Glioblastoma-infiltrating CD8⁺ T cells are predominantly a clonally expanded GZMK⁺ effector population

Abstract

Recent clinical trials have highlighted the limited efficacy of T cell-based immunotherapy in patients with glioblastoma (GBM). To better understand the characteristics of tumor-infiltrating lymphocytes (TIL) in GBM, we performed cellular indexing of transcriptomes and epitopes by sequencing (CITE-seq) and single-cell RNA sequencing (scRNA-seq) with paired V(D)J sequencing, respectively, on TIL from two cohorts of patients totaling 15 patients with high grade glioma, including GBM or astrocytoma, IDH mutant, grade 4 (G4A). Analysis of the CD8⁺ TIL landscape reveals an enrichment of clonally expanded GZMK⁺ effector T cells in the tumor compared to matched blood, which was validated at the protein level. Furthermore, integration with other cancer types highlights the lack of a canonically exhausted CD8⁺ T cell population in GBM TIL. These data suggest that GZMK⁺ effector T cells represent an important T cell subset within the GBM microenvironment and which may harbor potential therapeutic implications.

Introduction

Glioblastoma (GBM) is a highly aggressive brain cancer which is the most common primary malignancy of the central nervous system (CNS) in adults (1). Because the prognosis for patients diagnosed with GBM remains poor, there is a critical need for improved treatments. Given the increasing role and success of immunotherapies in many different cancer types, there has been tremendous enthusiasm to explore immune-based approaches to treat GBM. Despite promising results in early phase studies, there currently are no FDA approved immunotherapies for GBM due to lack of efficacy demonstrated in multiple randomized phase 3 clinical trials (2–4). For example, anti-PD-1 treatment, either as monotherapy or in combination, did not improve overall survival in newly diagnosed and recurrent GBM. Similarly, EGFRvIII-specific peptide vaccination also did not significantly improve overall patient survival (5).

Ongoing work is directed at trying to understand the barriers to treatments that are designed to license, or activate, endogenous T cell activity against brain tumor cells. Indeed, it is possible that both tumor intrinsic as well as microenvironmental characteristics of GBM tumors may prevent the immune system from being unleashed in a therapeutically effective manner. Recent work has revealed the significant somatic variant, neoantigen, and cellular intratumoral heterogeneity of GBM, underscoring the difficulty in targeting clonal antigens and specific cell types with disparate transcriptional and functional profiles (6–10). Furthermore, we speculate that this immunologically unfavorable environment represents a particularly severe example of cancer immunoediting in humans (11,12). A myriad of immune deficits have been cataloged including intrinsic immune suppression (13–15) and overexpression of several checkpoint molecules, such as PD-L1 (7,16). At the level of the tumor extrinsic microenvironment, the use of single cell analysis has demonstrated that GBM is largely immunosuppressive in both murine and human settings (17–20).

GBM is often described as a “non-inflamed” tumor because of both the quantitative lack of a robust de novo T cell infiltrate as well as the qualitative dysfunction of T cells that are present within the tumor. Indeed, the T cells in GBM are typically considered to be “exhausted” due to expression of surface receptors such as PD-1, LAG-3, and TIM-3 in human samples and functional hypoactivity in mouse models (21–23). One recent study has investigated the sex-biased difference of T cell exhaustion in patient GBM samples, highlighting the difference in frequency of progenitor exhausted T cells and TOX expression between tumor infiltrating T cells derived from male and female patients (24). However, prior studies describing T cell hypofunctionality in human GBM samples have not identified significant transcriptional exhaustion signatures. For example, Mathewson et al. (17) characterized a subset of T cells that expressed KLRB1 (encoding CD161). While inactivation of CD161 enhanced anti-tumor T cell function, there was a noticeable lack of RNA expression of several canonical exhaustion surface markers, such as TIGIT, LAG3, HAVCR2, and CTLA4, represented by an inhibitory signature. Additionally, a separate study by Abdelfattah et al. (25) performed single cell RNA sequencing (scRNA-seq) on several glioma types, including GBM, and similarly showed low expression of several exhaustion markers, such as LAG3, TIGIT, and CTLA4. Finally, a recent study by Ravi et al. (20) implicating the role of myeloid IL-10 as a driver of T cell exhaustion in GBM observed a very low frequency of T cells that harbored canonical markers of exhaustion, such as LAG3, HAVCR2, CTLA4, and PDCD1. In the aforementioned sex-biased T cell exhaustion study, the total frequencies of both progenitor and terminally exhausted T cells represented a minority of the overall total T cell landscape. Thus, these studies suggest that conventional T cell exhaustion may not be a predominant phenotype of GBM tumor-infiltrating lymphocytes (TIL) and highlight the need for further work to characterize the T cell populations found within brain cancers on a molecular level.

To better understand the T cell compartment within GBM, we sought to characterize T cell phenotypes and cell states using a combined multi-omics approach. To this end, we integrated single cell RNA sequencing (scRNA-seq) with paired antibody capture and single cell V(D)J sequencing on T cells isolated from a cohort of patients diagnosed with malignant gliomas comprised mostly of GBM but several cases of astrocytoma, IDH mutant, grade 4 (G4A) as well. Using several computational methods, we show that the canonical T cell exhaustion signature was weakly expressed in HGG TIL, which further supports prior studies (17,26). Notably, we observe that HGG TIL are enriched for CD8⁺ GZMK⁺ T cells that do not highly express classic cytotoxic markers such as PRF1, GNLY, and GZMB. Moreover, CD8⁺ GZMK⁺ T cells can be further subdivided with respect to NR4A2 expression. With the Xenium in situ platform, we performed sub-cellular resolution spatial transcriptomic analysis and observed not only the presence of GZMK⁺ T cells in GBM, but localization of these T cells with CD163⁺MHCII⁺ myeloid cells and vasculature. Using TCR clonotype data, we find that GZMK⁺ T cells are selectively expanded in TIL but not in matched PBMC. By integrating T cell expression data from a broad range of cancer types, including an independent GBM cohort, we gain further validation supporting the lack of exhausted T cells in HGG as well as the relative increased abundance of GZMK⁺ T cells in HGG and other types of cancer. Taken together, these data suggest that GZMK⁺ T cells may represent an effector subset distinct from canonically exhausted T cells and highlight this population for further study in GBM and other malignancies.

Results

Primary and recurrent high grade gliomas are infiltrated by a diverse repertoire of T cell states with a notable lack of exhausted phenotypes.

To characterize the tumor-infiltrating T cell states and clonotypic diversity in GBM to better understand why T cells in these tumors may not be poised for licensing by checkpoint blockade immunotherapies, we performed cellular indexing of transcriptomes and epitopes by sequencing (CITE-seq) and V(D)J sequencing on an initial cohort of high grade glioma (HGG) patient samples consisting of 9 IDH-WT GBM and 1 IDH-mutant G4A (7 primary GBM, 2 recurrent GBM, 1 recurrent G4A) along with 5 matched peripheral blood mononuclear cell (PBMC) samples (3 matched to primary tumor, 2 matched to recurrent tumor) (Supplementary Table S1). To recognize the 2021 WHO classification of CNS tumor accurately (27), we refer to “Glioblastoma, IDH-wildtype” as “GBM” and “Astrocytoma, IDH-mutant, grade 4” as “G4A” when referencing a specific diagnosis separate from each other but use the term high grade gliomas (HGG) when discussing both diagnoses as a single cohort. The use of CITE-seq and V(D)J sequencing allowed us to intersect transcriptomic, clonotypic, and proteomic information on a single cell basis. Specifically, T cells were enriched based on CD3 expression using fluorescence activated cell sorting, stained with TotalSeq antibodies, and subsequently characterized by CITE-seq (Fig. 1A). Unsupervised clustering and uniform manifold approximation and projection (UMAP) analysis was performed on 32,320 T cells, and T cell states were identified based on the expression of both gene and protein expression (Figs. 1B and 1C, Supplementary Table S2). Detected CD4⁺ T cell states included: naїve (T_N), central memory (T_CM), effector memory (T_EM), activated, GZMK⁺ effector (T_Eff), effector memory re-expressing CD45RA T cells (T_EMRA), and regulatory T cells (T_reg); detected CD8⁺ T cell states included: naїve (T_N), NR4A2^lo/hi GZMK⁺ effector (T_Eff), cytotoxic resident memory (T_RM), effector memory re-expressing CD45RA (T_EMRA), and MAIT-like T cells; other detected T cell states included: proliferating and stress signature T cells (Figs. 1C and 1D, Supplementary Table S2, Supplementary Fig. S1, Supplementary File S1: Sheet 1). In some cases, proteins were both highly expressed and correlated with RNA expression (e.g. CD8, Supplementary Fig. S2). In other cases, proteins were highly expressed but did not correlate with RNA expression (e.g. CD4 and CD45RA). Finally, several proteins were only moderately expressed and did not correlate well with RNA (e.g. CD27). These data highlight the utility of CITE-seq, especially with protein markers such as CD4 and CD45RA, for purposes of cell type identification. TIL derived from HGG mainly consisted of effector T cells, such as CD4⁺ T_EM, CD4⁺ and CD8⁺ GZMK⁺ T cells, cytotoxic T_RM, and T_EMRA, while T cells derived from matched PBMC samples mainly belonged to either a T_N or T_EMRA cell state (Figs. 1E and 1F).

Figure 1: Single cell preparation and sequencing shows a diverse TIL landscape in HGG.

(A) Illustration of single cell preparation of primary and recurrent HGG samples, consisting of both GBM and G4A, and subsequent isolation and analysis of CD3⁺ T cells. Created with BioRender.com. (B) UMAP visualization of 32,320 T cells in the first cohort of HGG and PBMC samples with respective cell states labeled. (C) Dot plot of RNA expression of select gene markers. (D) Violin plots of protein expression of select protein markers. (E) UMAP visualizations of T cells highlighted by sample type. (F) Bar plots of distribution of T cell states separated by sample type. (G, H) Violin plots of expression of select protein markers of activation and exhaustion. (I, J) UMAP visualization and violin plot of RNA Tirosh exhaustion score expression with each T cell state as colored in Fig. 1B.

To assess the phenotype of HGG TIL, we analyzed several cell surface protein co-stimulatory and co-inhibitory markers. High levels of CD69 and CD278/ICOS were expressed by several activated and cytotoxic T cell clusters (Fig. 1G). However, several common co-stimulatory surface receptors, including CD137/4–1BB, CD357/GITR, and CD134/OX40, were not well expressed. Finally, CD244/2B4, an NK cell receptor also expressed by T cells, was highly expressed on several CD8⁺ T cell clusters, including the GZMK⁺, cytotoxic T_RM, T_EMRA, and MAIT clusters. While PD-1 was expressed by several clusters, there were notably low levels of other exhaustion-associated markers, such as TIGIT, LAG-3, BTLA, and CTLA-4 (Fig. 1H). To determine if these PD-1 expressing clusters were represented by an exhausted transcriptional state, we applied a 28 gene expression-based exhaustion signature derived from melanoma TIL, as reported by Tirosh et al. (28) to our HGG TIL cohort. We scored each cell for expression of this signature (Materials and Methods) and observed a marginally higher expression in the CD8⁺ NR4A2^hi GZMK⁺ T_Eff (C10) and T_regs (C14) clusters (Fig. 1I and 1J), while the CD8⁺ NR4A2^lo GZMK⁺ T_Eff (C9) exhibited lower expression of this signature. These scores were similar between PBMC and TIL samples for each T cell cluster (Supplementary Fig. S3A). Similar findings have been reported in other GBM single cell data sets generated using different technologies, such as SMART-Seq2 and 10X 3’ sequencing, in which GBM-infiltrating CD8⁺ T cells exhibited a weak exhaustion-associated co-inhibitory receptor expression signature (consisting of PDCD1, CTLA4, HAVCR2, LAG3, and TIGIT) (17,25). When using a comparable inhibitory signature as these other studies (PDCD1, CTLA4, HAVCR2, LAG3, TIGIT, and, BTLA) (17), we found relatively high expression of the inhibitory signature in several clusters, most notably the T_regs (C14) cluster (Supplementary Figs. S3B–S3D). Overall, the transcriptional and paired protein expression of these inhibitory receptors suggest a lack of a canonical exhausted T cell population in HGG TIL.

Most CD8⁺ T cells within GBM TIL are GZMK⁺ and may develop separately from the T_EMRA cell state

Having observed the lack of a significant transcriptional exhaustion signature, we further characterized the transcriptional profile of HGG CD8⁺ T cells. A total of 10,749 CD8⁺ T cells were subsequently subsetted, reanalyzed, and visualized via UMAP (Fig. 2A, Supplementary File S1: Sheet 2, Supplementary Fig. S4). Strikingly, HGG CD8⁺ TIL were comprised more of GZMK⁺ T cells, either NR4A2^lo or NR4A2^hi (C9 and C10, respectively), than cytotoxic CD8⁺ T_RM (C11) and T_EMRA (C12). In PBMC, GZMK⁺ T_Eff (C9) were also present but T_N, T_EMRA, and MAIT-like cell states were the most prevalent cell types (Figs. 2B and 2C). Similar to GZMK⁺ T cells observed in aging mice, as published by Mogilenko et al. (29), CD8⁺ GZMK⁺ T cells expressed high protein levels of CD45RO, CD27, CD28, and PD-1 and low protein levels of IL7R (Fig. 2D). Transcriptionally, these cells expressed moderate levels of TOX and EOMES and low levels of TCF7, also previously observed by Mogilenko et al. in GZMK⁺ T cell populations derived from aged murine models (Supplementary Fig. S4). Notably, these T cells expressed high levels of GZMK and GZMH but did not express high levels of classic cytotoxic markers such as GZMB, IFNG, and PRF1. Differential expression of NR4A2 within the GZMK⁺ T cell population was discovered through unsupervised clustering (Materials and Methods) and notably has been previously characterized in TIL from non-small cell lung cancer samples (30). NR4A1 and NR4A3 exhibited low levels of expression and was not differentially expressed within the GZMK⁺ T cell population (Supplementary Figs. S5A–C). To validate whether GZMK is upregulated in TIL compared to PBMC on the protein level, we performed flow cytometry on prospectively collected GBM patient samples consisting of both TIL and PBMC (Figs. 2E and 2F). Gating on MAIT TCR⁻ CD45RO⁺ CD8⁺ T cells (Supplementary Fig. S6), as MAIT T cells express both GZMK and GZMB, we found that GZMK⁺GZMB⁻ and GZMK⁺GZMB⁺ T cell populations were enriched in TIL compared to PBMC (20.900 +/− 4.763% vs. 7.212 +/− 5.749%, 49.167 +/− 11.133% vs 18.688 +/− 12.760%, respectively, p ≤ 0.05, Mann-Whitney U-test). Meanwhile, GZMK⁻GZMB⁺ T cells were enriched in PBMC compared to TIL (59.500 +/− 21.318% vs. 19.167 +/− 7.920%, p ≤ 0.05, Mann-Whitney U-test). We did not observe GZMK⁺GZMB⁺MAIT⁻ T cells in our single cell data, highlighting the limitations of single cell RNA sequencing, which does not necessarily represent the true protein-level phenotype of analyzed cells. A similar observation has been made in patients with rheumatoid arthritis, in which two major CD8⁺ T cell subsets were observed at the transcript level, one with high GZMK⁺ expression and the other high GZMB⁺ expression, while three populations were observed at the protein level (GZMK⁺GZMB⁻, GZMK⁺GZMB⁺, and GZMK⁻GZMB⁺) (31).

Figure 2: GZMK⁺ effector T cells represent a significant proportion of CD8⁺ TIL and a separate development branch from T_EMRA.

(A) UMAP visualization of 10,749 CD8⁺ T cells in the first cohort of HGG tumor and PBMC samples with respective cell states labeled. (B) UMAP visualizations of T cells highlighted by sample type. (C) Bar plots of distribution of T cell states separated by sample type. (D) Violin plots of expression of protein markers selected from Mogilenko et al. (29). (E) Representative flow cytometry plots and (F) cumulative bar plot of frequency of GZMK and GZMB expression gated on CD45RO⁺CD3⁺CD8⁺MAIT⁻ cells from GBM tumor (n=3) and PBMC (n=5) samples. (G) Trajectory inference analysis of CD8⁺ T cells colored by T cell state, branch state, and pseudotime. (H) Stacked bar plot of relative frequency of occupied branch state per cell type. (I) Heatmap of gene expression dynamics along pseudotime progression. Bars in (F) represent mean +/− SD. Significance calculated using Mann-Whitney U-test. *p ≤ 0.05.

To gain a deeper understanding of GZMK⁺ T cell development in relation to other effector subsets, we performed pseudotemporal trajectory analysis using monocle2 on select CD8⁺ T cell clusters (Fig. 2G–2I) (32). This analysis revealed several cell states and branch points beginning with T_N (branch state 5), then developing into NR4A2^lo GZMK⁺ T_Eff and cytotoxic T_RM, both of which were evenly distributed across several branches. From these T cell states, CD8⁺ T cells split into either T_EMRA (branch state 4) or NR4A2^hi GZMK⁺ T cells (branch state 1) (Figs. 2G and 2H). These results suggest that NR4A2^lo GZMK⁺ and cytotoxic T_RM may represent a population from which terminally differentiated T cells may develop. After hierarchically clustering the top 200 differentially expressed genes plotted along the pseudotime axis, we identified several distinct gene signatures. First, genes related to the naїve T cell state, including CCR7 and SELL, were highly expressed at the beginning of the trajectory. In contrast, classic cytotoxic genes, such as GNLY, PRF1, and GZMB, were highly expressed towards the middle of the trajectory, while GZMK related genes, such as CCL4 and CCL5, were highly expressed together at the end. Genes highly expressed in the NR4A2^hi GZMK⁺ T_eff subset, such as JUNB and FOS which are associated with TCR activation, clustered separately and were highly expressed towards the end of the trajectory (Fig. 2I, Supplementary File S1: Sheet 2). These results, in conjunction with the GZMK protein validation data, suggest that NR4A2^lo GZMK⁺ T_Eff cells may arise from classic cytotoxic T cells (which express exclusively GZMB), transition through the GZMK⁺GZMB⁺ cell state, and terminally differentiate into NR4A2^hi GZMK⁺ T_Eff.

GZMK⁺ T cells localize near vasculature and CD163⁺ MHCII⁺ myeloid cells

Interested in understanding the spatial localization of GZMK⁺ T cells within the GBM tumor microenvironment, we performed sub-cellular resolution spatial transcriptomic analysis using the 10X Xenium in situ platform on one patient GBM tumor sample (Supplementary Table S1). Unsupervised clustering and UMAP analysis were performed on 329,254 detected cells, and cell type identification was performed using key marker genes and differentially expressed genes (Fig. 3A, Supplementary Fig. S7A, Supplementary File S1: Sheet 3). In this sample, 1,093 T cells were identified, with one cluster of 625 T cells expressing higher levels of GZMK (labeled GZMK^hi) and another cluster of 468 T cells expressing lower levels (labeled GZMK^lo). With the detection of concentrated GZMK transcripts in GZMK^hi T cells (Supplementary Fig. S7B), we showed through a third modality, which used probe-based detection methods, expression of GZMK in GBM TIL. Along with increased expression of GZMK, GZMK^hi T cells also expressed higher levels of TOX and EOMES compared to GZMK^lo T cells (Supplementary Fig. S7C, p ≤ 0.001, Mann-Whitney U-test), supporting our previous scRNA-seq observations. Neighborhood enrichment analysis was performed to identify spatial localization patterns with two major spatial categories emerging following hierarchical clustering: cell states consisting of, or localized near, tumor cells and those of, or near, vasculature (Fig. 3B, Supplementary File S2). GZMK^hi and GZMK^lo T cells were binned within the second category, with high enrichment scores with vasculature-related cell types and MHC.Myeloid.2 cells, myeloid cells which differentially expressed MHCII-related genes and CD163 (Supplementary Fig. S7A). Notably, GZMK^hi T cells localized with these myeloid cells more closely than did GZMK^lo T cells. Given CD163 expressing myeloid cells have been previously shown to drive T cell dysfunction (20,33,34), these data suggest that CD163 expressing myeloid cells may be a driver of either GZMK expression or initiating dysfunctionality of such cells. Meanwhile, GZMK^lo T cells were more closely associated with vasculature, specifically endothelial cells and VLMC.2 cells (Supplementary File S2). Through co-occurrence analysis, we orthogonally analyzed the top cell states which localized near GZMK^hi and GZMK^lo T cells (Fig. 3C, Supplementary Fig. S7D). In this analysis, the cell state with the highest probability of being observed given the presence of either GZMK^hi or GZMK^lo T cells was the GZMK^hi T cell state, suggesting that GZMK^hi T cells localize with all T cells. The next most probable cell states for GZMK^hi T cells included GZMK^lo T cells, VLMC.1 cells, and MHCII.Myeloid.2 cells (Fig. 3C). Similarly, the next most probable cell states for GZMK^lo T cells included VLMC.1 cells, GZMK^lo T cells, MHCII.myeloid.2 cells, and endothelial and VLMC.2 cells (Supplementary Fig. S7D). These two analyses showed that both GZMK^hi and GZMK^lo T cells aggregated with each other as well as MHC.myeloid.2 cells. However, while GZMK^lo T cells localized more with VLMC.2 and endothelial cells, GZMK^hi T cells specifically localized with VLMC.1 cells, distinguished by markers such as SULF1, SLIT3, and ADAMTS12. Notably, while both GZMK^hi and GZMK^lo T cells were detected throughout the sample (Supplementary Fig. S7E), neither were significantly localized near one specific tumor state. Visualization of GZMK^hi localization near vasculature was demonstrated by focusing on specific, representative regions within the Xenium spatial image (Figs. 3D and 3E). In these fields of view, we found that regions of dense GZMK^hi T cells (overlaid in blue), were limited to regions outside of the vasculature (overlaid in green), characterized by presence of VLMC and endothelial cells, interspersed with MHCII.Myeloid.2 cells. Similar findings have been made in ovarian cancer, in which GZMK⁺ T cells were enriched in immune excluded tumors, and generally localized to the stromal region (35). As a result, overall, these results highlight the restriction of T cells near vasculature. In addition, GZMK^hi T cells have been observed to co-localize with CD163⁺ myeloid cells more so than GZMK^lo T cells, which may play a role in regulating the function and localization of these T cells.

Figure 3: GZMK^hi T cells localized near vasculature regions and CD163⁺ MHCII⁺ myeloid cells.

(A) UMAP visualization of 329,254 cells detected in the GBM tumor Xenium spatial sample. (B) Neighborhood enrichment analysis of cell types to quantify spatial proximity between all cell type pairs. (C) Co-occurrence plot highlighting the conditional probability of observing a cell state given the presence of nearby GZMK^hi T cells. (D) Image of GBM tumor Xenium spatial sample. (E) Representative fields of view in (D) highlighting the localization pattern of GZMK^hi T cells in relation to vasculature and other cell types. Insets represent respective fields of view with only select cell types (Endothelial cells, VLMC.1, VLMC.2, GZMK^hi T cell, GZMK^lo T cell, MHCII.Myeloid.2).

TIL are enriched for GZMK⁺ T cells in both primary and secondary brain cancers

To further validate our findings and explore whether GZMK⁺ T cells are enriched in other brain malignancies, we used scRNA-seq to characterize T cells isolated from a second cohort of malignant brain tumor patient samples consisting of 4 GBM samples (4 primary GBM), 2 G4A samples (1 primary and 1 recurrent), and 5 brain metastases of variable histologies (Supplementary Table S1). These samples were integrated with the scRNA-seq data from the initial cohort of GBM samples, and T cell states derived from the initial cohort were applied to the integrated data set (Fig. 4A, Supplementary File S1: Sheet 4). Notably, several new cell states were observed, specifically exhausted CD4⁺ T cells which expressed high levels of PDCD1, LAG3, HAVCR2, and EOMES (C12), while other cell states, such as the distinction between NR4A2^lo/hi GZMK⁺ CD8⁺ T cells, were only resolved after further subsetting (Figs. 4B and 4C). These exhausted CD4⁺ T cells were mainly derived from BrMet027, a brain metastasis originating from breast cancer. Cells in this cluster scored highly for both the Tirosh RNA exhaustion signature and the RNA inhibitory signature (Figs. 4D and 4E, Supplementary Figs. S8A and S8B). These findings supported our previous observations that while GZMK⁺ T cells exhibited a low expression of exhaustion and inhibitory genes, they are not consistent with classically exhausted T cells according to conventional transcriptional profiles of exhaustion.

Figure 4: BrMet samples show exhaustion signature and enrichment of GZMK⁺ effector T cells.

(A) UMAP visualization of 51,400 T cells in both the first and second cohorts of HGG, BrMet, and PBMC samples with respective cell states labeled. (B) UMAP visualizations of T cells with sample type highlighted. (C) Bar plots of distribution of T cell states separated by sample type. (D, E) UMAP visualization and violin plot of RNA exhaustion score expression with each cell state colored as in Fig. 3A. (F) UMAP visualization of 18,650 CD8⁺ T cells in both the first and second cohorts of HGG, BrMet, and PBMC samples with respective cell states labeled. (G) Bar plots of distribution of T cell states separated by sample type.

We observed similar proportions of NR4A2^lo/hi GZMK⁺ T cells with the addition of these new HGG samples compared to the initial scRNA-seq data set (16.5% vs 15.0% of total T cells, respectively), further supporting the presence of GZMK⁺ T cells in HGG. Interestingly, BrMet samples also consisted of a similarly high proportion of GZMK⁺ T cells (13.7%) suggesting that the presence of GZMK⁺ T cells is not restricted to HGG but can be found in other brain malignancies (Fig. 4B and 4C). By further analyzing the CD8⁺ T cells, we resolved NR4A2^hi and NR4A2^lo GZMK⁺ T cells and highlighted the enrichment of GZMK⁺ T cells in BrMet samples with GZMK⁺ T cells occupying greater than 30% of the CD8⁺ T cell landscape (Figs. 4F and 4G, Supplementary File S1: Sheet 5). With the additional samples, we also were interested in potential differences in CD8⁺ TIL derived from IDH-WT GBM compared to those from IDH-mut G4A samples. While the frequencies of NR4A2^lo/hi GZMK⁺ T cells were higher in IDH-mut G4A samples compared to those from IDH-WT GBM samples, there were no notable differences in exhaustion scores nor transcriptional profiles (Supplementary Figs. S9A–S9C, Supplementary File S1: Sheet 6). Investigating the expression of GZMK in a separate data set with both CD8⁺ TIL derived from IDH-WT GBM and IDH-mut G4A SMART-Seq2 samples (17), we also found no significant difference in GZMK expression between these two TIL populations (Supplementary Fig. S9D). Similarly, we were interested in understanding the differences between CD8⁺ GZMK⁺ TIL derived from primary GBM compared to those from recurrent GBM. While the number of recurrent GBM samples in our data set was limited, a recent study by Lee et al. (36) analyzed TIL derived from 14 primary, 12 recurrent, and 14 anti-PD-1 treated recurrent GBM samples. Reanalyzing their scRNA-seq data set with our marker genes (Supplementary Table S2), we found similar T cell states which included progenitor-like, CD8⁺ NR4A2^lo and NR4A2^hi GZMK⁺, CD8⁺ cytotoxic T_eff, T_FH-like, T_GD-like, and T_reg cell states (Supplementary Figs. S10A & S10B, Supplementary File S1: Sheet 7). The relative frequency of CD8⁺ NR4A2^lo GZMK⁺ T cells was higher in primary GBM compared to recurrent GBM among all T cells and specifically within CD8⁺ T cells (Supplementary Figs. S10C & S10D). However, the transcriptional profiles of CD8⁺ GZMK⁺ T cells did not notably differ based upon recurrence status (Supplementary File S1: Sheet 8). These results suggest that while the frequencies of CD8⁺ GZMK⁺ TIL may vary depending on IDH and recurrence status, their transcriptional profiles do not vary among all conditions.

Given the overall high prevalence of GZMK⁺ T cells in HGG and brain metastases, we next aimed to define a gene signature specific to CD8⁺ GZMK⁺ T cells. To this end, we identified the top 25 differentially expressed genes (DEGs) by NR4A2^lo/hi GZMK⁺ T cells and used ToppGene (37) to perform a functional enrichment analysis (Supplementary Fig. S11, Supplementary File S3: Sheet 1). Associated biological pathways included “MHC class II protein complex assembly”, with enrichment of genes such as HLA-DRA, and “Lymphocyte mediated immunity”, with enrichment of genes such as CD27 (Supplementary File S3: Sheet 2). Given HLA-DRA expression is associated with T cell activation, these pathways suggest that GZMK⁺ T cells were activated, and antigen experienced. While several cell surface receptor genes were enriched in these T cells, they were not uniquely expressed and thus individually do not provide a sensitive marker for sorting. Regardless, these data support a high frequency of GZMK⁺ T cells within GBM and BrMet CD8⁺ TIL, and that they are enriched for genes associated with T cell activation and prior antigen exposure.

GZMK⁺ TIL are present in a broad range of cancer types

Given that GZMK⁺ T cells comprise a large percentage of CD8⁺ TIL, not just in GBM but also in heterogeneous BrMets, we explored whether GZMK⁺ T cells were present in cancers outside of the central nervous system. To this end, we integrated CD8⁺ T cells from our combined HGG and BrMet data set with additional data sets comprised of CD8⁺ T cells from 1) a pan-cancer TIL data set (38), 2) a melanoma data set (39), 3) an independent GBM data set (17), and 4) a pancreatic ductal adenocarcinoma (PDAC) data set (40) (Fig. 5A, Supplementary Fig. S12A, Supplementary File S1: Sheet 9) for a total of 20 different cancer types. These T cell states were annotated according to our gene signature and supported by original annotations for each respective study (Supplementary Fig. S12B). Notably, we found that GZMK⁺ TIL, as defined by our gene signature and supported by original annotations, were present in several tumor types and similarly separated into two distinct clusters, NR4A2^hi GZMK⁺ and NR4A2^lo GZMK⁺ (Figs. 5A and 5B). The NR4A2^hi GZMK⁺ T_Eff highly expressed several genes associated with TCR activation, such as CD74, CRTAM, and JUNB. Meanwhile, NR4A2^lo GZMK⁺ T cells did not express these particular genes, suggesting these T cells were not recently activated by antigen stimulation (Supplementary File S1: Sheet 9). A similar separation of GZMK⁺ T cells into these subsets has been observed in non-small cell lung cancer (NSCLC) patient samples, supporting our observations that NR4A2 expression may delineate different activation profiles (30). Compared to our data set, the independent GBM data set (17) had higher frequencies of NR4A2^lo GZMK⁺ TIL and similar levels of NR4A2^hi GZMK⁺ TIL, suggesting that our findings were generalizable across cohorts, processing protocols, and analysis pipelines. By comparing the frequency of GZMK⁺ T cells across cancer types, we found that HGG tumors harbored some of the highest frequencies of GZMK⁺ T cells, regardless of NR4A2 expression. As an exhausted T cell population was identified, using both canonical gene markers and original annotations (Supplementary Fig. S12B), and clustered separately from GZMK⁺ T cells, we were interested in comparing the transcriptional levels of exhaustion between exhausted T cells and GZMK⁺ T cells. Consistent with our previous observations, GZMK⁺ T cells (either NR4A2^hi or NR4A2^lo) (C3 and C4) expressed lower levels of the RNA inhibitory and Tirosh exhaustion signatures than did exhausted T cells (C8) (p ≤ 2.2e⁻¹⁶, Mann-Whitney U-test) (Supplementary Figs. S12C & S12D). In addition, the GBM/HGG data sets had some of the lowest frequencies of exhausted T cells of the 20 cancer types analyzed (7.67% and 7.93% in our data set and the published GBM data set, respectively) (Fig. 5C). Thirteen cancer types (red bars in Fig. 5C), including our BrMet samples, had significantly higher frequencies of exhausted TIL compared to our data set, with melanoma TIL having the highest frequency (27.6%, p ≤ 0.05, Fisher’s exact test). Only two cancer types (blue bars in Fig. 5C), colorectal cancer (CRC) and pancreatic ductal adenocarcinoma (PDAC), had significantly lower frequencies of exhausted TIL than did our data set. Within the exhausted T cell population, GBM/HGG derived TIL had the lowest expression of both the RNA inhibitory and Tirosh exhaustion signatures (Supplementary Fig. S12E). These levels were statistically significantly lower when compared to all other tumor types (p ≤ 2.2e⁻¹⁶, Mann-Whitney U-test) (Supplementary Fig. S12F). This observation supported and extended, by including additional cancer types, the recent conclusion made by Naulaerts et al. that when compared to tumor types with high levels of canonical exhaustion, such as melanoma, GBM has few T cells that express an exhaustion signature (26).

Figure 5: GZMK⁺ T cells are enriched in the many cancer types and cluster separately from exhausted T cells.

(A) UMAP visualization of CD8⁺ T cells derived from pan-cancer integration with respective T cell states labeled. (B) Stacked bar plot of relative frequency of NR4A2^hi GZMK⁺ and NR4A2^lo GZMK⁺ T cells among cancer types. (C) Bar plot of relative frequency of exhausted T cells among cancer types. The dotted lines represent the frequency of the respective T cell state in this study’s data set (labeled “Wang_HGG”). (D) Trajectory inference analysis of CD8⁺ T cells colored by T cell state, branch state, and pseudotime. (E) Stacked bar plot of relative frequency of occupied branch state per cell type. (F) Heatmap of genes differentially expressed at branch point 1. Genes upregulated on the right are genes upregulated in cells that follow the pseudotime progression to exhausted T cells while those on the left are upregulated in cells that follow the pseudotime progression to T_EMRA. In (C), red bars represent cancer types with exhausted T cell frequencies that are statistically significantly increased compared to that in this study’s data set. Blue bars represent cancer types with exhausted T cell frequencies that are statistically significantly decreased compared to that in this study’s data set. Significance calculated using Fisher’s exact test. Red or blue bars indicate p ≤ 0.05. Ind_GBM = Independent GBM data set (17). In (F), genes of interest are colored according to associated branch state (red represents association with the T_EMRA state while blue represents association with the exhausted state).

Finally, to understand how GZMK⁺ T cells may develop with respect to exhausted T cells, we performed pseudotemporal trajectory analysis on the integrated tumor data set. From this analysis, we observed two major terminal states branching from naїve T cells: the T_EMRA cell state (branch state 1) and the exhausted cell state (branch state 6) (Figs. 5D and 5E). While NR4A2^lo GZMK⁺ T cells (C3) were more associated with the terminally differentiated cell state, NR4A2^hi GZMK⁺ T cells (C4) were evenly split among the T_N, T_EMRA, and exhausted states. Consistent with previous observations, NR4A2^hi GZMK⁺ T cells may represent a branch point from which exhausted T cells arise (30). These results differ slightly from our previous pseudotemporal trajectory analysis on isolated HGG CD8⁺ T cells suggesting NR4A2⁺ GZMK⁺ T cells as a terminal state, likely due to the current presence of an additional terminal differentiated state (i.e. canonical exhaustion). In addition, as expected, cytotoxic T_RM (C5) were mostly associated with the terminal exhausted T cell state. Next, we investigated the genes that were differentially expressed at the initial branch point which split T_EMRA and exhausted T cells (Fig. 5F). As expected, several genes related to exhaustion were only upregulated on the terminal end of the right branch (“towards exhaustion” in Fig. 5F), such as TOX, LAG3, and ENTPD1. Meanwhile, genes such as LAIR2, KLRG1, and CX3CR1 were upregulated on the terminal end of the left branch (“towards T_EMRA” in Fig. 5F), which was associated with the T_EMRA cell state. GZMK was more highly expressed at the intermediate trajectory on the T_EMRA branch, while genes such as NR4A2 and NR4A3 were more highly expressed at the intermediate trajectory on the exhausted branch. The results from these multiple pseudotemporal trajectory analyses suggest NR4A2^lo GZMK⁺ T cells may represent an intermediary cell state that can develop into a T_EMRA population in the absence of antigen stimulation but alternatively, can differentiate into an exhausted T cell population via a NR4A2^hi GZMK⁺ T cell state with persistent antigen stimulation.

GZMK⁺ T cells within the tumor microenvironment are clonally expanded

Several prior studies have demonstrated that subsets of T cells within GBM are clonally expanded (10,17). We therefore investigated the relationship between GZMK expression and both clonotype diversity and expansion using single cell V(D)J sequencing. Of the 18,650 CD8⁺ T cells sequenced, 11,360 T cells (60.9%) contained paired αβ TCRs (Fig. 6A). T cell expansion states were defined as follows: TCRs with a count of 1 were labeled as singlets, with a count of 2 as doublets, with a count greater than 2 but less than or equal to 50 as expanded, and with a count greater than 50 as hyperexpanded. Among all T cell states, as expected, T_N (C0) were mostly single clones while around 50% of NR4A2^lo GZMK⁺ (C9), NR4A2^hi GZMK⁺ (C20), and cytotoxic T_RM (C10) clusters and up to 75% of the T_EMRA cluster (C12) were either expanded or hyperexpanded (Fig. 6B). When investigating the associated T cell states of the top 50 clonally expanded TCRs for each sample type, 2 of the top 3 expanded TCRs in both PBMC and HGG samples were invariant MAIT TCRs. Of the total number of cells associated with these clonally expanded TCRs, 34.1% and 11.9% were associated with the MAIT T cell state in PBMC and HGG, respectively. In PBMC, the T_EMRA state was the most frequent non-MAIT T cell state associated with these expanded TCRs (35.1%). NR4A2^lo GZMK⁺, NR4A2^hi GZMK⁺, and cytotoxic T_RM only occupied 9.80%, 11.6%, and 8.97% of the total expanded T cell landscape, respectively. In contrast, within HGG TIL, GZMK⁺ T cells comprised the majority of clonally expanded T cells with a relatively equal proportion of both NR4A2^lo and NR4A2^hi GZMK T cells (24.4% and 29.1%, respectively), and were significantly enriched compared to respective populations in PBMC (9.80% and 11.6%, respectively) (p ≤ 0.05, Fisher’s exact test). Meanwhile, cytotoxic T_RM and T_EMRA occupied only 9.86% and 20.4% of the clonally expanded T cell landscape in HGG samples. Finally, in BrMet samples, cytotoxic T_RM occupied the greatest proportion of clonally expanded T cells (29.2%) followed by NR4A2^lo and NR4A2^hi GZMK T cells (18.2% and 28.4%, respectively) (Fig. 6C). These results suggested that more GZMK⁺ T cells are associated with the most expanded TCRs compared with any other T cell subsets within the brain tumor microenvironment. Moreover, these enriched clonotypes were shared by both NR4A2^hi and NR4A2^lo GZMK⁺ T cells, supporting our hypothesis that NR4A2^hi GZMK⁺ T cells may develop from NR4A2^lo GZMK⁺ T cells. Interestingly, between tumor types, GZMK⁺ T cells occupied a greater proportion of the total T cell landscape in HGG compared to BrMet, with cytotoxic T_RM more abundant in the latter.

Figure 6: Expanded GZMK⁺ T cells are enriched in HGG TIL and are specific to the tumor site, and not matched peripheral blood.

(A) UMAP visualization of CD8⁺ T cells from Fig. 3E with respective TCR clonotype enrichment states labeled. (B) Stacked bar plot of distribution of TCR clonotype enrichment states by T cell state. (C) Distributions of the top 50 expanded TCR clonotypes colored by respective T cell states from Fig. 3E and separated by sample type. (D) Representative double sided bar plots representing the absolute count of the top 10 CD8⁺ NR4A2^lo/hi GZMK⁺ T_Eff (C9 & C20) TCRs from matched PBMC (top) or TIL (bottom) samples within the total population of CD8⁺ T cells in paired samples. T cell states associated with each TCR are colored as in Fig. 3E. (E) Heatmap of overlap coefficient for quantification of CD8⁺ T cell clonotype overlap among all five paired PBMC and TIL samples. (F) Stacked bar plot of distribution of TCR clonotype enrichment states by T cell state in the integrated pan-cancer data set.

To determine if expanded TCRs expressed by GZMK⁺ T_Eff were shared by other T cell states, we investigated the clonotype count, distribution, and associated T cell states of the top 10 expanded non-MAIT TCRs associated with GZMK⁺ T_Eff. Pairwise analysis was performed on autologous, matched PBMC and TIL from two patients (Fig. 6D). In the case of GBM106, the top 10 expanded TCRs associated with the GZMK⁺ T cell state in GBM106 PBMC were minimally expanded, with only one TCR detected more than once. This TCR was detected in a total of six PBMC associated T cells: five GZMK⁺ T cells and one T_EMRA cell. Overall, these same 10 TCRs were either not present in the TIL or detected with a single count. In comparison, the top 10 expanded TCRs associated with the GZMK⁺ T cell state in GBM106 TIL were detected at a much greater frequency. For example, the most highly expanded TIL clone was detected in 47 cells, with 17 and 27 of these cells associated with the NR4A2^lo or NR4A2^hi GZMK⁺ T cell state, respectively. Notably, all TCRs were mostly, if not entirely, associated with either the NR4A2^lo/hi GZMK⁺ T cell state or shared with the cytotoxic T_RM state. Importantly, these top expanded TCRs in the TIL were either undetected or detected only once in the matched PBMC sample. In the case of patient G4A112_Re PBMC, the top 10 expanded TCRs associated with the GZMK⁺ T cell state were detected at higher frequencies than TCRs in GBM106 PBMC. However, these expanded TCRs were mostly associated with the T_EMRA state, as seen with the top two most expanded TCRs. These TCRs were detected in TIL and either exclusively expressed by the T_EMRA state or shared with the NR4A2^hi/lo GZMK⁺ T cell state, suggesting that T_EMRA and GZMK⁺ T cell states may also share a developmental trajectory. Given the presence of these expanded TCRs in both PBMC and TIL, they may reflect infiltration of non-tumor-specific bystander T cells. Meanwhile, the top 10 expanded TCRs associated with the GZMK⁺ T cell state in G4A112_Re TIL were heavily expanded solely in the TIL and not the matched PBMC sample. Similarly, these TCRs were largely specific to the NR4A2^lo/hi GZMK⁺ T cell state, although several TCRs, the most expanded in particular, were also shared by cytotoxic T_RM. However, in TIL, we did not detect shared expression of specific TCRs between T_EMRA and GZMK⁺ T cells. The observation that GZMK⁺ TCRs were selectively expanded in TIL compared to PBMC was representative of the general clonotypic landscape of all matched samples as paired TIL and PBMC samples had low overlap (Fig. 6E). Given that T cells with the same TCRs are derived from the same progenitor cell, these results overall support our pseudotemporal trajectory analysis suggesting that cytotoxic T_RM and NR4A2^lo/hi GZMK⁺ T cells lie along the same developmental trajectory in the tumor microenvironment.

To determine whether expansion of GZMK⁺ T cells was specific to brain tumors, we investigated the TCR expansion states in our integrated tumor data set. Of the 20 tumor types examined, 13 contained V(D)J data that allowed for quantification of expansion states (Supplementary Figs. S12G & S12H). Notably, we observed that within the integrated tumor data set, both NR4A2^lo and NR4A2^hi GZMK⁺ T cells (C3 & C4) were expanded, with greater than 60% of cells either expanded or hyperexpanded (Fig. 6F).

From these results, we concluded that expanded GZMK⁺ TCRs were exclusively expanded in the tumor setting compared to matched blood; within the tumor microenvironment, they were mostly expressed by either GZMK⁺ T cells or cytotoxic T_RM. Furthermore, expansion of GZMK⁺ T cells was not specific to brain tumors but occurred in many different cancer types arising from diverse anatomical sites.

CD8⁺ GZMK⁺ TCRs are not detectable following ex vivo expansion

Adoptive cellular transfer of ex vivo expanded TIL derived from patient tumor specimens has been shown to be a promising immunotherapeutic approach in several cancer types (41,42). As antigen-specific T cells are enriched within the tumor microenvironment, isolated tumor specimens represent an ideal source of tumor-specific effector T cells for such immunotherapies. However, as many terminally differentiated T cell states lose capacity for self-renewal, it is unclear which effector T cell states are truly expanded from these ex vivo techniques. Recently, TCRs associated with GZMK⁺ TIL derived from PDAC patient samples have been shown to sustain similar frequencies following ex vivo culture, suggesting GZMK⁺ TIL retain proliferative potential (40). As a result, we were interested in assessing if these putative antigen-experienced, clonally expanded CD8⁺ NR4A2^lo/hi GZMK⁺ T_Eff derived from HGG patient samples were similarly capable of ex vivo expansion for potential therapeutic considerations. To this end, TIL from 6 different HGG patient samples were expanded in a gas permeable rapid expansion (G-Rex) culture system and subsequently analyzed via bulk TCR CDR3 beta-chain sequencing using the immunoSEQ platform (Fig. 7A). Among all paired samples, we observed an increase in overall clonal expansion following ex vivo culture represented by an increase in the occupied repertoire space by the topmost expanded TCRβ chains (Fig. 7B). However, when analyzing the TCR repertoire of each sample type, we found that cultured TIL TCR clonotypes were not representative of either D0 TIL or D0 PBMC. Such results are shown by representative alluvial plots of the top 20 TCRβ chains of D0 TIL, D25 cultured TIL, and D0 PBMC samples derived from patient sample GBM111_Re (Supplementary Fig. S13A). When investigating specifically the top 20 TCRβ chains of D0 TIL, we observed minimal enrichment of such TCRβ chains in D25 cultured TIL and some enrichment of such TCRβ chains in D0 PBMC. The inverse of such results is observed when investigating the top 20 TCRβ chains of D25 cultured TIL. These results were summarized across all samples with the use of the Morisita's overlap index, a measurement of the overlap among different data sets (Supplementary Fig. S13B), and similar results have been previously reported (43). Specifically analyzing the expansion status of TCRs associated with CD8⁺ GZMK⁺ T cells, we matched bulk TCRβ chains with those from CD8⁺ GZMK⁺ T cells within the single cell TCR data sets. In two representative samples, GBM111_Re and GBM113, we observed that these matched TCRβ chains were present in D0 TIL but were either not detected or unexpanded in cultured TIL, aside from one TCRβ chain which was expanded in cultured GBM113 TIL (Figs. 7C and 7D). Overall, these results suggested that the clonally expanded GZMK⁺ TIL present in the initial HGG microenvironment have limited expansion potential ex vivo using conventional expansion conditions. This potential loss of proliferative capacity, supported by low TCF7 expression (Supplementary Fig. S4), may also be HGG specific as expanded GZMK⁺ TIL derived from PDAC patient samples have been shown to continuously expand ex vivo (40).

Figure 7: TIL expanded ex vivo are not representative of the initial TCR landscape.

(A) Schematic of ex vivo growth of HGG TIL in G-rex and subsequent downstream analysis. Created with BioRender.com. (B) Bar plot of frequency of clonotype expansion states among various samples. (C, D) Alluvial plots highlighting the overlap between matched sample types of the β chains matched to single cell CD8⁺ GZMK⁺ TCRs in D0_TIL.

Discussion

In this study, we provided a combination of CITE-seq and scRNA-seq to explore the proteogenomic landscape of T cells within primary and recurrent HGG, including GBM and G4A, matched peripheral blood (PBMC), and brain metastases (BrMet). With paired single cell V(D)J sequencing, several observations emerged. First, using protein markers, we labeled T cell states that were otherwise unidentifiable by scRNA-seq alone and characterized the cell surface protein phenotypes of T cell states. Second, we observed an unexpected lack of canonical exhaustion markers at both the RNA and protein level, which we supported with additional computational methods. Similar results were reported by other published single cell analyses of GBM TIL, but thorough comparisons to other cancer types were not performed (17,43). Third, GZMK⁺ T cells comprised the largest percentage of CD8⁺ TIL within malignant brain tumors and were enriched in the tumor microenvironment compared to peripheral blood. Interestingly, this subset was further defined by the expression level of NR4A2. Fourth, through pseudotemporal ordering analysis, we inferred that NR4A2^hi GZMK⁺ T cells may represent a terminal developmental state derived from naїve CD8⁺ T cells and transitioning through NR4A2^lo GZMK⁺ T cells. Fifth, with sub-cellular resolution spatial transcriptomic analysis, we found that GZMK^hi T cells localized near vasculature and CD163⁺ MHCII⁺ myeloid cells. Sixth, CD8⁺ GZMK⁺ TIL comprised the most clonally expanded T cells relative to peripheral blood as well. Finally, when we integrated our data set, including both HGG and BrMet samples, with those from 18 other cancer types, we observed that GZMK⁺ TIL were found across all cancer types at varying frequencies, with higher frequencies in HGG. Taken together, these data further refine our understanding of the T cell populations within malignant brain tumors and across many other cancer types.

The comparative integrated analysis with additional cancer data sets harboring T cells with high-level exhaustion signatures, such as those found in melanoma, highlighted the profound differences in T cell exhaustion states between HGG and other cancers. In the overall T cell landscape of our data set, we observed high levels of inhibitory receptor PD-1 protein and PDCD1 RNA, moderate levels of TIGIT protein and RNA, and low levels of LAG-3, TIM-3, and BTLA protein and RNA. When applying an exhaustion gene signature based upon 28 genes upregulated in functionally exhausted T cells derived from melanoma (28), we found slightly elevated expression of this signature in GZMK⁺ T_Eff and T_reg clusters. Compared with 19 other cancer types, including BrMets and cancers known to harbor high levels of exhausted T cells, HGG, and in particular GBM, was infiltrated by some of the lowest frequencies of exhausted T cells. In this way, our conclusions support and expand the findings of a recent study by Naulaerts et al. which demonstrated that GBM TIL did not express strong exhaustion signatures (26). We also uniquely highlight the importance of supplementing cell surface phenotyping with additional approaches, such as transcriptional profiles, to measure exhaustion. Ultimately, however, further investigation into the epigenetic landscape and functional capacity of these T cells is required to understand how these T cells may differ from canonically exhausted T cells. Clinically, the low level of exhaustion coupled with the low overall frequency of T cell infiltration in GBM, could help explain the lack of efficacy of anti-PD-1/PD-L1 therapies in patients with GBM (3,4). Additional mechanisms driving T cell inhibition or dysfunction, such as activation of CD161 (17) or IL-10 release (20) by myeloid cells, could be more therapeutically compelling targets.

Our focus turned to NR4A2^lo/hi GZMK⁺ T_Eff that highly expressed GZMK, but did not express classic cytotoxic genes, such as GZMB, GNLY, and PRF1. While this T cell transcriptional state has been previously described in several single cell studies that encompass many different cancer histologies (26,31,35,40,44–47), several findings herein are novel. We found that GZMK⁺ T cells occupied a significant proportion of T cells in HGG both within our own data set and another published study, especially in contrast to other tumor types and matched blood and are further characterized based on NR4A2 expression. Moreover, these data support a developmental trajectory in which GZMK⁺ T cells may arise from classic cytotoxic T cells that acquire NR4A2 expression. Strikingly, tumor derived GZMK⁺ T cells were clonally expanded, whereas TCR matched GZMK⁺ T cells derived from blood were either not expanded or not detectable in matched blood, suggesting GZMK⁺ T cells either homed to the tumor site or developed in situ. While GZMK⁺ T cells were similarly abundant in metastatic brain samples, cytotoxic T_RM were more clonally expanded compared to GZMK⁺ T cells. The selective enrichment of clonally expanded GZMK⁺ T cells in the tumor microenvironment may suggest that these T cells are antigen-experienced and potentially tumor-specific. Further functional studies are needed to explore this possibility and identify the cognate antigen targets. Interestingly, a prior study in lung cancer demonstrated that GZMK⁺ T cells were equally clonally expanded in the blood and in lung tumors and suggested that further differentiation occurred in the tumor microenvironment (48). This observation, when compared to our findings, raises the intriguing possibility that T cell developmental fates may vary amongst tumor types and merits additional examination.

Given the mixed nature of our patient samples, which included both newly diagnosed and recurrent IDH-WT GBM and IDH-mut G4A, we were interested in understanding how CD8⁺ GZMK⁺ TIL may differ in these different settings. Using our available samples, we found that the relative frequency of CD8⁺ GZMK⁺ TIL was higher in IDH-mut G4A that IDH-WT GBM, and transcriptional levels of exhaustion do not noticeably differ between these populations. Analyzing TIL derived from IDH-WT GBM and IDH-mut G4A SMART-Seq2 patient samples, we found that the level of GZMK was similar in these two populations. Meanwhile, while we were unable to perform similar analyses on primary and recurrent IDH-WT GBM and IDH-mut G4A samples, we reanalyzed scRNA-seq data from a recent study by Lee et al. (36) that contained 14 primary, 12 recurrent, and 14 anti-PD-1 treated recurrent GBM patient samples. The relative frequency of CD8⁺ NR4A2^lo GZMK⁺ T cells was higher in primary GBM samples compared to recurrent GBM samples, while the relative frequency of CD8⁺ NR4A2^hi GZMK⁺ T cells was similar. In addition, there were no major differences in transcriptional profiles of CD8⁺ GZMK⁺ T cells based upon recurrence status. These results overall reinforce that GZMK⁺ T cells are the dominant CD8⁺ T cell state in high grade gliomas with similar transcriptional profiles regardless of IDH and recurrence status. However, given the relatively limited number of available data sets and matched functional studies, further investigation into how the transcriptional and functional profiles of TIL may differ based upon IDH and recurrence status in HGG is required.

Interested in understanding the spatial localization of GZMK expressing T cells, we performed sub-cellular resolution spatial transcriptomic analysis using the 10X Xenium in situ platform. This approach also demonstrated the significant proportion of GZMK^hi T cells in GBM. In line with previous work in ovarian cancer (35), we found that GZMK^hi T cells were generally restricted to vasculature regions and significantly associated with CD163 expressing MHCII⁺ myeloid cells, more so than GZMK^lo T cells. In ovarian cancer, GZMK⁺ T cells were found to be enriched in immune excluded tumors, compared to immune infiltrated tumors. Furthermore, within immune excluded tumors, GZMK⁺ T cells were found within the stroma, rather than the tumor region itself. Meanwhile, previous work has shown that CD163 expressing MHCII⁺ myeloid cells in GBM inhibit T cell functionality through IL-10 secretion (20) which can be prevented by anti-CSF1R therapy (34). In addition, tumor-associated macrophages (TAMs) in lung squamous-cell carcinomas, which express CD206 and CD163, have been shown to reduce T cell motility through long-lasting cell-to-cell interaction (49). Given the increased expression of TOX and EOMES in GZMK^hi T cells and their proximity to CD163⁺ MHCII⁺ myeloid cells, we suggest that these myeloid cells may be an important driver in regulating GZMK^hi T cell function and impairing their migration deeper into the tumor itself. Further work will however be needed to test these hypotheses.

As observed in the initial cohort of HGG and PBMC samples, when integrating our data set with several other cancer types, we noted two distinct populations of GZMK⁺ T cells: NR4A2^hi and NR4A2^lo GZMK⁺ T cells. NR4A2^hi GZMK⁺ T cells were enriched for many genes associated with, and downstream of, TCR activation, whereas these genes were significantly downregulated in the NR4A2^lo GZMK⁺ T cells. Given previous studies have shown that NR4A2 expression significantly increases within one hour following TCR stimulation (50) and is dependent on NFAT and AP-1 signaling (51), these data suggest that NR4A2^hi GZMK⁺ T cells have recently undergone TCR stimulation. NR4A2, which encodes NR4A2 and also known as NURR1, appears to perform several functions. In T cells, NR4A2 has been associated with resident memory differentiation (52,53). In addition, the NR4A family has been associated with modulating T cell function, as triple knock out (TKO) of the NR4A1, NR4A2, and NR4A3 transcription factors significantly increased the effector function of CAR T cells in vivo (54). Specifically, TKO of NR4A1–3 in CAR T cells promoted tumor regression and decreased expression of several inhibitory receptors while increasing the expression of effector genes. Furthermore, recent studies of the double knock out (DKO) of the TOX transcription factor family in murine models revealed a cross-regulatory relationship between the TOX and NR4A transcriptional families (55). Given that the most expanded GZMK⁺ TCRs were shared by both NR4A2^hi and NR4A2^lo transcriptional states, these results suggest that a switch in transcriptional state occurs. While we hypothesize that NR4A2^hi GZMK⁺ T cells develop from NR4A2^lo GZMK⁺ T cells, further study is required to first validate the presence of NR4A2 protein expression in NR4A2^hi GZMK⁺ T cells and second understand the developmental relationships and trajectories of these T cell subsets.

As adoptive T cell therapy (ACT) has become an increasingly developed technology and expanded GZMK⁺ TIL from PDAC patient samples have been shown to further expand ex vivo (40), we were interested in investigating whether clonally expanded GZMK⁺ TIL from HGG would likewise expand in ex vivo conditions currently used for ACT in the clinical setting. While TIL were expanded in this system, with enrichment of several clonotypes, they recapitulated neither the initial TIL nor PBMC clonotype landscape. Similar observations have been made by another group in which the TCR landscape of expanded TIL from GBM did not represent the original TIL TCR landscape (43). In our study, we additionally showed that GZMK⁺ TIL likewise demonstrated this pattern. Specifically, bulk TCRβ chains matched to single cell TCRs associated with the GZMK⁺ T cell state were expanded in initial HGG TIL but following ex vivo culturing were either minimally expanded or not detected. These results indicated that the use of current T cell expansion systems will likely need to be modified for further expansion of the most clonally expanded TIL present within the HGG tumor microenvironment. If these populations, which mostly include GZMK⁺ subsets, possess tumor-specific T cells, then an alternative approach to generating large numbers of these expanded clonotypes is required for ACT to be particularly effective for GBM treatment. Furthermore, specifically for the expansion of GZMK⁺ TIL, given expanded GZMK+ TIL from PDAC have been shown to further expand ex vivo with treatment of IL-2 and anti-CD3 and anti-4-1BB monoclonal antibodies (40), the lack of GZMK⁺ expansion in our samples may be due to ex vivo culturing conditions, and not T cell exhaustion. Recent work has shown that GZMK⁺ T cells frequencies increase optimally in the presence of cytokines only, including IL-2 and IL-12, and decrease following TCR activation (56). However, whether the loss of GZMK⁺ T cells is due to loss of GZMK expression or lack of T cell proliferation will require additional TCR sequencing. Overall, further work is needed to elucidate the proper ex vivo growth conditions for further expansion of clonally expanded glioma TIL and specifically GZMK⁺ T cells.

The emerging identification of both RNA and protein expression of GZMK within TIL raises the question of what the function of GZMK expression may be in these biological settings. Several studies have revealed the various functions of GZMK, both intracellular and extracellular. Intracellularly, GZMK has been shown to induce both caspase-dependent and -independent cell death in vitro, in addition to non-cytotoxic functions, such as inhibiting viral replication (57–61). Extracellularly, GZMK has been implicated in promoting a pro-inflammatory cytokine response (61,62). Notably, a recent study by Mogilenko et al. investigated the effects of age on the immune system and discovered the enrichment of GZMK⁺ T cells in older mice (29). These GZMK⁺ T cells developed in response to unknown extrinsic factors and expressed several exhaustion markers. Interestingly, GZMK was shown to induce a pro-inflammatory response, specifically stimulating production of IL6 and CCL2, both components of the senescence-associated secretory phenotype (SASP), by murine fibroblasts in vitro. In this way, GZMK⁺ T cells may contribute to age-related inflammation. Meanwhile, recent work showing that GZMK expression is affected by chronic antigen stimulation in both murine T cells (bioRxiv: 2023.04.17.537229) or human CAR T cells (63) underscores a potential role for its expression as a biomarker of antigen experience and also highlights a potentially distinct lineage separate from canonical exhaustion. In murine models, GZMK was initially expressed at day 0 (D0), downregulated at D4, and upregulated at D7. Meanwhile, GZMB continuously increased following chronic antigen stimulation. Moreover, human CAR T cells harboring the 4–1BB signaling domain were dominated by a GZMK⁺ transcriptional profile following 15 days of antigenic stimulation. In contrast, human CAR T cells containing the CD28 signaling domain were only enriched by a canonically exhausted transcriptional profile. These results suggest that the GZMK⁺ transcriptional profile, perhaps influenced by distinct signaling inputs, may represent a functional state that is distinct from the canonical exhaustion state. In our own study, we performed functional enrichment analysis on the top 25 DEGs by CD8⁺ GZMK⁺ T cells and found several pathways suggesting these T cells were activated. In conjunction with the selective expansion of these T cells in the tumor microenvironment and lack of exhaustion markers, we hypothesize that GZMK⁺ T cells are tumor specific and potentially functional. Finally, further work is also needed to understand how GZMK⁺ T cells may interact with other immune cell types within the microenvironment. Specifically, a recent study on non-metastatic colorectal tumors found that neutrophils were associated with GZMK⁺ T cells in the tumor microenvironment, which eventually promoted tumor relapse (64).

Finally, given the rising importance of immunotherapy in treating various types of cancers, the question of whether GZMK expression is associated with therapeutic response to immunotherapy remains unclear. In several studies and cancer types, including both liquid and solid cancers, GZMK⁺ TIL were enriched in patients responsive to anti-PD-1 therapy (30,65–67). In particular, in a recent study by Liu et al. (30) investigating the TIL states following anti-PD-1 therapy in non-small cell cancer biopsies, GZMK⁺ TIL were found to be enriched in patients responsive to anti-PD-1 therapy while GZMB⁺ TIL were enriched in non-responsive patients. In contrast, while GZMK⁺ TIL were detected in patients with head and neck cancer, there was no difference in proportion of GZMK⁺ TIL between responders and non-responders to neoadjuvant anti-PD-1 therapy (68). In GBM, neo-adjuvant pembrolizumab therapy has been shown to increase the frequency of CD8⁺ GZMK⁺ TIL in patients with recurrent GBM compared to untreated, non-matched primary GBM (36). However, an analysis of the clinical correlations associated with CD8⁺ GZMK⁺ TIL frequencies was not performed. Several studies have shown that specific tumor intrinsic factors, such as ERK1/2 phosphorylation or CHEK2 expression, can impact T cell activation and infiltration and response to anti-PD-1 treatment (69,70). While GZMK expression was not analyzed in these studies, further investigation into how GZMK expression may change in these settings is needed. Thus, these studies highlight the importance of better understanding whether GZMK⁺ TIL may correlate with immunotherapy responses.

Taken together, our findings provide a detailed analysis of the T cell states in HGG, consisting of both GBM and G4A, metastatic brain tumors, and matched PBMC samples. Given the unique and enriched clonal expansion of GZMK⁺ TIL, investigation of the antigen specificity of this T cell subset as well as how best to modulate, such as by licensing or expanding, this population to maximally augment its effector functions in patients with brain tumors may hold clinical relevance.

Materials and Methods

Patient recruitment and sample collection

Adult patients undergoing neurosurgical intervention at Barnes-Jewish Hospital and Massachusetts General Hospital were screened. Selection criteria included (1) age > 18 years and (2) presence of either primary or recurrent glioblastoma or high-grade glioma with clinical indications for surgical resection. Written informed consent was obtained from patients prior to surgery following the Washington University School of Medicine Institutional Review Board (IRB) Protocol #202107071, Washington University School of Medicine IRB Protocol #2011110011 and #202101040, the Dana Farber/Harvard Cancer Center IRB Protocol #10-417 and secondary Massachusetts General Hospital IRB Protocol #2022P001982, and the Massachusetts General Hospital IRB Protocol #2022P000955. Following surgical resection, specimens were placed in normal saline and maintained on ice. Characteristics of patients are summarized in Supplementary Table S1. All procedures and experiments were performed in accordance with the Helsinki Declaration.

Clinical Sample Processing: Single Cell RNA Sequencing

Tumor samples were processed as previously reported (10). Cohort 1 samples were stained with anti-CD45, anti-CD3, anti-CD11b, Zombie NIR Viability Dye, and TotalSeq-C antibodies (Supplementary Table S3). Live CD3⁺CD11b⁻ single cells were sorted by FACS on a BD FACSAria II (BD Biosciences) and a single cell suspension was submitted for sequencing. Cohort 2 samples were similarly stained with anti-CD45, anti-CD3, anti-CD11b, and Zombie NIR Viability Dye. Live CD45⁺ single cells were sorted by FACS on a BD FACSAria II (BD Biosciences) and a single cell suspension was submitted for sequencing.

Peripheral blood mononuclear cells (PBMCs) were separated through Ficoll-Paque PLUS density gradient (GE). The resulting buffy coat was collected, stained, and submitted for sequencing.

Clinical Sample Processing: Flow Cytometry

Tumor samples were collected from the operating room on ice and initially weighed. Samples were separated into aliquots up to 750 mg, as recommended by the Brain Tumor Dissociation Kit (Miltenyi Biotec). Aliquots were macerated on ice and disaggregated using the Brain Tumor Dissociation Kit (Miltenyi Biotec) on the gentleMACs Octo Dissociator with Heaters (Miltenyi Biotec). Samples then underwent Percoll (GE Healthcare Life Sciences) density gradient centrifugation to remove myelin contamination. The resulting pellets then underwent RBC lysis with ACK Lysis Buffer (Lonza Biosciences) and were passed through 70 micron filters. The tumor samples were then stained with anti-CD45RO, anti-CD3, anti-CD4, anti-CD8, anti-GZMK, anti-GZMB, anti-TCR Vα7.2, and Zombie Aqua Viability Dye (Supplementary Table S3) and analyzed on a Cytoflex LX (Beckman Coulter) or Sony MA900 (Sony).

Clinical Sample Processing: Xenium in situ

Tumor sample was collected from the operating room on ice and separated into several pieces. Each portion was fixed in 10% neutral buffered formalin for 24 hours, transferred to 70% ethanol, and embedded in paraffin wax.

Single Cell Library Preparation (CITE-seq/scRNA-seq/VDJ-seq)

Single cell gene expression and V(D)J libraries were prepared using the Chromium Next GEM Single Cell 5’ Reagents Kits v1 and Chromium Single Cell V(D)J Reagent Kits v1.1 (10X Genomics), respectively. cDNA was prepared after the GEM generation and barcoding, followed by the GEM-RT reaction and bead cleanup steps. Purified cDNA was amplified for 10–14 cycles before being cleaned up using SPRIselect beads. Samples were then run on a Bioanalyzer to determine the cDNA concentration. V(D)J target enrichment (TCR) was performed on the full length cDNA. Gene Expression, Enriched TCR and Feature libraries were prepared as recommended by the 10x Genomics Chromium Single Cell V(D)J Reagent Kits (v1 Chemistry) user guide with Feature Barcoding technology for Cell Surface Protein and Immune Receptor Mapping user guide, with appropriate modifications to the PCR cycles based on the calculated cDNA concentration. For sample preparation on the 10x Genomics platform, the Chromium Single Cell 5’ and Gel Bead Kit v2, 16 rxns (PN-1000006), Chromium Single Cell A Chip Kit, 48 rxns (PN-1000152), Chromium Single Cell V(D)J Human TCR Enrichment Kit (PN-1000005), Chromium Single Index Kit T, 96 rxns (PN-1000213), Chromium Single Cell 5' Feature Barcode library Kit, 16 rxns (PN-1000080) and Single Index Kit N Set A, 96 rxns (PN-1000212) were used. The concentration of each library was accurately determined through qPCR utilizing the KAPA library Quantification Kit according to the manufacturer’s protocol (KAPA Biosystems/Roche) to produce cluster counts appropriate for the Illumina NovaSeq6000 instrument. Normalized libraries were sequenced on a NovaSeq6000 S4 Flow Cell using the XP workflow and a 151×10×10×151 sequencing recipe according to the manufacturer protocol. A median sequencing depth of 50,000 reads/cell was targeted for each Gene Expression library and 5,000 reads/cell for each V(D)J and Feature library.

Single Cell RNA Sequencing Analysis

Raw sequencing data generated from 5’ gene expression libraries from both cohorts 1 and 2 were processed with the CellRanger pipeline (10X Genomics, default settings, version 3.0.1) and mapped onto a human genome reference (GrCh38). Downstream analysis was performed using the Seurat R package (71). Initial filtering excluded cells that had any of the following: a nFeature count less than 500, mitochondrial reads greater than 10%, or a nCount value greater than the 93rd percentile of each individual sample. All TCR related genes were then removed. Following this, doublets were initially identified using the DoubletCollection R package (72) in which several doublet packages were used to label doublets within each sample. The specific packages used included scds (73), scDblFinder (74), doubletCells (75), DoubletFinder (76), and Scrublet (77). Cells labeled as a doublet by three or more packages were then removed.

For samples within Cohort 1, which contained both RNA and antibody-derived tag (ADT) (i.e. or protein), single cell data, the RNA assay was first integrated using the reciprocal PCA (RPCA) method. Principal component analysis (PCA) on the integrated data set was then performed and the optimal number of principal components (PCs) was determined based upon elbow plots, jackstraw resampling, and PC expression heatmaps. Next, the ADT assay was similarly integrated using the RPCA method. The RNA and ADT assays were then integrated together using the Weighted Nearest Neighbor Analysis (WNN) (78). Dimensionality reduction and visualization were performed with the uniform manifold approximation and projection (UMAP) algorithm (arXiv: 1802.03426v3) (Seurat implementation) followed by unsupervised graph-based clustering. Clusters of T cells were iteratively filtered based upon expression of PTPRC and CD3E and a lack of expression of myeloid markers, such as CD14 and SPI1. Filtered cells were reanalyzed using the RPCA method followed by PCA for both the RNA and ADT assays. Finally, WNN was used to integrate the reanalyzed RNA and ADT assays followed by UMAP visualization. The final optimal number of PCs and clustering resolution were 30 and 0.9, respectively. T cell states were determined based upon select RNA and protein markers (Supplementary Table S2) and differentially expressed genes of each cluster determined using the Wilcoxon rank-sum test-based function. These annotations were supported by single cell gene enrichment analysis, which was performed as previously described using the escape R package (79–81). In brief, cluster specific gene signatures were derived from the top 50 differentially expressed genes from each cluster. Defined gene signatures from various single cell data sets of T cells were collected (17,38,81–83). Each cell was then assigned a score for each gene signature and the correlation of each cluster specific gene signature was assessed against other defined gene signatures. These correlations were used to support existing T cell annotations or were used to label specific T cell clusters (Supplementary File S4).

The RNA assay from Cohort 1 was integrated with the RNA assay from Cohort 2, which was subjected to the same initial filtering parameters, using the RPCA method. Similarly, PCA was performed followed by UMAP visualization. In addition, T cells were iteratively filtered as previously specified. The final optimal number of PCs and clustering resolution were 40 and 1.3, respectively. T cell states were determined based upon select RNA markers, differentially expressed genes of each cluster, and guided by T cell annotations from Cohort 1, which were derived from both RNA and protein markers.

CD8⁺ T cells were subsetted based upon both RNA and protein expression (cohort 1) or just RNA expression of CD8 (cohort 1 & cohort 2). Samples were reintegrated with the RPCA method followed by UMAP visualization. The final optimal number of PCs and clustering resolution were 24 and 1.1, respectively (cohort 1) and 25 and 2.5, respectively (cohort 1 + cohort 2). T cell states were determined based upon select RNA markers, differentially expressed genes of each cluster, and guided by T cell annotations from the parent data set.

Multiple cancer data set integration

CD8⁺ T cells from this current study of HGG and BrMet samples were integrated with CD8⁺ T cells from the following data sets: 1) a recent pan-cancer TIL data set (38), 2) a melanoma data set (39), 3) an independent GBM data set (17), and 4) a PDAC data set (40). Only samples with 5’ gene expression libraries (10X Genomics) were used for this integration. In several samples, genes in the Ensembl format were converted to respective gene names using clusterProfiler (84) and any outstanding discrepancies in gene names were corrected. Any genes related to T cell receptor genes, such as TRAV related genes, immunoglobulin genes, such as IGH related genes, and MALAT1 were removed as previously described (38). Only genes common to all samples were used to avoid any potential library inconsistencies, with a final 10,211 genes shared by all samples. Prior to integration, T cells were separated by tumor type and the total number of T cells from each tumor type was randomly subsetted to 1,904 cells for a total of 39,984 cells. All samples were then merged, normalized, the top 2000 variable genes were calculated, and scaled. To minimize noise from cell cycle and tissue dissociation, the S phase score and G2M phase score calculated by the function CellCycleScoring and mitochondrial gene percentage calculated with PercentageFeatureSet were regressed (28). PCA was then performed with the optimal number of PCs followed by integration using the Harmony R package (85). Dimensionality reduction and visualization were performed with the uniform manifold approximation and projection (UMAP) algorithm (arXiv: 1802.03426v3) (Seurat implementation) followed by unsupervised graph-based clustering. The final optimal number of PCs and clustering resolution were 50 and 1.0, respectively. T cell states were determined based upon select RNA markers (Supplementary Table S2) and differentially expressed genes of each cluster determined using the Wilcoxon rank-sum test-based function (Supplementary File S1: Sheet 9). These annotations were supported by the cell annotations specified in the meta data provided by the original authors.

Analysis of IDH-WT GBM and IDH-mut G4A SMART-Seq2 samples from independent study

CD8⁺ T cells from this current study of HGG and BrMet samples were integrated with CD8⁺ T cells from an independent GBM data set consisting of 11 IDH-WT GBM and 6 IDH-mut G4A SMART-Seq2 samples (17). SMART-Seq2 data was downloaded from GEO GSE163108 and CD8⁺ T cells were subsetted according to previously published annotations. Counts within the CD8⁺ T cells from this current study were initially converted to TPM using the NormalizeData function with normalization.method set to “RC” (Seurat R implementation). Only genes common to all samples were used to avoid any potential library inconsistencies. The data sets were then merged and to minimize noise from cell cycle and tissue dissociation, the S phase score and G2M phase score calculated by the function CellCycleScoring and mitochondrial gene percentage calculated with PercentageFeatureSet were regressed (28). PCA was then performed with the optimal number of PCs followed by integration using the Harmony R package (85). Dimensionality reduction and visualization were performed with the uniform manifold approximation and projection (UMAP) algorithm (arXiv: 1802.03426v3) (Seurat implementation). The final optimal number of PCs was 50.

Analysis of TIL from primary and recurrent GBM samples from independent study

scRNA-seq data was downloaded from GEO GSE154795 and CD3+ cells were subsetted according to CD3E and PTPRC expression (36). The data set was then re-integrated by sample using the RPCA method. Similarly, PCA was performed followed by UMAP visualization. The final optimal number of PCs and clustering resolution were 30 and 0.5, respectively. T cell states were determined based upon select RNA markers (Supplementary Table S2) and differentially expressed genes of each cluster.

Exhaustion module score

The two exhaustion scores were generated using AddModuleScore (Seurat implementation) and based upon select gene markers and the gene list previously published by Tirosh et al. (28) (Supplementary Table S2).

Pseudotemporal ordering analysis

This study’s HGG and BrMet data set:

Pseudotime trajectory analysis was performed with the monocle2 R package (32). Out of the CD8⁺ T cells, only the C2: CD8⁺ T_N, C9: CD8⁺ NR4A2^lo GZMK⁺ T_Eff, C10: CD8⁺ NR4A2^hi GZMK⁺ T_Eff, C11: CD8⁺ cytotoxic T_RM, and C12: CD8⁺ T_EMRA clusters were analyzed. The top 1,000 unique differentially expressed genes of these clusters were used to guide pseudotime trajectory analysis. The trajectory was rooted in the C2:CD8⁺ T_N cluster. The top 200 genes from the differentialGeneTest monocle2 function were used to generate the heatmap.

Integrated pan cancer data set:

Pseudotemporal ordering analysis was performed with the monocle2 R package (32). Out of the CD8⁺ T cells, only the C1: T_N, C2: IL7R⁺ T_EM, C3: NR4A2^lo GZMK⁺ T_Eff, C4: NR4A2^hi GZMK⁺ T_Eff, C5: cytotoxic T_RM, C7: T_EMRA, and C8: T_Exhausted clusters were analyzed. The top 200 unique differentially expressed genes of these clusters were used to guide pseudotime trajectory analysis. The trajectory was rooted in the C1:CD8⁺ T_N cluster. The top 200 genes from the differentialGeneTest monocle2 function were used to generate the heatmap.

Gene functional enrichment analysis

Gene functional enrichment analysis was performed through ToppGene (37). The top 25 DEGs expressed by CD8⁺ GZMK⁺ T cells were selected, and the top two unique gene ontology biological pathways were displayed. The pathways and associated genes are listed in Additional File 3: Sheet 2.

Xenium In Situ Workflow

Sample Preparation

Formalin-fixed, paraffin-embedded (FFPE) tumor sample was sectioned into 5-um sections using a microtome onto a Xenium slide according to specifications delineated by 10X Genomics in Protocol CG000578. The Xenium slide was then deparaffinized and decrosslinked according to the 10X Protocol CG000580 and as previously discussed (86). Both a pre-designed gene expression panel (Human Brain Probe set) and a custom probe set based on a previously published GBM data (bioRxiv: 2022.08.27.505439) was used during probe hybridization, ligation, and amplification according to 10X Protocol CG000582. The Xenium slide was then loaded into the Xenium Analyzer using 10X Protocol CG000584 and regions of interest (ROIs) were initially selected using the on-instrument software.

Downstream analysis

Xenium output files were processed using the Seurat R package (5.0.1). Initial filtering excluded cells that had any of the following: an nFeature count less than 5, an nFeature count greater than 150, an nCount value greater than 1,000, or a control probe count greater than 2. Counts were then log-normalized and PCA was performed with the optimal number of PCs determined based upon elbow plots, jackstraw resampling, and PC expression heatmaps. Dimensionality and visualization were performed with the UMAP algorithm (arXiv: 1802.03426v3) (Seurat implementation) followed by unsupervised graph-based clustering. The aggregated object containing all cells was initially subsetted into three categories: tumor and neural-related cells, vascular cells, and immune cells based on select marker genes (Supplementary Table S4). Each subsetted object was iteratively filtered with doublets being removed based on expression of canonical genes from non-overlapping cell types. Filtered cells were then reanalyzed with PCA and UMAP visualization. The final optimal number of PCs for tumor and neural-related cells, vascular cells, and immune cells was 30. The final optimal clustering resolutions were 0.3, 0.8, and 0.8, respectively. Cell states were annotated based upon select gene markers (Supplementary Table S4) and differentially expressed genes of each cluster determined using the Wilcoxon rank-sum test-based function. The aggregated sample was then filtered and annotated according to subsetted objects. PCA and UMAP visualization was performed again on the aggregated sample. The final optimal number of PCs and clustering resolution were 30 and 0.3, respectively.

Xenium output files were separately imported using the scanpy Python package (1.9.6) (87). Cells were filtered and annotated based on those present in the Seurat object. Neighborhood enrichment and co-occurrence analyses were performed using the squidpy Python package (1.3.1) (88). The “complete” clustering method (89) was used for the neighborhood enrichment analysis.

Single Cell TCR analysis

Raw sequencing data generated from V(D)J libraries from both cohorts 1 and 2 were processed with the CellRanger V(D)J pipeline (10X Genomics, default settings, version 2.0.0) and mapped onto a human genome reference (GrCh38–2.0.0). The scRepertoire R package (90) was used for all subsequent clonotype analysis. Clonotype states were defined as the following: single (0 < X ≤ 1), doublet (1 < X ≤ 2), expanded (2 < X ≤ 50), and hyperexpanded (50 < X), where X is the number of cells in which the clonotype is detected.

G-Rex Culturing & Bulk TCR Sequencing

Tumor samples were processed into 1 mm³ sections and incubated in a G-Rex 6M cell culture system (Wilson Wolf, Minnesota). Approximately 5–6 sections were incubated per well. Tumor samples were incubated in R10 media consisting of RPMI-1640 (Gibco) supplemented with 10% fetal bovine serum (FBS), 1% L-glutamine (Corning), 1% Pen/Strep (Gibco), 1% sodium pyruvate (Gibco), 0.5% sodium bicarbonate (Corning), 50 μM β-mercaptoethanol (Sigma-Aldrich), and 5,000 IU/mL recombinant human interleukin-2 (rhIL-2). Fresh IL-2 containing R10 media was added every 3 to 4 days. After approximately one month, cells were harvested by vigorous pipetting and then filtered through a 70 micron filter. Cells were then washed twice with R10 media and resuspended in ACK Lysis Buffer (Lonza Biosciences) if necessary. Cells were then frozen down in liquid nitrogen in 90% FBS with 10% DMSO.

Bulk T cell receptor (TCR) variable beta chain immunosequencing of genomic DNA from frozen T cells was performed using the ImmunoSEQ Assay (Adaptive Biotechnologies, Seattle, WA) on the following: TIL processed the day of surgical resection (D0), PBMCs isolated the day of surgical resection (D0), and expanded TIL (DX).

Bulk TCR from G-rex Analysis

Data generated from the G-Rex systems were analyzed using the immunarch R package (http://doi.org/10.5281/zenodo.3367200, RRID: SCR_023089, RRID: SCR_004129). Bulk TCRβ data was matched to single cell V(D)J data by matching bulk TCRβ sequences with single cell TCR clonotypes that contained the respective beta chains.

Data availability

FASTQ files are available in the NCBI Sequence Read Archive (SRA) (PRJNA1051638). Processed gene expression and V(D)J matrices and Seurat objects used for all analyses are available on the open-access data sharing platform Zenodo (https://doi.org/10.5281/zenodo.8198492, RRID: SCR_004129).

Statistical analysis

For comparison of groups, the Mann-Whitney U test was used. For comparison of frequencies, the Fisher’s exact test was used. For differential gene expression, the Wilcoxon rank-sum test (two-sided) was used with Bonferroni correction. p ≤ 0.05 was considered to be statistically significant. Statistical analyses were performed with R statistical language V 4.1.0 or Seurat.

Significance.

To understand the limited efficacy of immune checkpoint blockade in GBM, we applied a multi-omics approach to understand the TIL landscape. By highlighting the enrichment of GZMK⁺ effector T cells and lack of exhausted T cells, we provide a new potential mechanism of resistance to immunotherapy in GBM.