Lung tumor–infiltrating T_reg have divergent transcriptional profiles and function linked to checkpoint blockade response

Abstract

Regulatory T cells (T_reg) are conventionally viewed as suppressors of endogenous and therapy-induced antitumor immunity; however, their role in modulating responses to immune checkpoint blockade (ICB) is unclear. In this study, we integrated single-cell RNA-seq/T cell receptor sequencing (TCRseq) of >73,000 tumor-infiltrating T_reg (TIL-T_reg) from anti–PD-1–treated and treatment-naive non–small cell lung cancers (NSCLC) with single-cell analysis of tumor-associated antigen (TAA)–specific T_reg derived from a murine tumor model. We identified 10 subsets of human TIL-T_reg, most of which have high concordance with murine TIL-T_reg subsets. Only one subset selectively expresses high levels of TNFRSF4 (OX40) and TNFRSF18 (GITR), whose engangement by cognate ligand mediated proliferative programs and NF-κB activation, as well as multiple genes involved in T_reg suppression, including LAG3. Functionally, the OX40^hiGITR^hi subset is the most highly suppressive ex vivo, and its higher representation among total TIL-T_reg correlated with resistance to PD-1 blockade. Unexpectedly, in the murine tumor model, we found that virtually all TIL-T_reg–expressing T cell receptors that are specific for TAA fully develop a distinct T_H1-like signature over a 2-week period after entry into the tumor, down-regulating FoxP3 and up-regulating expression of TBX21 (Tbet), IFNG, and certain proinflammatory granzymes. Transfer learning of a gene score from the murine TAA-specific T_H1-like T_reg subset to the human single-cell dataset revealed a highly analogous subcluster that was enriched in anti–PD-1–responding tumors. These findings demonstrate that TIL-T_reg partition into multiple distinct transcriptionally defined subsets with potentially opposing effects on ICB-induced antitumor immunity and suggest that TAA-specific TIL-T_reg may positively contribute to antitumor responses.

INTRODUCTION

Immune checkpoint blockade (ICB) induces durable clinical responses in a subset of patients with non–small cell lung cancer (NSCLC); however, most tumors do not respond (1). A major frontier in cancer immunology is understanding the many factors in the tumor microenvironment (TME) that block versus promote antitumor immunity. ICB efficacy relies on endogenous tumor-reactive T cells for tumor eradication. In our ongoing efforts to understand mechanisms of ICB response versus resistance, we recently reported on the transcriptional programming of CD8⁺ tumor-infiltrating lymphocytes (TILs) in resectable lung cancers treated with neoadjuvant ICB (2). We found that the breadth and frequency of CD8⁺ TILs recognizing antigens derived from tumor somatic mutations (neoantigens) were not associated with pathologic response, thereby suggesting the existence of immunosuppressive mechanisms that can thwart otherwise effective antitumor immunity.

One candidate immune inhibitory cell population in the TME is the regulatory T cell (T_reg), characterized by expression of CD4 and the forkhead box protein P3 (FoxP3). These cells maintain systemic immune homeostasis by regulating peripheral tolerance and mitigating autoimmune disease (3) through a variety of immunosuppressive mechanisms (4). T_reg depletion experiments in mice demonstrate that this population can potently suppress antitumor effector T cell responses and prevent tumor clearance by endogenous tumor-specific effector T cells (5, 6). T_reg representation among total TILs in diverse murine and human cancers is generally much higher than in corresponding normal tissue and blood, and increased tumor infiltration by T_reg expressing specific genes was recently correlated with poor prognosis in NSCLC (7, 8). These findings have led to multiple clinical approaches to selectively deplete or inhibit T_reg in tumors, but none to date have been successful, possibly because of lack of T_reg specificity or failure to inhibit key functionally relevant T_reg subsets. Thus, an understanding of the phenotype and specificity of tumor-infiltrating lymphocyte T_reg (TIL-T_reg) is critical to target them for effective immunotherapy.

Contrary to conventional CD4 T cells (T_conv), whose self-specific repertoire is negatively selected, T_reg are selected for expression of T cell receptors (TCRs) that recognize self-antigen, thereby allowing them to maintain self-tolerance via targeted immune suppression. Under certain circumstances of pathologic autoimmunity, T_reg can lose FoxP3 and differentiate to acquire a T_conv program (so-called ex-T_reg) and thereby exacerbate autoimmunity via interferon-γ (IFN-γ) production and recruitment of T helper 1 (T_H1) cells (9). Although historically T_reg were dichotomized into thymic (natural T_reg) versus peripheral (induced T_reg) in origin, single-cell RNA sequencing (scRNA-seq) technology has enabled more refined analysis of these cells to reveal a level of T_reg heterogeneity that was not previously appreciated (10–16).

To study the transcriptional programming and function of TIL-T_reg at refined resolution, we used scRNA-seq of TIL-T_reg harvested from resected human lung cancers [and adjacent normal lung (NL)] from untreated patients and those treated with anti–PD-1 and, in parallel, integrated these data with an analysis of a murine tumor model wherein tumor-associated antigen (TAA)–specific T_reg were tracked. We identified two distinct T_reg subsets displaying opposite associations with ICB response. One is an activated, highly suppressive T_reg subset, characterized by a unique tumor necrosis factor receptor superfamily (TNFRSF) expression pattern and high levels of genes encoding multiple suppressive and tumor-homing molecules, that associates with ICB nonresponsiveness. The other is a unique population of TAA-specific T_reg, defined in murine tumors, that exhibits certain characteristics of “ex-T_reg,” with Foxp3 down-regulation and an acquired T_H1-like effector program. Using computational gene score homology analysis, we identified an analogous population in human lung cancers that was enriched in ICB-responding tumors. These findings reveal a previously unappreciated diverse substructure of TIL-T_reg that consists of distinct subsets that can either inhibit or enhance antitumor responses, defining potential T_reg targets for therapeutic modulation.

RESULTS

T_reg transcriptional programs in resected lung cancers treated with ICB

To better understand the transcriptional programs of TIL-T_reg subsets and their potential role in ICB responsiveness, we analyzed a dataset of coupled single-cell TCRseq/RNA-seq on CD3⁺ TIL isolated from resected lung cancers and adjacent NL tissue, derived from patients treated with neoadjuvant PD-1 blockade (NCT02259621) (17). We have previously reported on transcriptional programming of neoantigen-specific CD8⁺ T cells from this dataset (2). Of these 15 ICB-treated tumors, 6 had a major pathologic response (MPR), defined as ≤10% residual tumor at the time of surgery (table S1), which was recently associated with improved overall survival in lung cancers treated with neoadjuvant combination immunotherapy in the CheckMate 816 trial (NCT02998528) (18). Per standard convention, we refer to tumors with MPR as responders (R) and tumors without MPR as nonresponders (NR). In addition, coupled single-cell TCRseq/RNA-seq was performed prospectively on CD3⁺ T cells from tumor and adjacent NL from 10 treatment-naive patients with resectable lung cancer (table S2). After stringent quality control filtering, 851,674 CD3⁺ T cells passed quality control and were integrated to obtain broad CD4⁺ and CD8⁺ T cell subsets as visualized by uniform manifold approximation and projection (UMAP) dimensionality reduction (fig. S1A). Refined clustering of 375,001 CD8⁻, CD4⁺ T cells identified 11 subclusters, with three distinct T_reg clusters (fig. S1B). Computational selection of FoxP3⁺CD8^neg cells identified 73,882 T_reg. Consistent with known T_reg transcriptional programs (19–21), this FoxP3⁺CD8^neg population expressed higher levels of CTLA4 and IKZF2, which encodes the zinc finger protein Helios and lower levels of IL7R (CD127) relative to conventional CD4⁺ TIL (fig. S1C). This large number of individual human T_reg transcriptomes allowed us to cluster the TIL-T_reg with extremely high resolution; thus, we identified 10 distinct T_reg subclusters (Fig. 1A and table S3). Four “Activated” clusters were identified on the basis of high expression of IL2RA, CCR8, and ICOS, which are broadly associated with T_reg activation (22–24) and expression of published gene sets induced by TCR activation (KEGG; Kyoto Encyclopedia of Genes and Genomes) and IL-2 signaling (BioCarta) (Fig. 1B and fig. S1D). Cluster-defining genes and select functional markers are shown in Fig. 1 (B and C) and S1E. Activated (1)/ OX40^hiGITR^hi had a transcriptional program consistent with a unique activated state, as evidenced by high selective expression of the TNFRSF members TNFRSF4 (OX40) and TNFRSF18 (GITR) (Fig. 1, B and C, and fig. S1E) (7). Although these molecules have been reported to be generally expressed among TIL-T_reg, they have not previously been shown to be selectively expressed by a single T_reg subset, representing on average 19.9% of total TIL-T_reg (range, 9.75 to 41.67%; fig. S2A). A number of additional activation genes were shared with one of the other activated clusters, including LAG3 [Activated (2)] and TNFRSF9 (41BB), ICOS, and CCR8 [Activated (3)]. In contrast, there was an LN-homing cluster (SELL and S1PR1), two “resting” clusters based on expression of SESN3 and gene sets associated with resting T_reg (Fig. 1, B and C, and fig. S1, D and E), and an “inactive” population with overall low differential gene expression or gene set enrichment analysis association (Fig. 1B and fig. S1D). Other distinct clusters included a type I IFN-responsive cluster enriched for IFN-inducible genes such as IFI6 and MX1 and expression of an IFN-responsive gene set and an unusual subset characterized by expression of genes typically expressed by T_H1 cells, such as CCL4 (MIP-1β), CCL5 (RANTES), IFNG, and multiple granzyme genes typically found in activated CD8 cytotoxic lymphocytes (Fig. 1, B and C; fig. S1E; and table S3). Cell proportion analysis identified enriched representation of the Activated (1)/ OX40^hiGITR^hi (P = 0.0054), Resting (2) (P = 8.8 × 10⁻¹⁰), and Activated (4) (P = 0.014) clusters in tumor relative to adjacent NL, whereas the LN-homing cluster was enriched in adjacent NL (P = 4.5 × 10⁻⁸; fig. S2A). The distinct pathologic responses among different patients in the trial allowed us to ascertain whether specific T_reg subsets were associated with ICB response, but globally, there was no difference in TIL-T_reg subcluster proportions between treated and untreated tumors nor between R and NR (fig. S2B). Principal component analysis (PCA) of pseudobulk gene expression separated TIL-T_reg from those derived from adjacent NL (Fig. 1D), indicating distinct global transcriptional programs in tumor-derived versus adjacent NL T_reg, but could not distinguish R from NR (Fig. 1E) nor treated from treatment-naive (fig. S2C). These data suggest that more refined analyses of T_reg subsets might be needed to define transcriptional programs associated with PD-1 blockade resistance.

Fig. 1. Single-cell transcriptomic profiling of T_reg in treatment-naïve and neoadjuvant anti–PD-1–treated NSCLCs.

Coupled scRNA-seq/TCR-seq was performed on T cells isolated from resected tumor (n = 15), adjacent NL (n = 12), TDLN (n = 3), and a resected brain metastasis (n = 1) from patients with NSCLC treated with two doses of neoadjuvant anti–PD-1 as well as resected tumor (n = 10) and paired adjacent NL (n = 8) from treatment-naïve patients with NSCLC. (A) Two-dimensional (2D) UMAP projection of the expression profiles of the 73,882 T_reg that passed QC. T_reg subsets, defined by 10 unique clusters, are annotated and marked by color code. (B) Relative expression [average log₂(fold change)] for top differential genes for each cluster is visualized on a heatmap. Three thousand cells (or all cells in the cluster if cluster size <3000 cells) were randomly sampled from each cluster for visualization. Differential expression tests for T_reg cell subsets were performed with Wilcoxon rank sum test. Genes with >0.25 log₂ fold changes, at least 25% expressed in tested groups, and genes with Bonferroni-corrected P values < 0.05 were regarded as differentially expressed. (C) The expression of canonical T_reg subset marker genes and cell subset selective genes was visualized in red scale using UMAP projection. (D) PCA and canonical correlation of pseudobulk gene expression for individual tumor (yellow, n = 25) and adjacent NL (dark blue, n = 20) samples. Canonical correlation with tissue type = 0.56, P < 0.001. (E) PCA and canonical correlation of pseudobulk gene expression for individual nonresponder (red, n = 9) and responder (blue, n = 6) tumors. Canonical correlation with response status = 0.36, P = 0.50.

OX40^hiGITR^hi T_reg are highly suppressive and correlate with ICB resistance

Although OX40 and GITR have been targeted with agonist antibodies based on their costimulatory roles on T_conv, they are also expressed by TIL-T_reg (25), in which their role has yet to be clarified. Among all the TIL-T_reg, the Activated (1)/OX40^hiGITR^hi cluster exhibited markedly increased expression of multiple genes encoding molecules that mediate suppressive activity, such as FoxP3, IL2RA, LAG3, ENTPD1 (CD39), EBI3, and LAYN (7, 23, 26–28), when compared with all other T_reg clusters, suggesting a highly immunosuppressive phenotype (Fig. 1B and table S4). Unlike in prior work by Guo and colleagues (29), TNFRSF9 (4–1BB) was not particularly selective for this subset. Variations of this “activated” T_reg subset have been identified in prior lung cancer studies and have consistently shown coexpression of TNFRSF4 (OX40) and TNFRSF18 (GITR) (7, 29, 30). Although bulk LAG3 expression by TIL-T_reg has also been shown (7), LAG3 coexpression by this OX40^hiGITR^hi subset has not previously been reported. Moreover, the suppressive function of this subset and its association with ICB response are not known.

To determine whether the Activated (1)/OX40^hiGITR^hi T_reg displayed functionally superior suppressive capacity, we sorted OX40^hiGITR^hi and OX40^lowGITR^low T_reg from ICB-treated tumors to test their ability to suppress T_conv cells ex vivo. OX40^hi GITR^hi T_reg were markedly more suppressive than the composite remainder of TIL T_reg that expressed low levels of TNFRSF4 and TNFRSF18 (Fig. 2A and fig. S3). Although the proportion of individual T_reg subclusters was not different between our treatment and response groups (fig. S2B), we hypothesized that T_reg activation status could be affected because of the enrichment of the Activated (1)/OX40^hiGITR^hi cluster in TIL compared with adjacent NL T_reg (figs. S2A and S4A). We therefore developed an activated T_reg score derived from the top differentially expressed genes in Activated (1) relative to all other TIL-T_reg. In addition to TNFRSF4 and TNFRSF18, LAG3 stood out as the third most highly differentially expressed gene (Fig. 2B, fig. S1E, and table S4). TIL-T_reg expressing the activated T_reg score were localized within the Activated (1)/ OX40^hiGITR^hi cluster (Fig. 2C and fig. S4, B and C) and significantly more enriched in NR versus R tumors (P = 0.038; Fig. 2, C and D, and fig. S4, B and C). Treatment-naïve tumors had high representation of T_reg with a high activation score, similar to ICB NRs (P = 0.670), suggesting that contraction of this subset was a consequence of successful ICB treatment in responding tumors. In contrast, the unusual T_H1-like/cytotoxic cluster (Fig. 1, A and B) was not enriched in NR tumors. There was no difference in frequency of T_reg with a high activation score in adjacent NL among the different groups (P = 0.620), indicating that the difference between NR and R is specific to the TME (Fig. 2D). Enrichment of activated T_reg in NR tumors could be due to differentiation into a more activated, suppressive state, better maintenance of that state, and/or prevention of differentiation out of that state. Previous studies have used velocity trajectory analyses to demonstrate tumor TNFRSF⁺ T_reg to be a terminal differentiation state (12, 15). To address this question, we assessed TCR clonal sharing between each subcluster, based on the notion that clones shared between and among clusters can serve as lineage tracing markers and thus reflect developmental connectivity. This method has been previously used to study the development and differentiation of many T cell subsets, including T_reg in human and murine tumors (31–33). Of 1047 clones found in more than one cluster, 699 (66.8%) were shared between Activated (1)/OX40^hiGITR^hi and Activated (3) (Fig. 2E). Next, we performed diffusion trajectory and RNA velocity analysis on T_reg clones shared in the Activated (1)/OX40^hiGITR^hi and Activated (3) clusters to better understand potential developmental relationships between these clusters in NR versus R tumors. On the basis of this analysis, we found that TIL-T_reg within R tumors have a trajectory consistent with differentiating out of Activated (1)/OX40^hi/GITR^hi into Activated (3), contrasting with what is seen in NR tumors, in which the TIL-T_reg developmental trajectory is consistent with differentiation into the Activated (1)/OX40^hiGITR^hi cluster with essentially no differentiation out of that state (Fig. 2F). Together, these findings indicate that the paucity of Activated (1)/OX40^hiGITR^hi cells in R tumors could be due to a net differentiation flux out of this subset, in contrast to NR (Fig. 2F) (34). In support of this, the proportion of Activated (1) relative to Activated (3) was significantly decreased in R tumors relative to NR (P = 0.0115) and treatmentnaïve (P = 0.0136; Fig. 2G).

Fig. 2. Activated, OX40^hiGITR^hi T_reg are functionally suppressive and associate with nonresponse to PD-1 blockade.

(A) Functional analysis of OX40^loGITR^lo and OX40^hiGITR^hi TIL-T_reg–mediated suppression of conventional CD4⁺T cell (T_conv) proliferation. TIL-T_reg from patients MD017–0157, MD017–0092, MD017–0115, MD017–0124, and MD017–0116 were combined to ensure sufficient T_reg numbers for this experiment (n = 1). (B) Volcano plot showing differential expression between the Activated (1)/OX40^hiGITR^hi cluster (right) versus all other T_reg (left) when analyzing all tumor samples (n = 25). Each dot represents one gene. False discovery rate (FDR) < 0.05 is considered significant (blue/red dots). LAG3, TNFRSF18, and TNFRSF4 represent the top three most differentially expressed genes in the Activated (1)/OX40^hiGITR^hi T_reg (red dots). (C) Overlay of the Activated T_reg score on the TIL-T_reg UMAP for each response/treatment group. Red indicates higher expression; blue indicates lower expression. (D) The frequency of T_reg with a high T_reg activation score is shown for tumor (left) and adjacent NL (right) from nonresponders (NR; red), responders (R; blue), and treatment-naive patients (Tx-naive; green). Comparisons were performed at the individual patient level using Wilcoxon rank test. (E) Clonotype sharing pattern across T_reg subsets. The frequencies of T_reg TCR clones that were detected in at least two T_reg clusters were calculated and are shown on a heatmap. (F) Diffusion plot with RNA velocity for the Activated (1)/ OX40^hiGITR^hi and Activated (3) clusters (among which most clonotype sharing was observed). (G) Boxplots showing the relative proportion of Activated (1)/OX40^hi GITR^hi and Activated (3) clusters by R (blue), NR (red), and treatment-naïve status (green). (H) A violin plot shows TNFSF4 (OX40L) expression by neoantigen-specific (red) and flu-specific (blue) CD8 TILs from the same neoadjuvant-treated patients. (I) Violin plot comparing TNFSF4 (OX40L) expression by neoantigen-specific CD8⁺ TIL between nonresponding (red)and responding (blue) tumors. (J) Gene set enrichment analysis to evaluate differing biological functions of ex vivo OX40L, 41BBL, and GITRL agonism in sorted human TIL-T_reg. TIL-T_reg from patients MD043–011 and MD01–019 were combined to ensure sufficient T_reg numbers for this experiment (n = 1). NF-κB, nuclear factor κB.

Beyond being markers for a highly suppressive T_reg, TNFRSF4, TNFRSF18, and TNFRSF9—the three TNFRSF members most highly expressed by the Activated (1)/ OX40^hi/GITR^hi subset— likely play active roles in their function (34) that would be dependent on expression of their cognate ligands within the TME. We therefore broadly queried the expression of TNFR ligands on all intratumoral T cell and myeloid populations to define specific cell types that may be interacting with the TNFRSF4- and TNFRSF18-expressing T_reg. We found that in treated and untreated NSCLC, expression of TNFSF18 (GITRL) was very low across all cell types in the TME, thus precluding our analysis of this ligand. However, TNFSF9 (4–1BBL) was broadly expressed on both myeloid and T cells, and TNFSF4 (OX40L) was highly expressed on a limited number of T cells, most notably on CD8⁺ T cells (fig. S4D). Given that expression of TNFSF4 was restricted to CD8 T cells, we queried our previously published flu-reactive and neoantigen-specific CD8⁺ TILs in this patient cohort (2). There was significantly higher expression of TNFSF4 on neoantigen- versus flu-reactive TILs (P < 2.2 × 10⁻¹⁶; Fig. 2H). Furthermore, TNFSF4 expression on neoantigen-specific CD8⁺ TILs was much higher in NR compared with R tumors (P < 2.2 × 10⁻¹⁶; Fig. 2I). There was no significant difference in expression in TNFSF9 or TNFSF18 on neoantigen- versus flu-reactive TILs, regardless of response status (fig. S4E).

To assess the effect of TNFRSF signaling on T_reg transcriptional programs, we generated HEK-293T cell lines expressing each of the three TNFRSF ligands (OX40L, GITRL, and 41BBL; fig. S5). Sorted T_reg from two NR tumors were cocultured with the ligand-expressing 293T cells, followed by single-cell TCRseq/RNA-seq of the T_reg. In each ligand-stimulated condition, the noncanonical nuclear factor κB pathway was enriched (R-HAS-5676594, GSE18893; Fig. 2J and tables S5 to S7). Specifically, 4–1BB agonism initiated its unique gene program (BIOCARTA_41BB_PATHWAY; table S5), 4–1BB and GITR agonism showed enrichment for a gene set associated with TNFR2 signaling (R-HAS-5668541; tables S5 and S6) (35–38), and OX40 agonism strongly up-regulated a cell proliferation gene set (GO:0008283; table S7 and fig. S2J) (39). Given the particularly high expression of TNFSF4 on neoantigen-specific CD8⁺ T cells in NR tumors (Fig. 2I), our findings suggest that they may be providing a signal to expand this suppressive OX40^hi GITR^hi T_reg subset, thereby subverting their own potential antitumor activity. In summary, we show here that, although OX40⁺ and GITR⁺ T_reg have been identified previously in lung cancers (29, 30, 35, 40, 41), this subset is functionally highly suppressive ex vivo and is expanded in anti–PD-1 NR, possibly driven by TNFSF4 expression by neoantigen-specific CD8 T cells.

Tumor antigen-specific T_reg acquire a T_H1-like transcriptional signature in an ICB-responsive murine tumor model

In direct contrast to T_conv cells, T_reg are selected to recognize self-antigen (42). Naturally occurring T_reg are positively selected in the thymus by a broad set of self-antigens expressed in thymic epithelial cells (43–45). To model the biology of TAA-specific TIL-T_reg, we developed a mouse adoptive transfer model of T_reg with fixed specificity. We crossed the B6.Cg-Tg(TcrLCMV) TCR transgenic line that has a CD4⁺ TCR (termed SMARTA) specific for the H2-I-A^b–restricted LCMV GP_61–80 epitope (46) to mice that have expression of the LCMV glycoprotein (GP) controlled by the rat insulin promoter (RIP-GP) (47, 48). It has been well documented that RIP is also active in murine thymic epithelial cells (49, 50). Consistent with the developmental biology of T_reg, virtually no FoxP3⁺ T_reg bearing the SMARTA TCR develop in the absence of GP as a self-antigen, so this mating results in offspring with a much higher frequency of LCMV GP_61–80–restricted CD4⁺FoxP3⁺ T_reg because of thymic expression of GP (49, 51). We then crossed these mice to CD45.1;RFP-FoxP3 mice to generate a T_reg reporter mouse line with a transgenic TCR of known reactivity. CD45.1-marked transgenic CD25^hiRFP-FoxP3⁺ T_reg expressing the SMARTA TCR alpha and beta chains (Vα2.3-JαDK1 and Vβ8.3-Jβ2.5) were sorted, and 2 × 10⁵ SMARTA TCR⁺ T_reg per mouse were adoptively transferred to recipient B6 CD45.2 mice (fig. S6, A and B). As a tumor model, we used the ICB-responsive MC38, which either did or did not express GP. MC38-GP tumors grew more slowly than MC38wt tumors and were also more sensitive to PD-1 blockade (fig. S6A). Mice adoptively transferred with CD45.1-marked transgenic CD25^hiRFP-FoxP3⁺ T_reg expressing the SMARTA TCR were inoculated with either the MC38 parental line (MC38WT), which is an ICB-responsive tumor or an MC38-GP-expressing line (52). After 14 days of tumor growth, we harvested the tumors and sorted TIL-T_reg on the basis of high expression of CD25. The sorted population consisted of a mix of distinguishable tumor-reactive (TR)–T_reg that were adoptively transferred (SMARTA TCR⁺/RFP⁺) and endogenous T_reg (SMARTA TCR^neg/RFP^neg) (fig. S6C). Notably, fewer than 1% of total T_reg in the spleen were from the donor SMARTA TCR⁺ T_reg 14 days after adoptive transfer, a proportion similar to previous protocols using adoptively transferred T_conv (fig. S6D) (53).

Single-cell TCRseq/RNA-seq was performed on >30,000 murine TIL-T_reg, which identified five clusters (Fig. 3, A and B, and table S8). As expected, TR-T_reg accumulated in GP-expressing tumors and tumor-draining lymph nodes (TDLNs) to a significantly greater extent than in MC38WT tumors (P = 0.0159 and P = 0.0317, respectively), which indicates antigen recognition driving subsequent cell activation and proliferation (Fig. 3C).

Fig. 3. Defining the scRNA-transcriptome associated with T_reg tumor antigen reactivity using a transgenic TCR mouse model.

Coupled scRNA-seq/TCR-seq was performed on T_reg isolated from GP-expressing MC38 (MC38-GP, n = 5 mice per experiment) or parental MC38 (MC38WT, n = 10 mice per experiment) tumors, tumor-draining inguinal lymph node, and spleens on day 14 of tumor growth. (A) 2D UMAP projection of the expression profiles of the 30,325 T_reg that passed QC. T_reg subsets, defined by five unique clusters, are annotated and marked by color code. (B) Relative expression of three to five differential genes for each cluster is visualized on a heatmap. (C) Comparison of the frequency of SMARTA FoxP3⁺ T_reg of total T_reg defined by expression of CD45.1, FoxP3, and the SMARTA TCR between MC38WT (blue) and MC38-GP (red) tumor-bearing mice in each tissue compartment. Means with SEM error bars are shown for each mouse (n = 5). P values were obtained using Mann-Whitney test. (D) Homology to human T_reg clusters is shown by the average expression of genes scores built on the top 20 differentially expressed genes for each murine cluster queried in human clusters. (E) Quantification of cluster designation for all SMARTA clones from WT and GP-expressing MC38. (F) Cluster localization for SMARTA clones is shown with SMARTA TCR (red) overlaid on full 2D UMAP (gray) of T_reg isolated from MC38-GP (left) or MC38WT (right). (G) Averaged log₂-transformed and library size–normalized FoxP3 expression values are shown. Visualized in red scale using full UMAP projection. (H) TBX21 (Tbet) expression is shown for TR-T_reg (blue) and all other T_reg (red) from the GP-expressing MC38 tumor after 14 days in vivo. (I) Five thousand TR-T_reg and non–TR-T_reg were sorted and cocultured ex vivo with 1000 dendritic cells loaded with the LCMV GP peptide for 48 hours. Supernatants were harvested and assayed for IFN-γ using the Meso Scale Discovery platform. Data are shown as IFN-γ production above background, defined as T_reg cocultured with dendritic cells without peptide. Mann-Whitney test was used to compare average IFN-γ production between peptide-loaded and peptide-unloaded conditions (P = 0.33; n = 2). (J) Violin plot showing the TR-T_reg score of adoptively transferred SMARTA TCR^pos T_reg before adoptive transfer (pre-AT) and at 1, 7, and 14 days of tumor residence. (K) Violin plot showing FoxP3 expression by adoptively transferred SMARTA TCR⁺ T_reg before adoptive transfer (pre-AT) and at 1, 7, and 14 days of tumor residence. P values obtained by Student’s t test. The P value of the trend (P_trend) was calculated for the TR-T_reg score and normalized Foxp3 expression over time by testing the linear regression slope with null and alternative hypotheses: H0: β1 = 0 (the slope is equal to zero), HA: β1 ≠ 0 (the slope is not equal to zero). N = 5 to 10 mice per time point. (L) Volcano plots showing differentially expressed genes of SMARTA TCR⁺ T_reg in day 7 versus day 1. X axis shows log2(fold change), and y axis shows −log₁₀(FDR). Differentially expressed genes higher at day 7 are shown in red, and differentially expressed genes higher in 1 day are shown in blue. (M) TSDR methylation profile of sorted SMARTA TCR⁺CD25^hiRFP⁺ T_reg from female SMARTA;RIP-GP mice compared with control values (low, medium, and high). Residue numbers denote individual CpG motifs in reference to the transcription initiation site of FoxP3. Data are an average of n = 2 experimental replicates with cells from n = 3 mice pooled per experiment. Bar colors represent 0 to 100% methylation.

To compare our murine and human TIL-T_reg clusters, we used the top 20 genes from each of the murine T_reg clusters to develop gene scores. We then queried our human clusters to see which human clusters corresponded most to the murine clusters identified. The murine T_reg clusters were highly related to counterpart human TIL-T_reg clusters (Fig. 3D). Specifically, there was notable homology in the top cluster-defining genes for the resting and LN-homing clusters (CCR7, TCF7, KLF2, and S1PR1), the IFN-induced cluster (IFIT3, ISG15, and IRF7), and the T_H1-like cluster (CCL5, CCL4, and IFNG) (Fig. 3, B and D, and table S9). Murine TIL-T_reg also contained an activated cluster in which the top differentially expressed genes (TNFRSF4, TNFRSF18, TNFRSF9, IZKF2, EBI3, and CCR8) represent a compilation of several of the human activated clusters and was most analogous to Activated (1)/OX40^hi GITR^hi with expression of many suppressive T_reg genes (Fig. 3D) (23, 54). A notable difference was the proliferating cluster; although there were human TIL-T_reg with proliferation signatures, these were relatively few and did not segregate out into a separate cluster as they did in the mouse. This is likely because the relative kinetics of human versus murine tumor growth allows for a greater equilibrium to develop over the years of tumor development in humans. We next queried the murine single-cell data for coexpression of the SMARTA TCR alpha and beta chain to identify the transcriptional profile of TR-T_reg. In MC38-GP tumors, the TR-T_reg did not localize to the Activated (1)/OX40^hiGITR^hi T_reg cluster but rather acquired a transcriptional program characterized by up-regulation of T_H1-associated cytokine and proinflammatory chemokine genes such as IFNG, CCL3 (MIP1α), and CCL4 (MIP1β) (Fig. 3B and table S10). In contrast, the rare TR-T_reg cells harvested from the MC38WT tumors all resided in the resting/LN-homing cluster (Fig. 3, E and F), consistent with a lack of antigen stimulation. Because many of the genes associated with the TR-T_reg after 14 days’ residence in the tumor are not typically seen in conventional suppressive T_reg, we assessed the level of FoxP3 expression in each mouse T_reg cluster. The TR-T_reg cluster, although still RFP⁺ (indicating that the Foxp3 locus had been transcriptionally active), had relatively low FoxP3 mRNA expression compared with the other clusters (Fig. 3G). This is consistent with the long half-life of RFP that allows it to be detected after transcription of the linked gene is turned off (55). Antigen-dependent loss of FoxP3 expression in a minority of T_reg has been observed in models of autoimmunity and is associated with loss of T_reg-suppressive capacity and acquisition of T_conv activity (9). To determine whether the decreased FoxP3 mRNA was mirrored at the protein level, we sorted RFP⁺ T_reg directly from SMARTA (TR-T_reg) donor mice and compared intracellular FOXP3 protein expression with TR-T_reg isolated from recipient mice after the TR-T_reg had been in the MC38-GP expressing tumor for 14 days. Before adoptive transfer, an average of 94.2% (range, 92.4 to 95.3%) of the sorted population was FOXP3⁺, whereas only 22.8% were FOXP3⁺ after 14 days (fig. S6, C and E). Furthermore, the TR-T_reg cluster was highly enriched for TBX21 (Tbet) expression (Fig. 3H), which is a critical transcriptional regulator of the proinflammatory function of T_H1 cells and T_H1-like ex-T_reg observed in autoimmune settings (56–60). Although T_H1 cells are known to be proinflammatory, they only kill cells indirectly through activation of nearby phagocytic cells (61). In contrast, T_reg have been reported to directly kill cells through release of cytotoxic granules containing granzymes and perforin (62). Because our murine TR-T_reg up-regulate some cytotoxic granule molecules and NKG7 (63, 64), we asked whether they could kill MC38-GP tumor cells induced to express MHC2 through TCR:MHC binding. We harvested TIL TR-T_reg and endogenous T_reg on day 14 of tumor growth and plated them at varying ratios with IFN-γ–induced, MHC2⁺ MC38-GP tumor cells (fig. S7A). TR-T_reg did not exhibit any specific killing against MHC2⁺ GP-expressing tumor cells (fig. S7B), but they did functionally produce IFN-γ in an antigen-specific fashion when cocultured with LCMV GP_61–80 peptide–loaded dendritic cells ex vivo, although this was not significantly higher than background IFN-γ production (P = 0.33, Mann-Whitney test), likely due to the limited numbers of TIL-T_reg available for biological replicates (Fig. 3I). This observation suggests that T_H1-like T_reg could potentially assist antitumor immunity via a helper rather than cytotoxic mechanism. In addition, the up-regulation in the TR-T_reg cluster of TBX21, whose cognate protein, Tbet, has a known role in mediating the proinflammatory function of T_H1-like T_reg (Fig. 3H) (56–60), supports this notion. Together, the data demonstrate that tumor antigen-reactive T_reg acquire a distinct phenotype consistent with an “ex-T_reg” T_H1-like gene signature, accompanied by loss of FoxP3 and CD25 expression and upregulation of CCL5 and IFNG. We observed a similar trend of T_reg T_H1-like differentiation when adoptively transferring SMARTA TCR⁺ T_reg into RIP mice, in which the majority of TR-T_reg localized to the T_H1-like cluster (fig. S7, C to E, and tables S10 and S11), thus suggesting that the additional expression of LCMV-GP in the pancreas and kidneys did not affect the T_H1-like differentiation of these T_reg. Moreover, the majority of TR-T_reg acquired the T_H1-like gene program even in the absence of CD8⁺ T cells, which was achieved through antibody-mediated CD8⁺ T cell depletion of recipient mice before adoptive transfer (fig. S8, A to C, and table S12).

To evaluate the kinetics of acquisition of this T_H1-like phenotype, we performed single-cell TCRseq/RNA-seq on TR-T_reg harvested 1, 7, and 14 days after adoptive transfer. Although FoxP3 levels of the SMARTA TCR^pos T_reg were similar to those of pre-adoptive transfer donor T_reg after 1 day of tumor residence, these T_reg in MC38-GP tumors significantly lost FoxP3 expression and gained the “TR-T_reg” profile by day 7 (tables S8 and S13). When using the top 20 differential genes in the murine TR-T_reg cluster from the original murine dataset to generate a TR-T_reg score (tables S9 and 10), we observed a progressive increase in the TR-T_reg score (P < 2.2 × 10⁻¹⁶; Fig. 3J) together with a progressive decrease in FoxP3 expression (P < 2.2 × 10⁻¹⁶) over time (Fig. 3K). In addition, by 7 days after AT, TR-T_reg significantly up-regulate the T_H1-like genes CCL5 and IFNG and proliferating genes STMN1, MKI67, and TOP2A (Fig. 3L). The enhancement of proliferating genes concords with the progressive expansion in numbers of SMARTA⁺ T_reg over time in the MC38-GP tumors. In contrast, on day 1 of residence in the tumors, TR-T_reg express multiple resting/LN-homing genes, such as TCF7 and SELL, and the conventional T_reg genes FoxP3 and IL2RA (Fig. 3L and table S13). This type of T_reg differentiation has been reported in the context of autoimmune disease (9, 65, 66) but has not previously been demonstrated in the context of cancer.

In these analyses, it is possible that the SMARTA⁺ FOXP3^lo T_H1-like population was derived from the small percentage of FOXP3^neg T_conv cells present after SMARTA T_reg sorting (fig. S6E), which could have proliferated, turned on low levels of FoxP3, and taken over the SMARTA TIL population in vivo. To rigorously address this concern, we adoptively transferred a large number (5 × 10⁵) of pure GP-reactive T_conv cells (CD4⁺SMARTA TCR⁺CD25^lo RFP^neg) from CD45.1⁺ donor mice into CD45.2⁺ recipient mice bearing MC38-GP or MC38WT tumors. This represented roughly 50× more tumor-reactive T_conv CD4⁺ T cells than would be contaminating the sorted TR-T_reg used in Fig. 3. These were cells taken from the same SMARTA;RIP-GP;CD45.1;RFP-FoxP3 mice used for the T_reg adoptive transfers. In comparison with 22.8% of TR-T_reg expressing FOXP3 (fig. S6E), virtually none of tumor-reactive CD4⁺ T_conv TIL expressed FOXP3 after 14 days in vivo (fig. S9, A to C), although endogenous recipient-derived (CD45.2⁺ SMARTA TCR^neg) FOXP3⁺ T_reg could be readily detected in the tumor (fig. S9D). This lack of FOXP3 expression by CD4⁺ T_conv cells was also recapitulated in the spleen and TDLNs (fig. S9D). To further address the possibility that high-dose transgenic TCR T cell adoptive transfer could result in expansion of non-T_reg, T_conv cells leading to an artificial T_reg ➔ T_H1-like transition (67), we looked at the frequency of tumor-reactive SMARTA TCR⁺CD4⁺ T cells in the adoptively transferred population and in the tumor 1 day after adoptive transfer. Although 92.5% of adoptively transferred T_reg expressed the SMARTA (tumor-reactive) TCR, only 2.2% of the CD4⁺ TIL-T_reg were from the adoptively transferred SMARTA TCR⁺ cells 1 day after adoptive transfer, with the remaining 97.8% being endogenous T_reg from the recipient mouse (fig. S9E). This frequency of SMARTA TCR⁺ T_reg in the tumor after 1 day is consistent with prior studies of adoptively transferred CD4⁺ T cells (53).

To further support that the cells we adoptively transferred were TR-T_reg rather than T_conv cells, we performed epigenetic analysis of the T_reg-specific demethylated region (TSDR) in the RFP-FoxP3⁺ TR-T_reg and RFP-FoxP3^neg TR-T_conv cells directly from the SMARTA RIP-GP;RFP-FoxP3 mice as well as control wild-type T_reg (CD4⁺CD25^hi). The rationale behind this analysis is that T_conv cells will exhibit high methylation of the TSDR. Conversely, thymically derived T_reg have low levels of TSDR methylation that are in accordance with the sex of the donor from which the cells were derived, because FoxP3 TSDR methylation occurs on the inactive X chromosome (68, 69). Therefore, although FoxP3 TSDR methylation in male T_reg is typically <20%, this value in female T_reg ranges from 30 to 50% methylation. Conversely, T_conv cells have TSDR methylation >90% (70). To therefore confirm that the TR-T_reg obtained from our female mice were T_reg rather than T_conv cells, we analyzed the methylation status of TR-T_reg and TR-T_conv cells obtained from the same SMARTA mouse. The TR-T_reg isolated from female SMARTA mice have TSDR% methylation ranging from 39.6 to 48.9%, consistent with known FoxP3 TSDR values in T_reg from female mice due to X inactivation (71). The TR-T_conv cells, however, had methylation ranging from 79.3 to 95.9%, consistent with known FoxP3 TSDR values in T_conv cells. The female donor T_reg, which we used as a control, had methylation ranging from 55.1 to 69.1% (Fig. 3M and table S14). Together, we show that TR-T_reg with a T_H1-like signature have methylation patterns consistent with a typical female T_reg profile (72). We conclude that TCR stimulation of mature self-antigen-specific T_reg by TAA programs these cells to differentiate into a proinflammatory T_H1-like immune cell phenotype.

Murine TR-T_reg gene signature defines an orthologous human lung cancer subset enriched in anti–PD-1 R

The robust conversion of TIL-T_reg from resting to a T_H1-like program in MC38, which is an ICB-responsive tumor, raised the question of whether a human counterpart was associated with ICB responsiveness in our neoadjuvant cohort. To refine our search for the human ortholog to this TR-T_reg cluster identified in our murine tumor model, we used the gene score based on the top 20 differential genes defining the murine TR-T_reg cluster (tables S9 and 10). We identified the human T_H1-like/cytotoxic T_reg cluster (Fig. 1A) as the predominant population enriched for these genes (Fig. 4A and fig. S10A). To define a human transcriptional TR-T_reg ortholog at higher resolution, we performed refined subclustering on the human T_H1-like/cytotoxic T_reg cluster and identified six distinct subclusters. One of these, SC0, was dominantly enriched in the TR-T_reg gene score (Fig. 4, B and C, and fig. S10, A and B) and expressed high levels of IFNG, CCL4, and CCL5 (Fig. 4, B and C), which are three of the up-regulated genes associated with tumor reactivity in murine T_reg. SC0 also expressed multiple cytotoxic genes. It was not possible to assess antigen-specific killing activity of this population in human TIL-T_reg because their cognate antigens are not defined (as in the murine system) and they express no membrane molecules that would distinguish them selectively from other T_reg subsets. Although FOXP3 was expressed in all subclusters of the T_H1-like/cytotoxic T_reg (Fig. 4C), FOXP3 (P = 6.8 × 10⁻⁹) and IL2RA (CD25; P = 1.9 × 10⁻⁹) expression were significantly decreased in SC0 relative to all other T_reg, whereas TBX21 (Tbet) was significantly increased (P < 0.001) (Fig. 4D and fig. S10B; paired t test). Thus, SC0 has many T_H1 hallmarks of the TR-T_reg observed in the murine TIL-T_reg. In contrast to the Activated (1)/ OX40^hiGITR^hi subset, which was enriched in NR tumors, the TR-T_reg-score^hi region, SC0, was enriched in R tumors (Fig. 4, E and F, P = 0.066; Wilcoxon rank sum test). This enrichment in ICB R only occurs in the tumor and not in adjacent NL.

Fig. 4. Murine TR-T_reg gene signature defines an orthologous subset among human NSCLC TIL-T_reg that is enriched in anti–PD-1 R.

(A) Red scale overlay of the TR-T_reg gene score, with red indicating higher expression and blue indicating low expression. Red dotted line represents the UMAP region with highest expression. Refined clustering was performed on the T_H1-like/cytotoxic subset, and 2D UMAP projection of six unique subclusters (SC0-SC5) is visualized by UMAP and marked by color code. (B) Relative expression of the top 10 most differential genes for each subcluster is visualized on a heatmap. (C) The expression of biologically relevant genes from the TR-T_reg score is visualized in red scale. (D) Boxplots showing differences in expression of FoxP3 (P = 6.8 × 10⁻⁹), IL2RA (CD25; P = 1.9 × 10⁻¹), and TBX21 (Tbet, P = 0.00024) by TIL-T_reg in SC0 versus all other T_reg. Each dot represents an individual tumor sample per patient (n = 14). Comparisons were performed at the individual patient level using paired t test. Patient-averaged log₂-transformed and library size–normalized expression values are shown. (E) Cell density plots of the T_H1-like/cytotoxic T_reg subclusters stratified by response/treatment status. The TR-T_reg score^hi population is indicated with a dotted line. Increased density is represented by red scale, and decreased density is represented by green/blue. (F) Boxplots showing the frequency of SC0 in the tumors (left) and adjacent NLs (right) of NR (red) and responders (blue). Comparisons were performed at the individual patient level using Wilcoxon rank test. (G) Diffusion map with RNA velocity between SC0, Activated (1)/OX40^hiGITR^hi, and LN-homing clusters for NR (left) and R (right). (H) Cross-cluster sharing of T_reg TCR clonotypes detected in the SC0 subcluster with non-SC0 T_reg and T_conv clusters (as shown in fig. S1C). Red indicates a higher frequency of the clone in the relevant subcluster. T_cm, central memory T cells; T_em, effector memory T cells.

To better understand the relationship between SC0 and all adjacent clusters in the global T_reg UMAP [T_H1-like/cytotoxic, Resting, LN homing, Activated (1)/ OX40^hiGITR^hi, and Activated (3)], we performed diffusion trajectory and RNA velocity. The Activated (3), Resting, and LN-homing clusters aligned along a common trajectory, with the Activated (1)/ OX40^hiGITR^hi and T_H1-like/cytotoxic clusters diverging along two additional trajectories (Fig. 4G and fig. S10C). Because of these commonalities in diffusion trajectory and because of the relatively high clonal sharing between the T_H1-like/cytotoxic and Activated (1)/OX40^hiGITR^hi clusters as compared with the other UMAP-adjacent clusters (fig. S10D), we further refined this analysis to strictly assess the directionality between the LN-homing, Activated (1)/OX40^hiGITR^hi, and the TR-T_reg score^hi cluster (SC0). In RNA velocity analysis, SC0 cells in R and NR showed markedly different directionality of transcriptional gradients (Fig. 4G). In R SC0 T_reg, most velocity vectors moved away from the Activated (1)/OX40^hiGITR^hi and LN-homing clusters, suggesting a net differentiation into the SC0 cluster (Fig. 4G). This pattern was not observed in NR, where the small number of SC0 cluster cells appeared to be differentiating toward the Activated (1)/OX40^hiGITR^hi cluster. Last, to further support the relationship of the T_H1-like T_reg in the SC0 cluster to other T_reg rather than CD4⁺ T_conv cells, we analyzed the transcriptional heterogeneity of individual TCR clones that were found in the SC0 subcluster. If the T_H1-like T_reg were derived from activated T_conv CD4⁺ TIL rather than T_reg, then one would expect a high level of TCR clonal sharing between the T_H1-like T_reg and T_conv CD4⁺ TIL. Of the 441 clones that comprised SC0, 179 (40.6%) were not detected in any other CD4⁺ TIL subset, whereas 262 (59.4%) were found in at least one other CD4⁺ T cell cluster (Fig. 4H and fig. S10, E and F). Of these “shared” clones, 173 (66.0%) were solely detected in other, non-SC0 T_reg subsets, whereas only 43 shared clones (16.4%) were shared solely with T_conv subsets, and 46 clones (17.6%) were shared with both T_reg and T_conv subsets. The dominant sharing by SC0 clones with non-SC0 T_reg was observed across all patients, and the 16.4% of shared clones that were found in T_conv subsets were heterogeneous in nature, with no obvious predominance for a specific T_conv subset. These analyses provide support for the relative enrichment and maintenance of the T_H1-like cluster in R tumors and the origination of these cells from T_reg rather than T_conv CD4⁺ TIL.

DISCUSSION

In the present study, we resolved multiple transcriptionally distinct TIL-T_reg subsets—10 in human and five in mouse—with remarkable cross-species transcriptomic matching. One of the activated subsets, characterized by selective expression of TNFRSF4 (OX40), TNFRSF18 (GITR), and high LAG3 and multiple immune suppression-associated genes such as ENTPD1 (CD39) and EBI3 (IL27B), was functionally far more suppressive in vitro than all other TIL-T_reg and correlated with ICB resistance. In an ICB-responsive murine tumor model, T_reg specific for a TAA down-modulated FoxP3 and progressively developed into a population resembling T_H1 cells, which was dependent on expression of cognate antigen. We found a clear transcriptional ortholog to these TAA-specific T_reg in human NSCLC TIL-T_reg that is more highly represented among pathologic R to anti–PD-1.

Although OX40 and GITR expression by TIL-T_reg has been reported for a number of human cancers (73, 74), our findings define a single specific subset that expresses these TNFRSF members along with the highest levels of multiple suppressive genes among the overall population. Guo and colleagues focused on a subset of 4–1BB^hi T_reg, showing that its proportion correlated with worse outcomes in patients with lung cancer treated with anti–PD-1 (29); although this subset did express TNFRSF4, TNFRSF18, and other suppressive T_reg markers, LAG3 was notably absent, potentially missed with the relatively small number of T_reg profiled in that study. OX40, GITR, and LAG3 have all been independently shown to mark activated T_reg, and in general, activated T_reg tend to be more suppressive (30, 75); however, we had the unique opportunity here to functionally study a distinct subset coexpressing all three of these molecules. The particularly high LAG3 expression, which promotes T_reg suppression in murine models (76–78), on the Activated (1)/OX40^hiGITR^hi subset raises the possibility that the success of LAG3 blockade in human cancer immunotherapy may be related to inhibition of suppressive function by this OX40^hi GITR^hi T_reg population. Furthermore, it is notable that a major source of TNFSF4 in the TME is the neoantigen-specific CD8⁺ population, with much higher expression in TILs from anti–PD-1–resistant tumors. It is probable that the TME variability between R and NR contributes to the maintenance of these T_reg in NR tumors. This has been shown in studies of hepatocellular carcinoma in which TNFRSF4 expression correlated with poor prognosis overall, including association with tumor growth characteristics and the functions of macrophages, plasmacytoid dendritic cells, and T_reg (79, 80).

The remarkable similarity among TIL-T_reg subsets between a commonly used murine cancer and human NSCLC was validated by comparative gene set enrichment analysis. The most unexpected finding from our murine model is that when cognate antigen is expressed in the tumor, TR-T_reg accumulate, down-modulate FoxP3, and progressively develop from a resting state into a population expressing TBX21, T_H1-like cytokines and chemokines, and multiple cytotoxic molecules. We rigorously ruled out the possibility that the TAA-specific T_H1-like T_reg population comes from expansion and intratumoral accumulation from the few (<5%) TAA-specific T_conv CD4 cells contaminating the T_reg at the time of adoptive transfer that might transiently up-regulate FoxP3. The development of this population in an inflamed tumor (representative of half of human NSCLC) is reminiscent of ex-T_reg that develop in autoimmune settings and can exacerbate auto-inflammatory responses. Although these cells have an established functional role in the exacerbation of autoimmune responses, their contribution to antitumor immunity has not previously been studied. Apropos of this notion, the instability of FoxP3 expression by T_reg is well established (81), but its association with ex-T_reg phenotype and function has not been previously reported. In the present study, we find that virtually all the TAA-specific T_reg make this conversion in the MC38 tumors over a 2-week period, in contrast to the minority of T_reg reported to convert to ex-T_reg in autoimmunity settings (48, 49, 51–53). T_H1-like Tbet⁺ T_reg have been shown to develop in inflamed tissues (79). The expression of Tbet and IFN-γ by T_reg is required for the development of autoimmune colitis in murine models (59) and human multiple sclerosis (60), and this subset is enriched in patients with type 1 diabetes (56). These prior findings are consistent with the higher incidence of immune-related adverse events in immunotherapy-sensitive tumors (82). Together with the previous literature, our findings lead us to hypothesize that TR-T_reg may contribute positively to antitumor immunity while also facilitating development of therapy-induced autoimmunity.

By generating a gene profile for these tumor-specific T_reg in the mouse, we identified a human TIL-T_reg ortholog, which, in contrast to the OX40^hiGITR^hi subset, is more highly represented among anti–PD-1–sensitive tumors. We acknowledge that mouse models cannot recapitulate human biology; however, the presence of distinct T_reg subsets with putatively opposing function regarding antitumor immunity begs the key question of which signals in the TME drive incoming T_reg to one or the other population. Elucidation of these signals will likely define important targets for immunotherapy intervention. Another key endeavor is definition of the antigen specificity of the human T_reg subsets. In contrast to a recent study of melanoma TIL-T_reg (83), our analysis of the more highly represented clones in the T_H1-like TIL-T_reg subset via TCR gene transfer and screening of predicted MHC2-binding neoantigenic peptides failed to reveal any such recognition (fig. S10G and tables S15 and S16). Fundamentally, one might expect that tumor recognition by T_reg would preferentially target self-antigens rather than mutation-associated neoantigens, because the natural T_reg repertoire is positively selected on self-antigens in the thymus (44, 45). Because TAA-specific T_reg in the murine model differentiate exclusively into a T_H1-like program rather than the activated OX40^hiGITR^hi program, we postulate that this suppressive subset may enter the tumor having already developed a suppressive program because of recognition of self-antigen not expressed by the tumor. The markedly different clonality between melanoma and lung cancer concurs with the apparent differences in frequency of neoantigen-specific T_reg observed by Oliveira et al. (83) in melanoma as opposed to our findings in lung cancer. We postulate that the distinct repertoires of T_reg antigen recognition in melanoma versus lung cancer reflect the different microenvironmental cues between the two tumor types, as witnessed by their different responses to immunotherapy.

Our in vitro overexpression system used to study TNFR super-family ligation on T_reg has limitations because of the potential for supraphysiologic signaling; thus, follow-up studies will be necessary to confirm the impact of in vivo OX40 and GITR ligation on immunosuppressive T_reg. In addition, it is thus far unknown how OX40^hi− GITR^hi T_reg are induced or recruited mainly in NR patients. Given that this population is more frequent in NR and untreated tumors relative to R tumors, it is possible that this is a tumor-intrinsic characteristic unrelated to ICB treatment; however, our study is limited by the lack of paired pre- and post-ICB tumor specimens from the same patients that would enable us to make this distinction. Although others have reported on the association of generally high expression of TNFRSF4 and TNFRSF18 among T_reg with immunosuppression and poor prognosis (7, 29, 30, 40, 41, 79, 80, 84–86), the coexpression of LAG3 and the association of this specific subset with ICB resistance have not previously been identified. Our conclusions are that these “Activated T_reg” genes and the presence of this immunosuppressive T_reg subset may be used as a prognostic marker; however, additional investigation is warranted to evaluate the specific mechanistic role they may play in the lack of response to anti–PD-1. Regarding the T_H1-like T_reg, although they exhibited higher expression of TBX21 than all other T_reg subsets, the data shown here do not distinguish whether TBX21 expression drives the T_reg transition to a more proinflammatory phenotype or is up-regulated as a consequence of other signals that drive this transition in the TME. Follow-up work should focus on understanding the role of Tbet and other transcriptional regulators in differential T_reg development. Last, we are limited by the number of human TIL-T_reg available to perform additional functional studies; thus, future work should aim to fully understand the function of these subsets as it relates to anti–PD-1 response. Nonetheless, a major strength of our study is the number of TIL-T_reg profiled at the single-cell level, which enabled us to study refined T_reg subsets at unprecedented resolution.

The primary goal of this study was to better understand the associations of T_reg functional programming, especially those that are actively recognizing tumor antigens, with response to ICB in lung cancer. Our findings regarding the strong proliferative and T_reg activation programs among the NSCLC TIL-T_reg induced by OX40L, and previous evidence that OX40 signaling stabilizes T_reg (34), support the notion that tumor-specific CD8⁺ T cells have the potential to directly enhance T_reg suppression as a form of feedback inhibition. If so, the agonistic OX40 antibodies under clinical investigation may be counterproductive; rather, OX40 pathway blockade might be considered as an immunotherapy strategy. Moreover, as combination ICB becomes more standard, it is likely that different therapeutic regimens will differentially affect TIL-T_reg function, and we postulate that targeted T_reg modulation, i.e., specific inhibition of highly suppressive subsets or promotion of T_H1-like T_reg development, could significantly improve ICB response while decreasing toxicities related to T_reg-mediated immune-related adverse events.

MATERIALS AND METHODS

Study design

The overall goal of this study was to better understand the associations of T_reg functional programming, especially those that are actively recognizing tumor antigens, with response to ICB in lung cancer. The sample size for human lung cancer TIL studies was limited by the number of patients enrolled in the trial from whom viable TILs were obtained. Tumors from murine experiments were randomized by tumor size on day 7 and were combined for transcriptomic analysis. The number of mice used/tumors harvested per experimental condition was predicated on requiring >30,000 TIL-T_reg from all pooled samples per experimental condition. Preliminary studies determined that to achieve this number of TIL-T_reg, 5 to 10 mice per experimental group were required. The investigators were not blinded to allocation during experiments and outcome assessment.

Patients and biospecimens

This study was approved by the Institutional Review Boards (IRBs) at Johns Hopkins University and Memorial Sloan Kettering Cancer Center (IRB Number NA_00092076) and was conducted in accordance with the Declaration of Helsinki and the International Conference on Harmonization Good Clinical Practice guidelines. The patients described in this study provided written informed consent. All biospecimens were obtained from patients with stage I to IIIA NSCLC who were enrolled in a phase 2 clinical trial (NCT02259621) evaluating the safety and feasibility of administering two doses of anti–PD-1 (nivolumab) before surgical resection. Pathological response of primary tumors were done as previously reported (17, 87). Tumors with no more than 10% residual tumor cells were considered to have an MPR.

Mice

SMARTA-1 mice [B6.Cg-Ptprc^a Pepc^b Tg(TcrLCMV)1Aox/PpmJ] have a CD4⁺ T cell repertoire that recognizes the H2-IAb–restricted LCMV GP61–80 epitope as previously described (46). To generate mice with a sufficient number of T_reg, SMARTA-1 mice were crossed to RIP-GP mice that have the LCMV GP protein expressed under the RIP, allowing for expression of these epitopes in the β cells of the pancreas and in the thymus for positive selection. This cross allows for thymic generation of SMARTA-1 T_reg as described (49, 51). SMARTA-1;RIP-GP mice were then crossed to RFP-FoxP3 reporter mice (C57BL/6-Foxp3^tm1Flv/J). These mice were used for adoptive transfer into C57BL/6J wild-type mice. All mice used in this study, donors and recipients, were females 6 to 8 weeks of age. SMARTA-1, RFP-FoxP3, and C57BL/6J were obtained from Jackson Laboratories (Bar Harbor, ME). RIP-GP mice were obtained from the laboratory of M. von Herrath (La Jolla Institute, CA). All mice were housed and bred under specific pathogen–free conditions at Johns Hopkins University Animal Care and Use Facility in Cancer Research Building I under protocol number M022M67. The Institutional Animal Care and Use Facility at Johns Hopkins University approved all animal experiments. All mouse models are available by request.

Tumor cell lines

Parental MC38 (MC38WT; American Type Culture Collection) and MC38-GP (Bosselut lab, National Institutes of Health) tumor lines were grown in Dulbecco’s modified Eagle’s medium (DMEM) 10% fetal bovine serum (FBS), 10 mM Hepes (Gibco), L-glutamine (Gibco), penicillin (100 U/ml), and streptomycin (100 μg/ml; Quality Biologic). Cells were passaged once using 0.05% trypsin before flank injection. Cells (6 × 10⁵) were injected in 100 μl of phosphate-buffered saline (PBS) into each mouse flank using Intradermal TB needles (BD, 309623).

Tumor cell line growth and response to anti–PD-1

MC38WT and MC38-GP tumor lines were injected into C57BL/6J mice as above. Mice were randomized into anti–PD-1–treated and anti–PD-1–untreated (isotype control) groups on day 8 of tumor growth. The anti–PD-1 treatment course began 2 days after tumor volume reached an average of ~100 mm³ (MC38WT: days 10, 12, 14, and 16; MC38-GP: days 14, 16, 18, and 20). One hundred micrograms of in vivo monoclonal antibody (mAb) anti-mouse PD-1 (BioXCell; BE0146) or in vivo mAb rat IgG2a isotype control (BioXCell; BE0089) was injected intraperitoneally into each mouse every other day for a total of four doses. Tumor growth was measured every other day for the duration of the experiment. Pre-treatment time point is day −2 of first anti–PD-1 dose, and post-treatment time point is day +4 of the last anti–PD-1 dose.

Coupled single-cell TCRseq/RNA-seq

Cryobanked T cells were thawed and washed twice with pre-warmed RPMI with 20% FBS and gentamicin. Cells were resuspended in PBS and stained with a viability marker (LIVE/DEAD Fixable Near-IR; Thermo Fisher Scientific) for 15 min at room temperature in the dark. Cells were then incubated with Fc block for 15 min on ice and stained with antibody against CD3 (BV421, clone SK7) for 30 min on ice. After staining, highly viable CD3⁺ T cells were sorted into 0.04% bovine serum albumin in PBS using a BD FACSAria II Cell Sorter. Sorted cells were manually counted using a hemocytometer and prepared at the desired cell concentration (1000 cells/μl) when possible. The Single-Cell 5’ V(D)J and 5’ digital gene expression (DGE) kits (10X Genomics) were used to capture immune repertoire information and gene expression from the same cell in an emulsion-based protocol at the single-cell level. Cells and barcoded gel beads were partitioned into nanoliter-scale droplets using the 10X Genomics Chromium platform to partition up to 10,000 cells per sample followed by RNA capture and cell-barcoded cDNA synthesis using the manufacturer’s standard protocols. Libraries were generated and sequenced on an Illumina NovaSeq instrument using 2 × 150-bp paired end sequencing. The 5’ VDJ libraries were sequenced to a depth of ~5000 reads per cell, for a total of 5 million to 25 million reads. The 5’ DGE libraries were sequenced to a target depth of ~50,000 reads per cell. Further detail can be found in Supplementary Materials and Methods.

Jurkat reporter cell line

A gBlock was created with the full TCRα and TCRβ chains separately and human constant regions and was synthesized [Integrated DNA Technologies (IDT)]. To generate a Jurkat reporter cell in which we could transfer our TCRs of interest, the endogenous TCR α and β chains were knocked out of a specific Jurkat line that contains a luciferase reporter driven by a nuclear factor of activated T cells (NFAT)–response element (Promega) using the Alt-R CRISPR system (IDT). Two sequential rounds of CRISPR knockout were performed using crDNA targeting the TCRα constant region (AGAGTCTCTCAGCTGGTACA) and the TCRβ constant region (AGAAGGTGGCCGAGACCCTC). crDNA and tracrRNA (IDT) were resuspended at 100 μM with Nuclear-Free Duplex Buffer. They were duplexed at a 1:1 molar ratio according to the manufacturer’s instructions. The duplexed RNA was cooled to room temperature before mixing with Cas9 nuclease at a 1.2:1 molar ratio for 15 min. Forty picomoles of Cas9 RNP complexed with gRNA was mixed with 500,000 cells in 20 μl of OptiMEM, loaded into a 0.1-cm cuvette (Bio-Rad), and electroporated at 90 V and 15 ms using an ECM 2001 (BTX, Holliston, MA). Cells were transferred to complete growth medium and expanded for 7 days. Limiting dilution was used to acquire single-cell clones, and genomic DNA was harvested using the Quick-DNA 96 Kit (Zymo Research, Irvine, CA). The regions flanking the CRISPIR cut sites were polymerase chain reaction–amplified (TCRα forward primer: GCCTAAGTTGGGGAGACCAC, reverse primer: GAAGCAAGGAAACAGCCTGC; TCRβ forward primer: TCGCTGTGTTTGAGCCATCAGA, reverse primer: ATGAACCACAGGTGCCCAATTC) and Sanger-sequenced. Only TCRα⁻/β⁻ clones were selected. Complete knockout was confirmed by failure to restore CD3 expression on electroporation with only a TCRα or TCRβ chain and successful CD3 expression on electroporation with both TCR chains.

CD8 was transduced into the TCRα⁻/β⁻ Jurkat reporter cells using the murine stem cell virus (MSCV) retroviral expression system (Clontech). gBlocks (IDT) encoding CD8α and CD8β chains separated by a T2A self-cleaving peptide were cloned into the pMSCVpuro retroviral vector by HiFi DNA assembly (New England Biolabs). The plasmid was then cotransfected with a pVSV-G envelope vector into the GP2–293 packaging cell line per the manufacturer’s instructions. The viral supernatant was harvested 48 hours after transfection and concentrated 20-fold using Retro-X Concentrator (Clontech). For transduction, nontissue cultured treated 48-well plates were coated with 150 μl of retronectin (Clontech) in PBS at 10 μg/ml overnight at 4°C. Plates were then blocked with 10% FBS for 1 hour at room temperarture followed by washing once with PBS. After removing PBS, viral particles and 2 × 10⁵ of TCRα⁻/β⁻ Jurkat reporter cells were added to each well in a total volume of 500 μl of cell culture media. Plates were spun at 2000g for 1 hour at 20°C and then incubated at 37°C. Selection with puromycin (1 μg/ml; Thermo Fisher Scientific) began 3 days later. Single-cell clones were established by limiting dilution, and clones were subsequently screened for CD8 expression by flow cytometry. To generate a Jurkat reporter line that expresses both CD4 and CD8, CD4 viral particles were produced and transduced into the CD8-expressing Jurkat reporter cells using similar procedures.

Jurkat TCR transfer

TCRs of interest were introduced into the CD4/CD8 TCRα⁻/β⁻ Jurkat reporter line by cloning the TCRα and TCRβ chains separately into the pCI vector (Promega) by HiFi DNA assembly (New England Biolabs). The two plasmids were co-electroporated into the TCRα⁻/β⁻ Jurkat reporter line using 4-mm cuvettes (Bio-Rad) and 275 V for 10 ms for three pulses at a 0.1 interval between pulses. Cells were rested in RPMI 10% FBS at 37° for 24 hours. TCR expression efficiency was assessed by CD3 expression using flow cytometry. After a 24-hour rest, live Jurkat cells were counted and plated at a 1:1 ratio with a patient-matched lymphoblastoid cell line and peptide pools (JPT Peptide Technologies and Sigma-Aldrich). Neoantigen peptides were plated and pooled (three to five peptides per pool) at 50 μg/ml per peptide to assess TCR reactivity to neoantigens. A peptide pool (20 μg/ml) spanning the S protein of SARS-CoV-2 (JPT) was used to stimulate a known SARS-CoV-2–reactive TCR as positive control (88). Cells and peptide were cocultured for 24 hours. TCR activity was assessed by NFAT-luciferase reporter readout using the Bio-Glo Luciferase Assay System (Promega).

Statistics

All statistics were performed in R (version 4.2.1) and GraphPad Prism (version 9.3.0). Wilcoxon signed rank test or t test (Student’s and paired) was used to determine significance when relevant. Benjamini-Hochberg adjustment was done when needed. Specific statistical information for each experiment is found in the relevant figure legends and Supplementary Methods section. Additional methodology can be found in Supplementary Materials and Methods.