multiple washes with FACS buffer and Milli-Q water, cells were filtered through a 35-μm nylon mesh and then analyzed on a Helios-upgraded CyTOF 2 Mass Cytometer (Fluidigm). The normalization beads were EQ Four Element Calibration Beads (Fluidigm), and the normalization procedure was carried out using the processing function within the Fluidigm CyTOF acquisition software.

Whole blood was diluted 1:1 with sterile PBS then layered over Ficoll. Samples were centrifuged at 800 × g (room temperature, 20 min with slow acceleration and deceleration). Interphase cells were removed and washed twice with sterile PBS. Freshly isolated PBMCs were cultured in six-well plates in RPMI 1640 (+10% FCS). Stimulated cells were incubated with 0.05 μg/ml IL-2 (Sigma-Aldrich). Anti-CD3 and anti-CD28 Abs (Sigma-Aldrich) were added to desired wells at 15 μg/ml. LPS (Sigma-Aldrich) was added to the cells at 2 μg/ml. Cells were cultured for 24 h at 37°C, 5% CO₂.

PBMCs or tissue suspensions were centrifuged at 1700 rpm for 4 min and resuspended in PBS. Live/Dead Zombie NIR or Live/Dead Fixable Red (BioLegend, San Diego, CA; Invitrogen) was added to the sample and incubated on ice in the dark for 30 min. Cells were resuspended in FACS buffer (PBS plus 0.5% FCS plus 0.01% sodium azide [PRO-LAB, Geldenaakaeban, Leucen]). The optimal quantity of each Ab was added to each tube, as determined by titration (Table IV). Fluorescence minus one controls were also prepared. Cells and Abs were incubated on ice in the dark for 20 min. Cells were then fixed in 1% PFA for at least 45 min before resuspension in FACS buffer and storage at 4°C. Cells were incubated in 0.5% saponin (Sigma-Aldrich) for 60 min with gentle agitation. Primary intracellular Abs were added in 0.5% saponin and incubated for 90 min with gentle rocking. Secondary anti-rat IgG–PE (BD Biosciences) was added with saponin and incubated for 90 min.

Compensation was calculated using OneComp eBeads (eBioscience, San Diego, CA) for Ab stains (eBioscience OneComp eBeads protocol). Flow cytometric acquisition was performed on an LSR FORTESSA (BD Biosciences, San Diego, CA) using FacsDIVA (version 6.2; BD Biosciences). Data were exported as FCS 3.0 files and analyzed using FlowJo (version X.0.7; Tree Star, Ashland, OR). All gates were set based on fluorescence minus one controls. Median fluorescence intensities were calculated using FlowJo software.

Analyses were performed using the validated CITRUS (19) and VorteX (20) packages as per the developers’ instructions. Preprocessing of data was performed in FlowJo (version X.0.7), and basic reorganization of data was performed in R prior to graphing in GraphPad Prism 7.0a (GraphPad, La Jolla, CA). For both VorteX and CITRUS, cells were clustered on expression of FOXP3, GATA-3, CD3, BLIMP-1, CD4, CD8, CD45RO, T-bet, RORγT, CD127, CD69, and CD25. An equal number of events was sampled from each tumor or NTB file.

The x-shift analysis was performed within the VorteX Java application. All nonclustering markers were included as functional only during sample importation. Angular distance was used for calculations. The n nearest neighbors value was used as a density estimate and the iterative k value for the x-shift algorithm ranged from 395 to 35 in 30 steps; this was determined automatically. Elbow-point validation was then used to determine the most representative (k) cluster number.

CITRUS was performed within the Cytobank web application. The minimum population frequency was set at 5% to prevent unnecessary division of populations. A 1% false detection rate was used to only include the most significant differences between tumor and NTB. A significance analysis of microarrays (SAM) test was used as the association model.

All non–cluster-based statistical analyses were calculated using GraphPad Prism 7.0a. A Wilcoxon signed-rank test was used for paired data. Correlations were determined using a Spearman rank correlation coefficient. Additional statistical methods are described in figure legends where appropriate.

Fig. 1 shows the workflow for mass cytometry data collection and analysis for this study. Colorectal tumor tissue and NTB were removed during surgery (Fig. 1, panel 1). These tissue samples were processed to a single cell suspension as described in 2Materials and Methods (Fig. 1, panel 2). Live immune cells were enriched using density gradient centrifugation (Fig. 1, panel 3). Samples were labeled with heavy metal–conjugated Abs and processed by a mass cytometer (Fig. 1, panels 4 and 5). Mass cytometry data were output as .fcs files that were then passed into a variety of analysis software as with flow cytometry data. Data points representing beads, debris, and dead cells were removed (18, 24). The data were also scrutinized for the presence of biologically impossible populations (for example, coexpression of mutually exclusive markers; Fig. 1, panel 6).

VorteX (20) was used to perform iterative, weighted k-nearest neighbor clustering (Fig. 1, panel 6a). Each iteration used a different k value for the clustering. Importantly, VorteX included an elbow-point validation step. This step calculated which iteration (and therefore which specific k value) best modeled the data into appropriate clusters. Practically, this ensured that each cluster represented a real population of cells and reduced the chance of overclustering, in which multiple clusters represented a single unified population. As such, VorteX effectively predicted the number of biologically unique populations in a data set.

Both force-directed (FD) and minimal spanning tree (MST) plots were used to visualize clusters. FD plots used gravitational forces to display nodes (cells) and edges (the relationship between the nodes). Cells that were similar across all variables grouped together, whereas unrelated cells were more distant. Tightly grouped regions of cells became specific clusters that represented a single phenotypic population. MST plots were simplifications of these clusters in which each node now represented a single cluster. In MST plots, if two nodes were close together they were most likely similar; however, distant nodes were not necessarily dissimilar. As such, MST plots functioned well as simplifications but were less indicative of population relationships than FD plots.

CITRUS (19) was used to cluster the data into nodes (Fig. 1, panel 6b). Classifying features of each sample were then determined using a SAM. A classifying feature referred to a node with a unique phenotype that was significantly increased or decreased in a given sample type. This highlighted phenotyped clusters that were significantly different in abundance between sample types. These were called features (Fig. 1, panel 6b). The density plots shown in Fig. 1, panel 6b show the relative expression of each marker in each cluster (red) against the average expression of the cell assessed (purple).

There are numerous studies showing the anticancer effect of T cells (reviewed in Ref. 25); however, the protective effect is not consistent across patients, cancer types, and cancer stages. Our finding that the infiltrate of a subset of Tregs (BLIMP-1⁺ eTregs) was associated with improved prognosis (7) highlights the complexity of the antitumor immune response and the need to understand the heterogeneous populations in the tumor and NTB.

T cell populations (Cisplatin⁻, DNA⁺, CD3⁺) in human colorectal tumor and associated NTB from the same patients were assessed for expression of 34 surface markers, transcription factors, and secreted proteins using mass cytometry (see 2Materials and Methods). These data sets were analyzed using an x-shift, k-nearest means cluster analysis in VorteX to identify cell phenotypes within the tumor and NTB infiltrating T cell populations. Sixteen unique T cell phenotypes were designated by the algorithm using the elbow-point k value of 221 as present in either or both tissue types (Fig. 2). Fig. 2A shows an FD layout of the clusters colored by cluster membership. Color assignment is categorical and randomly assigned to each cluster. Each dot represents a single cell. Clusters can be clearly distinguished by their focal point clustering into visual nodes. The lines between each dot are edges that indicate relationships between cells. The eight largest clusters have been identified by a cluster identification number (1–8). Only the eight largest were labeled, given the relatively low abundance of the remaining eight clusters. Fig. 2A acts as a reference for these clusters in later figures.

As a validation of the clustering integrity, CD4 and CD8 expression were examined within each node using an MST representation of the clusters (Fig. 2B). Each circle in these plots represents one of the clusters in the FD plot in Fig. 2B. There is a clear bipartite division of the MST plot into CD8⁺ populations at the bottom (Fig. 2B, left) and CD4⁺ populations at the top (Fig. 2B, right). To determine whether we could detect differences in cluster abundance between tumor and NTB we plotted the same MST plots divided into tumor- and NTB-derived cells (Fig. 2C). Cluster 2 was most abundant in the tumor, whereas Cluster 1 was most abundant in the NTB.

The x-shift algorithm identified 16 unique T cell clusters present in the tumor and NTB. As an initial overview of the phenotypic differences between all the clusters, we produced a heatmap of all marker expression across all 16 clusters (Supplemental Fig. 1). Next, we produced FD plots and overlaid expression of broad phenotypic and functional markers to determine the functionality of these populations. The VorteX plots shown identify the same 16 clusters (the largest eight labeled 1–8) as shown in Fig. 2A. CD4 and CD8 expression clearly separate into different clusters in the FD plot (Fig. 3A, 3B) as they did in the MST plots (Fig. 2B). IFN-γ was expressed in both CD4⁺ and CD8⁺ populations (across multiple clusters; Fig. 3C). Granzyme B was expressed in many clusters; however, the highest expression was in a cluster of CD8⁺ T cells and a cluster of CD4/CD8 double-negative cells (Fig. 3D). There was little variation in granzyme B expression among CD4⁺ T cells. IL-2 expression was mostly limited to two independent but closely related CD4⁺ clusters (Fig. 3E). Interestingly, T-bet expression mimicked granzyme B expression patterns almost perfectly (Fig. 3F), whereas RORγt expression predominantly coincided with IL-2 expression. RORγt was also expressed by a subset of the two granzyme B high clusters, the CD8⁺ and CD4/CD8 double-negative clusters. GATA-3 was predominantly expressed in CD4⁺ T cell clusters, particularly the cluster at the topmost point of the plot, although there was also some GATA-3 expression in the IFN-γ⁺ CD8⁺ cluster.

A CITRUS analysis was used to determine if any unique T cell populations were differently enriched in tumor or NTB. The CITRUS algorithm produced 32 unique clusters as opposed to the 16 predicted by VorteX. Of the 32 populations identified, 10 were highlighted as significantly different in abundance between tumor and NTB when using a 5% minimum population size and SAM with a 1% false detection rate (Fig. 4A). Table V indicates the frequency of each population in the NTB or the tumor (for example, of all of the cells represented in Cluster 2383, 8.2% were in the NTB and 91.8% were in the tumor). Of these populations, two were highlighted as features that were increased in the tumor compared with the NTB. Red nodes indicate a significant difference in abundance for that cluster between tumor and NTB tissue; blue nodes indicate no difference in abundance. Orange rings around red nodes indicate an increase in the abundance of this cluster in the tumor compared with NTB; green rings indicate a decrease. One of the most striking findings was that both clusters that were observed at a greater abundance in the tumor than NTB were CD4⁺ rather than CD8⁺ T cells. Furthermore, these cells appeared to be a locally stable population, based on their expression of activation and exhaustion markers (Table V), Clusters 2383 and 2392. In addition, the ratio of CD4⁺/CD8⁺ T cells in the tumor was significantly increased compared with NTB (Fig. 4B). A large proportion of the T cells that were enriched in the NTB were naive-like CD4⁺ and CD8⁺ populations that were potentially transient T cell populations in the vasculature. This was implied by their lack of markers indicating Ag experience (CD45RO, CD127), exhaustion (PD-1, TIM-3, LAG-3), or activation (CD69, CD25); however, the most enriched populations in NTB were effector T cells expressing IFN-γ, TNF-α, and granzyme B. This was unexpected, as effector-like CD8⁺ populations are the presumed protective T cells in most cancers, although the presence of the tumor may indicate a failure to mount an appropriate CD8⁺ T cell response and may explain the lack of these cells in the NTB samples.

Cluster 2383 was the cluster most enriched in the tumor compared with NTB: 91.8% of the cells in Cluster 2383 were from tumor samples (Fig. 4C; 2.2% of total CD3⁺ lymphocytes). This cluster was CD4⁺, highly expressed CD25 and FOXP3, and was CD127⁻ (Fig. 4B). These markers suggested that the cluster was a Treg population. Cluster 2383 also expressed CTLA-4 and PDL-1 (CD274), further supporting this identification as a Treg population. In addition to the Treg markers, Cluster 2383 highly expressed CD45RO, ICOS, BLIMP-1, PD-1, PD-L1, and TIM-3 (Fig. 4B). This phenotype suggested an activated or effector-like population (10), and we identified these cells as eTregs. Cluster 2392 was the second-most enriched population in the tumor tissue. This cluster shared a very similar phenotype to Cluster 2383 and may be the result of overclustering by the CITRUS algorithm (Table V). Regardless of overclustering, this still demonstrates that the defining T cell feature of a colorectal tumor is a relative increase in the frequency of eTregs. Fig. 4C shows the relative abundance of Cluster 2383 for each patient. The equivalent data for all 10 clusters identified as different between tissues in Fig. 4A are shown in Supplemental Fig. 3. Although there is interpatient variation, Cluster 2383 is at a greater abundance in all but one patient. Furthermore, the CITRUS algorithm only displays significant differences in abundance. As such, both eTreg populations are significantly increased in the tumor compared with the NTB despite patient variation.

To validate the findings from the CITRUS analysis, matched colorectal tumor and NTB samples from patients were assessed from two independent cohorts. In Fig. 5A, 5B, we acquired data with flow cytometry using an Ab panel designed to detect the BLIMP-1⁺ eTregs described by the CITRUS algorithm. These samples were from the cohort of patients described in Table II, but processing and data analysis were performed in the same manner as for previous experiments. There was a higher frequency of BLIMP-1⁺ eTregs in tumor tissue compared with NTB (Fig. 5A). Furthermore, FOXP3 expression was confined to BLIMP-1⁺ T cells in the tumor tissue (Fig. 5B).

As a further validation, we used the VorteX FD plots of the mass cytometry data assessed earlier (from patients described in Table I) to determine if the phenotypes of the clusters highlighted by CITRUS represented genuine cell populations. As observed with Clusters 2383 and 2392, CITRUS may overcluster the data, whereas the x-shift analysis performed using VorteX uses cluster elbow-point validation to avoid overclustering. The tumor-enriched cluster regions are shown in purple. The VorteX plots shown identify the same 16 clusters (the largest eight labeled 1–8) as shown in Fig. 2A. These clusters were predominantly on the top-center/right of the cluster map (Fig. 5C). From Fig. 3, the right and left regions represent CD4⁺ and CD8⁺ T cells, respectively. The majority of the tumor tissue–derived cells fell within the CD4⁺ region (Figs. 2A, 6A), which is in agreement with the CITRUS findings (Fig. 4, Table V). Of the eight large clusters, two clusters were significantly enriched in the tumor (Clusters 2 and 5) and two were enriched in the NTB (Clusters 1 and 6) (Fig. 5D). Interestingly, IFN-γ and IL-2 production was almost entirely limited to the tumor-enriched clusters. (Figs. 3C, 3E, 5C). The difference between the abundance in NTB and tumor of the eight largest clusters identified by number in Fig. 2A is shown in Fig. 5D.

We have highlighted two populations within the tumor-enriched regions of the FD plot (Figs. 5, 6). The VorteX plots shown identify the same 16 clusters (the largest eight labeled 1–8) as shown in Fig. 2A. The most interesting finding, which also validated the CITRUS finding, is highlighted by the red-outlined region (Cluster 5, from Fig. 2A). The CD4⁺ T cell cluster in the center of the cluster map (Cluster 5) expresses FOXP3 (Fig. 6E), CD45RO (Fig. 6C), CD25 (Fig. 6F), BLIMP-1 (Fig. 6G), and ICOS (Fig. 6H) and is negative for CD127 (Fig. 6D). Supplemental Fig. 2 shows a negative correlation between FOXP3 and CD127 in this tumor-enriched cluster. This pattern of expression matches Cluster 2383 from the CITRUS analysis and supports the hypothesis of a BLIMP-1⁺ eTreg-defined tumor environment (Fig. 4, Table V). This VorteX cluster represented 11.4% of total CD4⁺ T cells in the tumor. These data are also represented as a heat map in Supplemental Fig. 1.

Interestingly, these plots also revealed an activated CD4⁺ population (Cluster 2) that was enriched in the tumor. Cells in this region expressed an activation phenotype of CD45RO⁺ BLIMP-1⁺ and FOXP3^−/lo, as well as IFN-γ⁺and IL-2⁺, as also shown in Fig. 3.

Interpatient variation is important to address with a small patient cohort. We assessed the cells in the Treg cluster on a single cell basis to determine the effect of interpatient variation (Fig. 7A, 7C). There was a significant increase in the frequency of BLIMP-1⁺ Tregs out of total T cells in the tumor tissue compared with the NTB in every patient (Fig. 7B). Furthermore, T cells infiltrating the tumor had higher expression of BLIMP-1, ICOS, CTLA-4, and CD45RO than the T cells from the NTB (Fig. 7C–F). BLIMP-1⁺ Tregs from tumor tissues, from the second cohort (analyzed by flow cytometry), also had higher expression of CD45RO, CTLA-4, and ICOS than BLIMP-1⁻ Tregs (Fig. 7G–I). BLIMP-1⁺FOXP3⁺Tregs produced more IL-10 than BLIMP-1⁻FOXP3⁻ Tregs (Fig. 7J). CD4⁺ T cell IL-10 production correlated with the frequency of BLIMP-1⁺ Tregs in colorectal tumor tissue (Fig. 7K). It is possible that these BLIMP-1⁺ Tregs may be the primary T cell producers of IL-10 in the TME; in the mass cytometry cohort, there was a strong correlation between BLIMP-1 expression and IL-10 production in VorteX clusters (Fig. 7L).

Given the role of these cells in the TME by producing IL-10 and their significantly reduced frequency in NTB, we hypothesized that local factors in the TME promote this phenotype. In mice, the promotion of BLIMP-1 expression in Tregs has been observed to occur after a strong TCR signal (26), and stimulation of Tregs with IL-2 both in vivo and in vitro increased BLIMP-1 expression in Tregs (12). To determine if IL-2 was involved in the expression of BLIMP-1 in human Tregs in the TME, PBMCs from healthy donors were stimulated in the presence of IL-2. PBMCs were used because it is not possible to isolate tumor-naive T cells from tissue samples to perform these experiments. We have previously shown that healthy PBMCS respond similarly to IL-2 as PBMCs from people with CRC (27). With IL-2 in the stimulation medium, frequencies of BLIMP-1⁺FOXP3⁺ increased compared with cultures stimulated without IL-2 (Fig. 7M). IL-2 can also increase the expression of FOXP3 in T cells (28); thus, the frequency of BLIMP-1⁺ T cells (of the FOXP3⁺ population) was measured. The frequency of Tregs that were BLIMP-1⁺ increased after stimulation with IL-2 (Fig. 7N, 7O).

We used mass cytometry to confirm that the TME in CRC is distinct from NTB from the same patients, supporting previous data (3–6). Importantly, in addition to providing a detailed overview of the infiltrating immune response in CRC, we identified a population of eTregs that produce IL-10 as a distinguishing feature of the tumor that was almost completely absent from NTB; these eTregs have previously been associated positively with patient outcome (7). In vitro, we showed that IL-2 can promote this eTreg phenotype in the presence of anti-CD3 stimulation, which may indicate that IL-2 in the TME promotes enrichment of this population. Unexpectedly, we found that the tumor was dominated by CD4⁺ T cells as opposed to CD8⁺ T cells. Our data acquisition and analysis approach highlight the diversity within the TME and validates high-dimensional data as a discovery tool in cancer.

There are two important considerations concerning eTregs in this study. First, eTregs were significantly enriched in the tumor tissue compared with NTB, representing ∼11% of total infiltrating CD4⁺ T cells (data not shown). BLIMP-1⁺FOXP3⁺ T cells were associated with longer disease-free survival in a small cohort of stage II CRC patients (7), indicating these cells may have a direct antitumor role. Recently, two distinct populations of Tregs were defined in CRC patients: those expressing high levels of FOXP3 and those expressing low levels of FOXP3 (29). A high infiltrate of FOXP3^lo but not FOXP3^hi Tregs was associated with good patient prognosis. The FOXP3^lo Tregs also produced inflammatory cytokines, implying that they may be a population of eTregs. Populations of eTregs have also been implicated in patient survival in other cancers. In patients with B cell lymphoma, eTregs were identified by the coexpression of FOXP3 and CTLA-4, and high infiltrates of these cells were associated with more advanced tumor stage (30). High infiltrates of eTregs, characterized by CD45RA⁻FOXP3^hi expression, were associated with late stage CRC (31) and worse disease-free survival in head and neck cancer (32). In CRC, the infiltrate of CD4⁺FOXP3⁺ Tregs has been associated with both poor and favorable outcomes despite being mainly associated with poor outcomes in other human cancers (10). It has been hypothesized that in CRC CD4⁺FOXP3⁺ Tregs may be associated with favorable outcomes because their involvement in the suppression of inflammatory responses (33), which is associated with unfavorable outcomes in CRC patients (6).

The second important consideration regarding eTregs is that we showed insight into their mechanism of protection via IL-10 production. Is IL-10 production by eTregs protective in CRC? There is evidence in mice that a lack of functional IL-10 results in the development of severe intestinal inflammation and spontaneous CRC, indicating that IL-10 is essential for the control of inflammatory-driven CRC (34). More recently, others have also demonstrated the importance of IL-10–mediated suppression of inflammatory responses in the colon and tumors (35, 36). BLIMP-1 expression in T cells in the colon has been associated with lower frequencies of IL-17–producing cells in mice (37). Therefore, it is possible that the antitumor role of BLIMP-1⁺FOXP3⁺ T cells in CRC may be the inhibition of an inflammatory response via the production of IL-10, although specific experiments addressing this hypothesis are essential.

CD4⁺ T cells dominated the tumor T cell infiltrate. This CD4⁺ T cell population in the tumor was heterogenous, but almost all populations expressed markers associated with prolonged activation. This phenotype existed regardless of whether the CD4⁺ T cells were defined as eTregs. CD45RO, which is upregulated upon Ag experience (38), was expressed on all CD4⁺ T cell populations in the tumor. Markers such as BLIMP-1 and ICOS, which are upregulated on activated cells (10, 11), were also upregulated in these populations. Although the eTreg populations did not express high levels of cytokines, the remaining non-eTreg CD4⁺ T cell populations produced effector cytokines such as IFN-γ, TNF-α, and IL-2 in addition to IL-6 and IL-10, which are important mediators of the TME (35). These cells may represent the FOXP3^lo population identified by Saito et al. (29).

All CD4⁺ T cells in the tumor tissue expressed inhibitory receptors (PD-1, LAG-3, and TIM-3), although there was significant variability in which inhibitory receptors were expressed between populations. Inhibitory receptors, such as PD-1, are commonly used as markers of T cell exhaustion. T cell exhaustion refers to a loss of cytokine production and proliferation compared with equivalent cells that are inhibitor receptor negative or to cells before chronic stimulation (39). This expression pattern has led to the implication that inhibitory receptors are causative of exhaustion when, in fact, they may only be “guilty by association” (40). It has been shown that inhibitory receptors can also be upregulated on healthy activated cells. PD-1^hi T cells in both healthy human and mouse breast cancer models retain cytokine production, indicating that PD-1 expression does not directly inhibit cytokine production (41, 42). Instead, these inhibitory receptors may simply be indicative of activation state. Another inhibitory receptor, TIM-3, was preferentially expressed on naive T cells and was downregulated upon Ag experience (40, 43). In this study, TIM-3 was only expressed on the eTreg populations in the tumor. These findings give rise to two unanswered questions in the context of this study: how do inhibitory receptors affect the function of eTregs, and how does this affect patient outcome with regards to checkpoint blockade therapies? Answering both of these questions is critical to improving the efficacy of immunotherapies in CRC.

A high infiltrate of CD8⁺ T cells is associated with improved patient prognosis in most tumors, including CRC (2). Despite this, we showed that CD4⁺ T cells were the dominant T cell population in the tumor. This may indicate that CD4⁺ T cells have a greater impact on tumor progression than CD8⁺ T cells. However, the fact that there is a tumor present may indicate that the immune infiltrate represents failed immune control of the tumor. As such, the high CD4⁺/CD8⁺ T cell ratio may be beneficial to tumor progression. Contrary to this, eTregs make up a large proportion of the CD4⁺ T cell infiltrate, and these cells have been associated with improved disease-free survival (7), suggesting that they hinder tumor progression. Collectively, these data highlight the importance of studying the heterogeneity of the immune response not just to better define the TME but because individual T cell populations may have considerable therapeutic influence.

Mass cytometry opens the possibility of examining the expression of a substantial number of proteins across an entire population on a per cell basis. This allowed us to perform a detailed functional analysis of multiple specifically defined subsets simultaneously. Moreover, all experiments were performed on primary tissue, either patient tissue samples or PBMCs. We then validated our primary findings using flow cytometry. As a further validation of our findings, we used multiple cluster-based analysis approaches to determine which cell populations were defining of the colorectal tumors. These approaches demonstrate the strength of cluster analyses for highlighting the diverse range of T cell populations present in human tissue. Using the elbow-point clustering validation in VorteX provides some mathematical certainty as to the actual number of unique populations present, whereas other approaches may under- or overrepresent the number of populations in the data.

As a consequence of using freshly acquired tissue, it was not possible to include associations between tumor-infiltrating cells and clinical outcomes such as disease-free survival. This study lacked direct functional readouts to confirm the function of eTregs in CRC; however, we did show that they produce IL-10 and express several markers associated with activation. There are several other limitations to this study. First, the sample size is small and heterogeneous with respect to important variables such as site of disease and stage. Second, the study was performed with freshly collected specimens, and there are no outcome data available at this time. Nevertheless, the immune features described have been associated with outcomes in previous studies and the current work provides important insights into possible mechanisms.

We propose that eTregs represent a substantial, protective population of CD4⁺ T cells in colorectal tumors. These cells could be targeted by new immune therapies to improve patient survival. We have demonstrated the heterogeneity of T cell populations in CRC that appear to be primarily variants of activated-phenotype CD4⁺ T cells. We have added a new level of specificity to our understanding of T cell infiltration into CRC, which in turn gives us the possibility of identifying more immune targets for treatment or prognosis. Together, we have demonstrated the strength of using mass cytometry and cluster-based analysis to examine immune cell populations in cancer. This work provides a platform for future investigations and demonstrates the necessity of high-dimensional analysis of complex immune populations.

We thank our donors for contributing to this study, Andrew Mitchell for mass cytometry support, and Alex McLellan for critical review of the manuscript.

The authors have no financial conflicts of interest.