Pan-cancer single cell transcriptomic clustering reveals heterogeneous CD8⁺ exhausted T cell populations with different immune checkpoint inhibitor responses

ABSTRACT

In most cancers, T lymphocytes comprise an essential cellular component of the non-neoplastic microenvironment, where they have the capacity to both suppress and support tumor growth. One specialized T lymphocyte population is the CD8⁺ exhausted T cell, which has been intensely studied as an actionable therapeutic target. Unfortunately, there is currently no uniformly accepted classification scheme for these specialized T cells. To provide a potential model for classifying CD8⁺ exhausted T cells, we leveraged single cell transcriptomic analysis of a diverse collection of both human (n = 8) and mouse (n = 4) cancers to identify unique subpopulations shared across tumor types and species. By integrating data from both human and mouse cancer studies, as well as previously described CD8⁺ exhausted T cell subsets, we provide an integrated framework to characterize the heterogeneity of exhausted CD8⁺ T cells. As such, one of these subpopulations (cluster C1) increases following immune checkpoint inhibitor treatment in the setting of cancer in mice and patients. Taken together, this proposed classification scheme may be useful for the design and interpretation of current and future immune-based therapy studies.

Introduction

Initially identified as a functional state of CD8⁺ T cells in the setting of chronic infection,¹ CD8⁺ exhausted T cells were later discovered to represent a targetable immune cell population in cancer.^2,3 In this context, effector cytotoxic CD8⁺ T cells expressing specific receptors (e.g., PD-1, TIGIT) engage with ligands expressed on tumor cells (e.g., PDL1, CD155) to become quiescent (exhausted). These exhausted CD8⁺ T cells can be therapeutically reactivated to kill tumor cells by abrogating these ligand/receptor interactions.⁴ Examples of such immune checkpoint inhibitor (ICI) therapies include ipilimumab (targeting the CTLA4 receptor), nivolumab and pembrolizumab (targeting the PD1 receptor), and tiragolumab (targeting TIGIT receptor interactions).⁵ While some of these treatments have shown efficacy in clinical trials, others have been less successful.⁶ We hypothesize that these responses may reflect the heterogeneity of CD8⁺ exhausted T cell populations in cancer.

To gain insight into the landscape of CD8⁺ exhausted T cells in cancer, we leveraged nine single cell RNA sequencing (scRNAseq) datasets representing eight distinct human cancer histotypes. Using this approach, we identified five subpopulations of CD8⁺ PD1⁺ (PDCD1-expressing) exhausted T cells common to all eight human cancers, four of which were shared in mice. We then integrated our subpopulations with previously reported exhausted T cell classifications and synthesized these findings into a new contextual framework.

Materials and methods

Mouse ICI treatments

This manuscript adheres to ARRIVE guidelines. All experiments using mice (Mus musculus) were performed under an active Institutional Animal Care and Use Committee (IACUC) protocol (#22–0361) at the Washington University School of Medicine (Washington University in St. Louis Institutional Animal Care and Use Committee). Mice were maintained as a continuous breeding colony at Washington University on a strict C57BL/6 background and were housed in a specific pathogen-free barrier facility with controlled temperature (21 °C), humidity (55%), light-dark cycles (12:12 h) and ad libitum access to food and water. According to the ethics committee at Washington University, any animals exhibiting impaired motion, poor eating habits, or an unhealthy appearance must be euthanized. No animals were euthanized due to tumor burden or due to the treatments administered. For the ICI therapy studies, male and female mice harboring a neomycin cassette inserted into exon 31 (Nf1^+/-) and hGFAP-Cre-mediated Nf1 loss in neuroglial progenitors (Nf1^flox/neo; hGFAP-Cre) (Nf1-OPG mice⁷), were randomly assigned to receive anti-PD1 (BE0146; 200 μg/dose, i.p., twice per week, n = 10 mice) or anti-TIGIT (BE0274; 200 μg/dose, i.p., twice per week, n = 10 mice) antibody treatment at 12 weeks of age for four consecutive weeks prior to processing for single cell RNA sequencing as previously reported.⁸ Mice were euthanized at the end of these experiments by an overdose of Fatal Plus Pentobarbital solution (86.6 mg/ml). Investigators were blinded to group allocation during both data collection and analysis.

Single cell RNA sequencing (scRNAseq)

Nf1-OPG optic nerves were collected from ten mice of both sexes and pooled to prepare a single cell suspension as previously described.⁸ Live cells were collected and sent for scRNAseq.

Data collection and quality controls

We established inclusion and exclusion criteria to identify datasets for analysis. Inclusion required: (1) publicly available datasets from peer reviewed publications, (2) availability of raw count data, (3) scRNAseq datasets from solid cancers, and (4) sequenced biopsies from treatment-naïve patients. Exclusion criteria included: (1) UMAPs lacking a T cell cluster, (2) missing or incomplete patient data, and (3) datasets containing high mitochondrial counts, indicating lower quality cells. Single cell transcriptomic data were downloaded from GEO (https://www.ncbi.nlm.nih.gov/geo/). Cell quality was evaluated by number of reads, detected gene count, and percentage of mitochondrial reads. Cells with <200 counts or >5% mitochondrial gene percentages were removed. The use of human datasets previously deposited in the GEO repository were performed in strict accordance with the principles stated in the Declaration of Helsinki. Since these datasets were generated by investigators outside of our institution, do not contain personal health information (PHI), and no direct contact with the subjects is possible, this study was deemed exempt by the Institutional Review Board at Washington University.

To ensure analytical rigor, we applied quality control measures in addition to the above inclusion and exclusion criteria, and retained datasets that met the following criteria: (1) used in published studies for which experimental details were available, (2) passing previously established QC thresholds (see above), and (3) containing CD8⁺ exhausted T cells. After applying these criteria, only 9 of the available 19 human datasets and 6 of the available 20 mouse datasets were included. The datasets used and rejected are listed in Supplementary Tables 1–4.

For GEO datasets containing mice treated with anti-PD1 antibodies or patients treated with anti-PD1/PDL1 antibodies, only those that met the following criteria were used in addition to the preliminary inclusion and exclusion criteria: (1) QC compliance, (2) matched pre- and post-treatment samples, and (3) not including other treatments (chemotherapy, radiation, or other ICI therapies). After applying these criteria, 3 human and 8 mouse datasets remained. Unfortunately, only one human and one mouse dataset contained > 1500 CD8⁺ exhausted T cells, which is the minimal number of cells required for the algorithm to obtain precise subclustering. The datasets used and rejected are listed in Supplementary Tables 5–6.

Data normalization, integration, and clustering

Data were analyzed using Seurat (v5.1.0) in R (v4.4.1). SCTransform (v2) (SCTransform with vars.to.regress = c(‘percent.mt,’ ‘percent.rib’)) was applied for data normalization, followed by PCA. Harmony was applied for data integration using Seurat IntegrateLayers with ‘method = HarmonyIntegration.’ FindNeighbors with top 30 PCs followed by FindCluster with resolutions set by Clustree were used to identify clusters. To determine the clustering resolution for the human and mouse datasets, we employed the Clustree R package,⁹ which visualizes how cells transition between clusters across different resolution parameters. By plotting the hierarchical relationships among cluster assignments at successive resolutions, Clustree enabled detection when cluster identities become stable. Specifically, it examined how clusters split or merge across resolutions and evaluated the stability and biological coherence of these transitions. The optimal resolutions given by Clustree package are 0.5 and 0.8 for the human and mice cancer datasets respectively. UMAPs were generated using Harmony embeddings to visualize cell clusters.

CD8⁺ exhausted T cell identification and subtypes classification

In selecting CD8⁺ exhausted T cells for analysis, we used the following code in R: exhausted T < −subset(T cells, subset = CD3E > 0 & CD8A > 0 & PDCD1 > 0 & MKI67 = = 0 & CD4 = = 0, slot = ‘counts’). We selected cells based on gene expression, measured as SCTransform corrected counts: CD3E > 0 (CD3E was detected in cells with at least one RNA molecule captured), CD8A > 0 (CD8A was detected in cells with at least one RNA molecule captured), and PDCD1 > 0 (PDCD1 was detected in cells with at least one RNA molecule captured). To ensure that the analysis includes only CD8⁺ exhausted T cells, we excluded MKI67 expressing cells (MKI67 = = 0, meaning that MKI67 was not detected in that cell and the expression level was zero) and CD4-expressing cells (CD4 = = 0, meaning that CD4 was not detected in that cell and the expression level was zero). Gene markers for each CD8⁺ exhausted T cell subpopulations were identified by the FindMarker using Wilcoxon rank-sum tests, filtered by FDR < 0.05 and log2(fold change) > 1 and ranked by log2(fold change) (Supplementary Tables 7–8). AverageExpression and DoHeatmap were applied to visualize the genes in heatmap representations. DotPlot was used to visualize gene expression by dot plots. Pathway analysis was performed using the DAVID pathway platform (https://david.ncifcrf.gov/home.jsp.) employing the top 25 genes in each subtype, filtered by FDR < 0.1, and visualized by R ggplot. Human cancer, mouse cancer, and mouse lymphocytic choriomeningitis virus (LCMV) datasets were analyzed independently using the workflow above.

Doublet removal

After selecting CD8⁺ exhausted T cells, doublets were identified and removed using DoubletFinder. We then used differential gene expression to remove clusters that do not contain bona fide CD8⁺ exhausted T cells. We detected subclusters with low expression of CD3E and high expression of AIF1 or CD79A/CD79B in both the human and mouse datasets, which clustered separately from the CD8⁺ exhausted T cells. We removed those subpopulations, as they likely represented monocytes or B cell doublets.

Comparative analysis with the Zheng atlas

To validate the identity of the CD8⁺ exhausted T cell populations in our study, we performed a correlation analysis between our datasets and a tumor infiltrating lymphocyte atlas.¹⁰ We selected their CD8⁺ exhausted T cell clusters (CD8.c11.Tex.PDCD1, CD8.c12.Tex.CXCL13, CD8.c13.Tex.myl12a, and CD8.c14.Tex.TCF7) for comparison with our CD8⁺ human cancer exhausted T cells. Both datasets were normalized using SCTransform, and the analysis was restricted to genes present in both datasets (Supplementary Table 9). We calculated Pearson correlation coefficients between individual cells from each dataset using SCT-normalized gene count matrices. Cell-to-cell correlations were visualized as heatmaps using pheatmap. To assess overall transcriptional similarity, we computed average gene expression profiles for each cell population and performed Pearson correlation testing between the averaged profiles. Statistical significance was determined using cor.test() in R. Results were visualized as a scatter plot, where each point represents one gene and its average expression in both populations.

Cell type compositional analysis

We applied scCODA, a Bayesian compositional model, to identify statistically significant changes in CD8⁺ exhausted T cell subtype proportion between conditions (control and after ICI therapy). The false discovery rate (FDR) value was set to 0.4 to be able to detect subtle, yet biologically relevant, changes, with an automatic reference selection.¹¹

Exhaustion score

Exhaustion module score was calculated using the Seurat AddModuleScore and visualized using violin plots by Seurat VlnPlot with red crossbars indicating average scores. To assess statistically significant differences in exhaustion scores between groups using R, we performed a normality test using Shapiro-Wilk test (sample sizes between 3 and 5000 cells) or Kolmogorov-Smirnov test ( >5000 cells) followed by Wilcoxon rank-sum test. Only non-parametric tests were used due to violations of normality in some groups. Resulting p values were adjusted using the Bonferroni correction to account for multiple comparisons.

Hierarchical clustering

To explore transcriptional similarity between CD8⁺ exhausted T cell clusters, hierarchical clustering was performed based on the average expression of significantly differential expression gene signatures for each CD8⁺ exhausted T cell subtype. Normalized expression values from the SCTransform (SCT) assay were used to compute the average expression for each gene across clusters using Seurat AverageExpression, resulting in a matrix of cluster-wise gene expression. Pearson correlation-based distance matrix (1 - correlation) was computed to capture pairwise cluster similarity. Hierarchical clustering was then performed using the Ward.D2 method, and the resulting dendrogram visualized to assess the relationships among clusters based on their transcriptional profiles.

Results

Single cell RNA sequencing analysis reveals five populations of CD8⁺ exhausted T cells

Leveraging nine human cancer datasets representing eight different cancer histotypes (Figure 1A), we initially used CD3E to capture all T cells followed by CD8A and PDCD1 to identify all populations of CD8⁺ exhausted T cells (Figure 1B). These criteria were specifically chosen to allow us to broadly identify all PDCD1⁺ CD8⁺ T cell populations, potentially representing different exhausted T cell subpopulations (e.g., terminally exhausted, progenitor exhausted, etc.).

Figure 1.

Selection criteria for capturing CD8⁺ exhausted T cells.

(A) Table detailing the tumor types, number of tumors, and GEO ID numbers for the eight different human cancer histotypes used for scRNAseq analysis (nine total datasets). Created in BioRender. Chatterjee, J. (2025)

(B) Schematic representation of the selection criteria used to capture CD8⁺ exhausted T cells. Created in BioRender. Chatterjee, J. (2025)

https://BioRender.com/xxe6q9k

(C) UMAP visualization of scRNAseq data from Zheng and colleagues, color coded by cell type. The dotted red circle denotes the CD8⁺ exhausted T cell population.

(D) UMAP visualization of scRNAseq data from Zheng and colleagues, color coded by PDCD1 gene expression. The dotted red circle denotes the CD8⁺ exhausted T cell population.

(E) Histogram showing the percentage of CD8⁺ exhausted T cells from the eight different human cancer histotypes. Each dot represents one tumor sample.

(F) Cell-to-cell correlation matrix comparing CD8⁺ exhausted T cells from Zheng and colleagues relative to those in the current study. Each cell represents the Pearson correlation coefficient between individual CD8⁺ exhausted T cells from the two studies.

(G) Scatter plot showing the average gene expression levels of CD8⁺ exhausted T cells in the current study relative to those from the Zheng et al. study (r = 0.919, P value < 2.2e-16).

First, to ensure that our selection criteria did not exclude relevant CD8⁺ exhausted T cells and included only authentic CD8⁺ exhausted T cells, we utilized data from the tumor-infiltrating T cell atlas¹⁰ (http://cancer-pku.cn:3838/PanC_T/.) to map PDCD1 expression across their different CD8⁺ T cell populations (Figure 1C). PDCD1 was specifically enriched in the CD8⁺ exhausted T cell subsets, but not in other CD8⁺ T cell populations, supporting the validity of our selection criteria (Figure 1D). Moreover, PDCD1 expression was observed in all CD8⁺ exhausted T cell subtypes and was largely restricted to these populations. These results confirm that no exhausted CD8⁺ T cells were excluded from our analysis.

Second, in addition to PDCD1, we also examined the expression of other exhaustion-associated genes (LAG3, TIGIT, CTLA4, TOX, and ENTPD1) in the Zheng atlas. Similar to PDCD1, these other genes were uniquely enriched in the CD8⁺ exhausted T cell population relative to non-exhausted CD8⁺ T cells, further supporting our classification of this population as exhausted CD8⁺ T cells (Fig. S1A-E).

Third, to evaluate the relationship between our CD8⁺ PD1⁺ exhausted T cells (Figure 1E, Fig. S2A-C) and the CD8⁺ exhausted T cell populations in the Zheng atlas, we performed a correlation analysis. This analysis revealed a positive correlation between the average gene expression profile of the CD8⁺ exhausted T cells selected in our study and those from the Zheng atlas, demonstrating strong transcriptional similarity (r = 0.919, p value < 2.2e-16) (Figure 1F-G).

Taken together, these three complementary analyses confirm that our selection approach is both comprehensive and accurate, ensuring the inclusion of bona fide CD8⁺ exhausted T cell populations across human cancer datasets.

To determine the clustering resolution for our analyzed datasets, we employed the Clustree R package, which visualizes how cells transition between clusters across different resolution parameters. By plotting the hierarchical relationships among cluster assignments at successive resolutions, Clustree enabled detection when cluster identities become stable, examining how clusters split or merge across resolutions, assessing the stability and biological coherence of these transitions. While this approach yielded a numerous subclusters, many of these identified subgroups lacked differentially expressed genes and did not exhibit clear or distinct biological identities. To overcome this problem, a secondary analysis was performed in which subclusters that shared similar gene expression profiles were merged, resulting in five subpopulations (Figure 2A, Fig. S2D) that were not unique to any one cancer and existed in varying numbers across all cancers examined (Figure 2B-D) with significant differentially expressed genes for each cluster (denoted as “C”; Fig. S2E, Supplementary Table 7). Importantly, each CD8⁺ exhausted T cell cluster harbored a distinct set of differentially expressed genes, increased expression of checkpoint receptors,^12–14 cytokines and effector molecules associated with exhaustion,¹⁰ and transcription factors,¹⁵ and different pathway enrichments: C1 (CCL3, CCL4), enriched in chemotaxis pathways and expressing high levels of exhaustion markers (TIGIT, LAG3, TOX) and terminal exhaustion markers (HAVCR2, ENTPD1); C2 (collagen and extracellular matrix-associated genes), enriched in extracellular matrix (ECM) pathways with lower expression of exhaustion markers except for LAG3 and HAVCR2; C3 (IL7R, TCF7, CCR7, SELL), with no discernable pathway enrichment and lower expression of exhaustion markers, but with high CD69 expression; C4 (heat shock protein genes), enrichment in unfolded protein response and heat shock genes with lower expression of exhaustion markers and high CD69 expression; C5 (IFIT and OAS-associated genes), enrichment in interferon pathways and high expression of exhaustion markers and terminal exhaustion markers (Figure 3A–C).

Figure 2.

Five subpopulations of CD8⁺ exhausted T cells are present across the eight different cancer histotypes.

(A) UMAP visualization of scRNAseq data from the eight different human cancer histotypes (corresponding to the datasets in Figure 1A), based on the CD8⁺ exhausted T cell cluster (C1, C2, C3, C4, C5).

(B) UMAP visualization of scRNAseq data from the eight different human cancer histotypes (corresponding to the datasets in Figure 1A), based on the type of cancer (brain tumor, thyroid cancer, lung cancer, pancreatic cancer, ovarian cancer, colorectal cancer, peripheral nerve sheath tumor, skin cancer).

(C) Stacked bar plot representing the percentage of each CD8⁺ exhausted T cell cluster (C1, C2, C3, C4, C5) in the eight different human cancer histotypes (brain tumor, thyroid cancer, lung cancer, pancreatic cancer, ovarian cancer, colorectal cancer, peripheral nerve sheath tumor, skin cancer).

(D) Stacked bar plot showing that the five CD8⁺ exhausted clusters (C1, C2, C3, C4, C5) included cells from all eight analyzed cancer histotypes (brain tumor, thyroid cancer, lung cancer, pancreatic cancer, ovarian cancer, colorectal cancer, peripheral nerve sheath tumor, skin cancer).

Figure 3.

Unique gene signatures and pathway enrichments are associated with each CD8⁺ exhausted T cell subpopulation.

(A) Heat map showing the top differentially expressed genes for the five CD8⁺ exhausted T cell clusters (C1, C2, C3, C4, C5) (corresponding to the datasets in Figure 1A).

(B) Dot plot showing the average expression level and expression percentage of previously reported genes used to identify exhausted T cell genes, including checkpoint receptors, cytokines and effector molecules associated with exhaustion, and transcription factors in the five CD8⁺ exhausted T cell clusters (C1, C2, C3, C4, C5) (corresponding to the datasets in Figure 1A).

(C) GO pathway analysis of the five CD8⁺ exhausted T cell populations. Fold enrichment is shown in the graph, where the color depth represents the FDR. Nucleotide-binding oligomerization domain containing 2, NOD2. Anti-microbial, AM. Extracellular matrix, ECM. Created in BioRender. Chatterjee, J. (2025) https://BioRender.com/d7zy7fy.

Importantly, none of the CD8⁺ exhausted T cell clusters expressed genes associated with an effector CD8⁺ T cell phenotype (CXCR1, CD44, SPN, FAS, KLRG1, RUNX3)^16,17 or OSR2, which augments T cell exhaustion¹⁸ (Fig. S3A). Moreover, genes previously used to classify different subpopulations of CD8⁺ exhausted T cells, such as CD101 and CD69 (terminal CD8⁺ exhausted T cells) or SLAMF6 and CXCR5 (progenitor CD8⁺ exhausted T cells)^19–22 failed to consistently identify the same subpopulations when applied to the five clusters. In this respect, CD101 and CD69 were expressed in clusters C1, and C3/C4, respectively, SLAMF6 had higher expression in cluster C5, and CXCR5 exhibited low expression in all five clusters (Fig. S3B).

To improve the robustness of our analysis, two additional analyses were performed. First, we showed that the population of MKI67⁺ CD8⁺ T cells clusters independently by generating a UMAP with all the CD8⁺ T cells including the five subpopulations of CD8⁺ exhausted T cells (C1, C2, C3, C4, C5), CD8⁺ proliferating (MKI67⁺) T cells, and CD8⁺ non-exhausted T cells. The MKI67-expressing CD8⁺ T cells do not cluster with the five clusters of CD8⁺ exhausted T cells, revealing the clustering of CD8⁺ exhausted T cells was not affected by the exclusion of MKI67⁺ CD8⁺ T cells (Fig. S3C). Second, we compared the expression of exhaustion markers, which includes immune checkpoint receptors, cytokines and effector molecules associated with exhaustion, transcription factors, and exhaustion score in CD8⁺ exhausted T cells with the context of other CD8⁺ T cells, including CD8⁺ proliferating T cells and CD8⁺ non-exhausted T cells. The subpopulations of CD8⁺ exhausted T cells have increased expression of exhaustion markers and higher exhaustion scores compared to other CD8⁺ T cells, including CD8⁺ proliferating T cells, and CD8⁺ non-exhausted T cells (Fig. S3D-F).

CD8⁺ exhausted T cell subpopulations are largely shared between human and mouse cancers

To extend our CD8⁺ exhausted T cell subpopulation analysis to murine experimental systems, we leveraged six scRNAseq datasets from mice harboring four distinct cancer histotypes (brain tumor, optic pathway glioma, Nf1-OPG; skin cancer, melanoma; pancreatic cancer; colorectal cancer) (Figure 4A). We found that CD8⁺ exhausted T cells expressing Cd3e, Cd8a, and Pdcd1 were found in all five mouse cancer datasets (Figure 4B, Fig. S4A-C). Five CD8⁺ exhausted T cells subpopulations were detected in all of the mouse cancers in varying proportions (Figure 4C-F, Fig. S4D, Supplementary Table 8). Using differential gene expression to include heatmaps, dot plots, and pathway analysis (Figure 4G, Fig. S5A-C), we identified four of the five human CD8⁺ exhausted T cell clusters (C1, C2, C3, and C5) in the murine cancers. Interestingly, the human cluster C4 was not detected in the mouse cancers, while cluster C6 was unique to the murine datasets.

Figure 4.

CD8⁺ exhausted T cell subpopulations are conserved across human and mouse cancers.

(A) Table describing the tumor type, number of datasets, and GEO ID numbers of the four different mouse cancer histotypes used for scRNAseq analysis (six total datasets). Brain tumor (optic pathway glioma, OPG); genetically engineered mouse model, GEM; injected tumor cell line model, Inj. Created in BioRender. Chatterjee, J. (2025)

https://BioRender.com/pqnmwfe

(B) Histogram showing the percentage of CD8⁺ exhausted T cells in the four different mouse cancer histotypes. Each dot represents one tumor sample.

(C) UMAP visualization of scRNAseq data from the four different mouse cancer histotypes (corresponding to the datasets in Figure 4A), based on the CD8⁺ exhausted T cell clusters (C1, C2, C3, C5, C6).

(D) UMAP visualization of scRNAseq data the four different mouse cancer histotypes (corresponding to the datasets in Figure 4A), based on the type of cancer (brain tumor, skin cancer, pancreatic cancer, colorectal cancer).

(E) Stacked bar plot representing the percentage of each CD8⁺ exhausted cluster (C1, C2, C3, C5, C6) in the four different mouse cancer histotypes (brain tumor, skin cancer, pancreatic cancer, colorectal cancer).

(F) Stacked bar plot showing that the five CD8⁺ exhausted cluster (C1, C2, C3, C5, C6) included cells from all analyzed cancer histotypes (brain tumor, skin cancer, pancreatic cancer, colorectal cancer).

(G) GO pathway analysis of the five CD8⁺ exhausted T cell populations. Fold enrichment is shown in the graph, where the color depth represents the FDR. Sarcoplasmic reticulum, SR; Type I interferon, IFN. Created in BioRender. Chatterjee, J. (2025) https://BioRender.com/pqnmwfe.

None of the CD8⁺ exhausted T cell subpopulations expressed Osr2 or genes characteristic of effector T cells (Fig. S5D). Similar to the human cancers, genes previously used to classify different subpopulations CD8⁺ exhausted T cells failed to consistently identify the same subpopulations when applied to the five mouse clusters. In this regard, Cd101 and Cd69 were enriched in clusters C3 and C5, respectively, Slamf6 had high expression in cluster C3, and Cxcr5 had low expression in all five murine cancer subclusters (Fig. S5E). Additionally, we generated a UMAP with all the CD8⁺ T cells, including the five subpopulations of CD8⁺ exhausted T cells (C1, C2, C3, C5, C6), CD8⁺ proliferating (MKI67⁺) T cells, and CD8⁺ non-exhausted T cells. Similar to the human datasets, CD8⁺ proliferating T cells did not cluster with CD8⁺ exhausted T cells (Fig. S6A). CD8⁺ exhausted T cells had higher expression of exhaustion markers (immune checkpoint receptors, cytokines and effector molecules associated with exhaustion, transcription factors) and exhaustion scores compared to the CD8⁺ proliferating T cells and CD8⁺ non-exhausted T cells (Fig. S6B-D).

CD8⁺ exhausted T cell clusters exhibit different levels of exhaustion

To assess the exhaustion state of these CD8⁺ exhausted T cell populations, we calculated exhaustion scores based on the expression of immune checkpoint receptors, cytokines and effector molecules associated with exhaustion, and transcription factors in both the human and mouse cancer datasets. Using these exhaustion scores, we ranked the CD8⁺ exhausted T cell subpopulations: In the human cancers, clusters C2 and C1 had the highest exhaustion scores, followed by clusters C4 and C5 with “intermediate” exhaustion scores, and cluster C3 harboring the lowest exhaustion score (Figure 5A). In the mouse cancers, clusters C1 and C6 had the highest exhaustion scores, followed by clusters C2 and C3 (intermediate exhaustion scores), and cluster C5 with the lowest exhaustion score (Figure 5B).

Figure 5.

CD8⁺ exhausted T cells exhibit different exhaustion states in human and mouse cancers.

(A) Violin plots showing the exhaustion scores based on the expression of immune checkpoint receptors, cytokines and effector molecules previously associated with exhaustion, and transcription factors expressed in CD8⁺ exhausted T cells, CD8⁺ proliferating T cells, and CD8⁺ non-exhausted T cells in the human cancers. Exact P values obtained by Wilcoxon rank-sum test with Bonferroni correction are shown.

(B) Violin plot showing the exhaustion scores based on the expression of immune checkpoint receptors, cytokines and effector molecules associated with exhaustion, and transcription factors expressed in CD8⁺ exhausted T cells, CD8⁺ proliferating T cells, and CD8⁺ non-exhausted T cells in the mouse cancers. Exact P values obtained by Wilcoxon rank-sum test with Bonferroni correction are shown.

(C) Cluster similarity matrix based on Pearson correlation of gene expression profiles in five human CD8⁺ exhausted T cell clusters (C1, C2, C3, C4, C5). Each cell represents the Pearson correlation coefficient between a pair of clusters, with values ranging from –1 (blue, strong negative correlation) to + 1 (red, strong positive correlation), as indicated by the color scale. Statistically significant correlations (P values ≤ 0.05) are denoted with asterisks (*). Hierarchical clustering dendrograms above and to the left of the matrix illustrate the relative similarity and grouping of clusters.

(D) Cluster similarity matrix based on Pearson correlation of gene expression profiles in five mouse CD8⁺ exhausted T cell clusters (C1, C2, C3, C5, C6). Each cell represents the Pearson correlation coefficient between a pair of clusters, with values ranging from –1 (blue, strong negative correlation) to + 1 (red, strong positive correlation), as indicated by the color scale. Statistically significant correlations (P values ≤ 0.05) are denoted with asterisks (*). Hierarchical clustering dendrograms above and to the left of the matrix illustrate the relative similarity and grouping of clusters.

(E) Schematic representation of the “exhaustion states” of CD8⁺ exhausted T cell subpopulations in the human and mouse cancers. Created in BioRender. Chatterjee, J. (2025) https://BioRender.com/d1llwj4.

Next, we performed hierarchical clustering using the differentially upregulated genes for each CD8⁺ exhausted T cell subpopulation identified in the human and mouse cancer datasets combined with a calculation of the Pearson correlation coefficients of the different CD8⁺ exhausted T cell subsets to explore the transcriptional similarity of the different clusters. The dendrogram of the human CD8⁺ exhausted T cells subsets showed hierarchical relationships among the clusters, grouping clusters C1 and C2 together with cluster C5 branching separately, and clusters C3 and C4 forming another subgroup. All the clusters showed a significantly negative correlation with each other with the exception of clusters C3 and C4, which exhibited a low positive correlation (r = 0.17), suggesting that the five CD8⁺ exhausted T cell clusters differ substantially in their gene expression patterns (Figure 5C). In the mouse tumors, the dendrogram revealed grouping of clusters C1 and C5, clusters C2 and C6, with cluster C3 branching separately. Notably, most clusters exhibited a negative correlation suggesting highly divergent transcriptional profiles. However, C2 showed a positive correlation with C6 (r = 0.33), suggesting distinct, yet partially overlapping, gene expression patterns. Weak or negligible correlations were observed between clusters C1 and C3 (r = −0.05), as well as between clusters C1 and C5 (r = 0.06), indicating minimal shared transcriptional identity (Figure 5D).

Finally, we then combined the calculated exhaustion score, cytotoxic potential (expression of genes and pathways associated with cytotoxicity), and the expression of progenitor-associated genes (combined with pathway analysis) to order the CD8⁺ exhausted T cell subpopulations from early to terminal exhaustion stages (Figure 5E). We defined clusters C3 (human and mouse) and C4 (human only) as “early exhausted” CD8⁺ T cells, since these two clusters have either lower exhaustion scores or express progenitor markers. We next identified an “intermediate exhausted” state that included cluster C5 (human and mouse), based on the expression of exhaustion markers (LAG3, ENTPD1, TIGIT), albeit at lower levels than the more terminally exhausted subsets, cytotoxic pathways related to interferon responses, and no progenitor markers. Lastly, clusters C1 (human and mouse), C2 (human and mouse), and C6 (mouse only) were classified as “late exhaustion” CD8⁺ T cells, based on the highest exhaustion scores (C1, C2, C6), expression of cytotoxic genes (C1, C2), and an absence of T cell progenitor gene expression.

Anti-PD1 antibody treatment increases C1 CD8⁺ exhausted T cells

To determine the response of these CD8⁺ exhausted T cells subpopulations to ICI treatment, we performed three sets of experiments. First, we utilized scRNAseq datasets from acute and chronic LCMV-infected mice to determine which CD8⁺ exhausted T cell subpopulations were increased following chronic LCMV infection, the setting in which CD8⁺ exhausted T cells were originally identified (Figure 6A). Although there was no equivalent for the human cluster C4 in the mouse cancer datasets, we were able to identify all five human CD8⁺ exhausted T cell subpopulations in the LCMV dataset (C1, C2, C3, C4, C5), as well as an additional cluster (C6). Importantly, the different subpopulations of CD8⁺ exhausted T cells identified in cancer undergo similar dynamic changes after chronic (exhausted state), relative to acute, infection.

Figure 6.

Cluster C1 CD8⁺ exhausted T cells are increased following ICI antibody treatment in chronic viral infection and cancer.

(A) Stacked bar plot representing the percentage of each CD8⁺ exhausted T cell cluster (C1, C2, C3, C4, C5, C6) in mice following acute and chronic LCMV infection.

(B) UMAP visualization of scRNAseq data from mice infected with LCMV (LCMV CTL), based on the CD8⁺ exhausted T cell cluster (C1, C2, C3, C4, C5, C6).

(C) UMAP visualization of scRNAseq data from mice chronically infected with LCMV and treated with anti-PDL1 antibodies (LCMV+anti-PDL1), based on the CD8⁺ exhausted T cell cluster (C1, C2, C3, C4, C5).

(D) Stacked bar plot representing the percentage of each CD8⁺ exhausted T cell cluster (C1, C2, C3, C4, C5, C6) from mice chronically infected with LCMV with or without anti-PD1 antibody treatment.

(E) Stacked bar plot representing the percentage of each CD8⁺ exhausted cluster (C1, C2, C3, C5, C6) in untreated (CTL), anti-PD1 antibody (Nf1-OPG+anti-PD1) treated, and anti-TIGIT antibody (Nf1-OPG+anti-TIGIT) treated, Nf1-OPG mice.

(F) Stacked bar plot representing the percentage of each CD8⁺ exhausted cluster (C1, C2, C3, C4, C5) in patients with colorectal cancer (CRC) prior to PD1 therapy (CTL) and after anti-PD1 (CRC+anti-PD1) antibody therapy.

Second, we analyzed chronic LCMV-infected mice treated with anti-PDL1 antibody therapy to determine which CD8⁺ exhausted T cell populations respond to ICI treatment (Figure 6B-C). Treatment with anti-PDL1 antibodies resulted in a 1.85-fold increase in the proportion of cluster C1 in LCMV-exposed mice (percentage of cells in the C1 cluster as a proportion of the total number of CD8⁺ exhausted T cells; LCMV, 21.42%; LCMV + anti-PDL1, 39.57%) (Supplementary Table 10; Figure 6D).

Third, to compare the changes in CD8⁺ exhausted T cell subpopulations after ICI therapy, we employed scRNAseq data from a genetically engineered mouse low-grade glioma model (Neurofibromatosis type 1 optic pathway glioma, Nf1-OPG) and patients with colorectal cancer (CRC) treated with ICI therapy. In the murine ICI study, Nf1-OPG mice were treated with anti-PD-1 or anti-TIGIT antibodies for 4 weeks, beginning at 12 weeks of age, when low-grade optic gliomas are fully developed in this murine preclinical model.^23–25 Whereas anti-PD-1 antibody treatment reduced tumor proliferation (%Ki67⁺ cells), there was no effect of anti-TIGIT antibody treatment.⁸ Similar to the anti-PDL1 antibody-treated mice following chronic LCMV infection, anti-PD-1 antibody treatment, but not anti-TIGIT antibody treatment, resulted in an increase in the proportion of cluster C1 relative to the control group (Figure 6E). In Nf1-OPG mice, cluster C1 increased by 1.42-fold after anti-PD1 antibody treatment (percentage of cells in the C1 cluster as a proportion of the total number of CD8⁺ exhausted T cells (Supplementary Table 10; CTL, 15.26%; anti-PD1, 21.63%); however, no expansion of cluster C1 was detected following treatment with anti-TIGIT antibodies (0.84-fold change; percentage of C1; CTL, 15.26%, anti-TIGIT, 12.75%).

Fourth, we analyzed scRNAseq datasets from metastatic colorectal cancers following anti-PD1 antibody ICI therapy. In this study, patient tumors were biopsied before and after successful ICI treatment (tumor regression and no recurrence of the cancer at the end of the study). Similar to Nf1-OPG mice treated with anti-PD1 antibodies, cluster C1 increased by 1.9-fold change in the colorectal cancer patients who responded to anti-PD1 ICI therapy (percentage of C1; CRC CTL: 11.68%, CRC + anti-PD1: 22.21%) (Figure 6F, Supplementary Table 10).

Discussion

The identification of CD8⁺ exhausted T cells and their role in tumor immunology has provided new opportunities for cancer therapeutics, culminating in the use of ICI therapies to activate these tumoricidal lymphocytes. Unfortunately, there is currently no uniformly adopted classification scheme to study this critical population of immune cells. Herein, we employed single cell transcriptomics across numerous human and mouse cancers to identify several conserved clusters, one of which was consistently increased by ICI therapy. These findings offer a synthesized conceptual framework for studying populations of cancer-relevant CD8⁺ T lymphocytes and suggest new avenues for future investigation.

First, the lack of a uniform CD8⁺ exhausted T cell population classification scheme makes it challenging to harmonize results from different studies and species relevant to clinical translation. In this regard, CD101 and CD69 are often used to identify terminally exhausted T cells,^26,27 while SLAMF6 and CXCR5 are frequently used to mark progenitor exhausted T cells.^28–30 Applying these genes to the nine human datasets used in this study, we found that CD101 was enriched in cluster C1, whereas clusters C3 and C4 had the highest expression of CD69. Similarly, C3, a cluster expressing progenitor exhausted T cells genes (e.g., TCF7 and IL7R), had low expression of SLAMF6 (highest expression in C5) and CXCR5 (low expression in all subpopulations). Moreover, we observed differences between the expression of these genes in human versus mouse cancers: CD101 was expressed in cluster C1 from human tumors but cluster C3 in the mouse tumors, while CD69 expression was highest in clusters C3 and C4 from human tumors but cluster C5 in the mouse tumors. Similarly, SLAMF6 was enriched in cluster C5 from human tumors but cluster C3 in the murine tumors, whereas CXCR5 exhibited low expression in cancers from both species. For these reasons, we sought to establish an integrated CD8⁺ exhausted T cell subpopulation classification system, such as the one described herein, recognizing that future studies using this framework in both mouse and human cancers will be required to confirm its broad applicability.

Second, we showed that the proportions of exhausted T cell subpopulations are not static but rather undergo dynamic changes in response to ICI therapy. In the setting of both chronic infection in mice and cancer in human and mice, cluster C1 was expanded following successful ICI therapy. In this regard, it is also important to note that prior studies have also attempted to correlate positive responses to ICI with specific T cell signatures. In the B16-OVA murine melanoma cancer model, PD1 blockade increased the frequency of TCF1⁻ TIM-3⁺ terminal exhausted CD8⁺ T cells, which exhibit potent cytotoxic activity, and increased their granzyme and perforin-mediated anti-tumor function.³¹ An increase in cytotoxic activity (Granzyme B, Tissue Necrosis Factor, Interferon gamma) was also seen in mouse glioblastoma models after anti-PD1 antibody treatment.³² In human clinical settings, CD8⁺ exhausted T cells with increased expression of cytotoxic-activity (Perforin, Granzyme B) expand following successful anti-PD1 ICI therapy in breast cancer.³³ Collectively, these findings suggest that one mechanism by which anti-PD1 antibody therapy eradicates tumors could involve an expansion of cluster C1 containing CD8⁺ exhausted T cells with cytotoxic activity.

Third, while we used all cancer datasets available with sufficient numbers of T cells, one limitation of this study is the paucity of human ICI therapy scRNAseq datasets with matched pre-treatment controls. For this reason, additional studies will be required to determine whether refinement of the cluster C1 signature can be used to predict patient outcome in a larger number of cancers, including human datasets with treatment-responsive and non-responsive ICI therapy. Moreover, to assess compositional changes, we performed scCODA analysis in LCMV-infected mice treated with anti-PDL1 antibodies, and both mouse and human cancers post ICI treatment. Cluster C1 was increased in LCMV-infected mice treated with anti-PDL1 antibody therapy and in patients with metastatic colorectal cancer following anti-PD1 antibody ICI therapy.

Fourth, comparing our CD8⁺ exhausted T cell subpopulations to those reported in existing large-scale human and mouse T cell atlases, we identified common features between the clusters identified in our study and those previously described in the literature. In this respect, our manuscript builds upon prior analyses that aim to describe the diversity of CD8⁺ exhausted T cells within the tumor microenvironment. Rather than identifying entirely novel subpopulations, we now provide a unified framework that integrates previously reported CD8⁺ exhausted T cell populations into a single classification scheme (Figure 5). Clusters C1 and C3 closely correspond to the GZMK⁺ and TCF7⁺ exhausted T cells described by Zheng et al.¹⁰ and show strong alignment with terminal exhausted and progenitor exhausted CD8⁺ T cell populations identified by Chu et al., Pritykin et al., and Andreatta et al.^34–36 Tietscher et al. identified a CD8⁺ terminal exhausted population, which is analogous to cluster C1.³⁷ Skelly et al. described a CD8⁺ T cell subset expressing exhaustion markers without cytotoxicity, similar to cluster C2.³⁸ Liu et al. identified a range of CD8⁺ subsets that parallel the findings by Chu et al, including a CD8⁺ exhausted T cell subset (similar to cluster C2, but lacking GZMK expression) and a CD8⁺ terminally exhausted population (similar to cluster C1).³⁹ Additionally, Cheng et al. identified terminally exhausted CD8⁺ T cell populations that align with cluster C1,³⁹ while Sade-Feldman et al. described a CD8⁺ terminally exhausted population expressing cytotoxic genes similar to cluster C1, and a CD8⁺ exhausted population expressing heat shock proteins akin to cluster C4.⁴⁰

Taken together, by integrating data from both human and mouse cancer studies, chronic viral infection in mice, and previously described, but often conflicting subpopulation assignments, reported by others, we provide an integrated framework that consolidates the heterogeneity of exhausted CD8⁺ T cells to allow for a better understanding of their roles across tumor contexts.

Funding Statement

This work was partly funded by a grant from the National Cancer Institute [5R01CA261939] to D.H.G. This manuscript is the result of funding in whole or in part by the National Institutes of Health (NIH). It is subject to the NIH Public Access Policy. Through acceptance of this federal funding, NIH has been given a right to make this manuscript publicly available in PubMed Central upon the Official Date of Publication, as defined by NIH.

Disclosure statement

No potential conflict of interest was reported by the author(s).

Data availability statement

The single cell RNA sequencing datasets generated in this study have been deposited in GEO database under accession code GSE282401. All published datasets were previously deposited in the GEO repository under the accession numbers listed in Supplementary Tables 1 , 2, and 5 (human), Supplementary Tables 3 , 4, and 6 (mouse), and Supplementary Table 9 (Zheng atlas). The code used to generate the figures was deposited on https://github.com/Rui-MU-bids/CD8-exhausted-T-cell-manuscript. All datasets generated in the Gutmann laboratory will be made available upon request to Dr. Gutmann (point of contact: gutmannd@wustl.edu). All other datasets are publicly available from the GEO repository as detailed above.

Supplementary material

Supplemental data for this article can be accessed online at https://doi.org/10.1080/2162402X.2025.2540504.