Identification of human exT_reg cells as CD16⁺CD56⁺ cytotoxic CD4⁺ T cells

Abstract

In atherosclerosis, some regulatory T (T_reg) cells become exT_reg cells. We crossed inducible T_reg and exT_reg cell lineage-tracker mice (FoxP3^{eGFP−Cre-ERT2} ROSA26^{CAG-fl-stop-fl-tdTomato}) to atherosclerosis-prone Apoe^−/− mice, sorted T_reg cells and exT_reg cells and determined their transcriptomes by bulk RNA sequencing (RNA-seq). Genes that were differentially expressed between mouse T_reg cells and exT_reg cells and filtered for their presence in a human single-cell RNA-sequencing (scRNA-seq) panel identified exT_reg cell signature genes as CST7, NKG7, GZMA, PRF1, TBX21 and CCL4. Projecting these genes onto the human scRNA-seq with CITE-seq data identified human exT_reg cells as CD3⁺CD4⁺CD16⁺CD56⁺, which was validated by flow cytometry. Bulk RNA-seq of sorted human exT_reg cells identified them as inflammatory and cytotoxic CD4⁺T cells that were significantly distinct from both natural killer and T_reg cells. DNA sequencing for T cell receptor-β showed clonal expansion of T_reg cell CDR3 sequences in exT_reg cells. Cytotoxicity was functionally demonstrated in cell killing and CD107a degranulation assays, which identifies human exT_reg cells as cytotoxic CD4⁺T cells.

Atherosclerosis is a chronic inflammatory disease of the artery walls with clear evidence of a systemic CD4⁺T cell-mediated autoimmune response to apolipoprotein B (ApoB) epitopes in mice^1–5 and humans^3,6,7. T_reg cells play a protective role via anti-inflammatory cytokines and contact-dependent mechanisms^8,9 and represent a promising therapeutic tool. Under inflammatory conditions, T_reg cells become plastic, that is, they acquire lineage-defining transcription factors in addition to FoxP3 (called T_H-like T_reg cells)^10–12, like T-bet for T_H1 or RORγt for T_H17. In chronic inflammation, T_reg cells become unstable and lose CD25 and FoxP3 expression¹³. Such exT_reg cells have been identified in mouse disease models such as in type 1 diabetes¹⁴, experimentally induced autoimmune encephalomyelitis¹⁵, arthritis¹⁶ and atherosclerosis^4,7,10,11,17. In some studies, exT_reg cells have been reported to express RORγT and produce interleukin (IL)-17A^14,16,18, or T-bet and interferon (IFN)-γ^14,15,19 or Bcl6 and CXCR5 (ref. 17). However, complete transcriptomes of exT_reg cells have not been reported. Here, we use the inducible T_reg cell lineage-tracker mouse strain FoxP3-eGFP⁻Cre-ERT2 (ref. 20) ROSA26CAG-fl-stop-fl-tdTomato, crossed into atherosclerosis-prone Apoe^−/− mice. Tamoxifen injection induces the generation of tdTomato and GFP-expressing Foxp3⁺ T_reg cells. Converted exT_reg cells are detectable as GFP⁻tdTomato⁺CD4⁺T cells.

Human exT_reg cells are suspected to exist, but have not been described. In a previous study, transcriptomic analysis of tetramer-sorted APOB-specific human CD4⁺ T cells, which are expected to contain T_reg cells and exT_reg cells based on the mouse studies^4,7,11,17, revealed that exT_reg cells are not similar to T_reg cells⁷. In fact, exT_reg cells mapped widely across a uniform manifold approximation and projection (UMAP) plot constructed of T_reg cells, T_H1 and memory T cells. The present study was designed to identify human exT_reg cells by using an integrated approach of mapping gene signatures from sorted mouse exT_reg cell transcriptomes to human scRNA-seq with cellular indexing of transcriptomes and epitopes sequencing (CITE-seq) data, followed by T cell antigen receptor sequencing (TCR-seq), bulk RNA-seq and functional characterization of sorted human exT_reg cells. Bulk RNA-seq of sorted T_reg cells and exT_reg cells from the lineage-tracker mice on the Apoe^−/− background yielded high-quality transcriptomes identifying hundreds of differentially expressed genes (DEGs). Then, we extracted gene signatures from exT_reg cells and mapped them to a published and publicly available human peripheral blood mononuclear cell (hPBMC) scRNA-seq dataset²¹. This dataset also had information on cell surface phenotype by CITE-seq. This identified surface markers on CD4⁺T cells that expressed the exT_reg cell signature genes. These were in turn used to sort the putative exT_reg cells from human blood, obtain high-quality, deep transcriptomes and assess their function. To identify the provenance of human exT_reg cells, we used TCRβ sequencing. We reasoned that exT_reg cells derived from T_reg cells might show clonal expansion of TCRβ sequences found in T_reg cells. TCR diversity was assessed by analyzing TCR CDR3 sequences directly and using GLIPH²².

Results

Differentially expressed mouse exT_reg and T_reg cell-classifying genes

To track exT_reg cells in mice under conditions of chronic sterile inflammation, we used a recently published mouse model⁷, in which we crossed an inducible FoxP3-Cre lineage-tracker mouse strain (FoxP3^{eGFP-Cre-ERT2}ROSA26^{CAG-fl-stop-fl-tdTomato}) into the atherosclerosis-prone Apoe^−/− mouse strain. Both strains are in a C57BL/6 background. Apoe^−/− mice develop atherosclerosis with chronic inflammation, which can trigger T_reg cell instability and exT_reg cell formation. In tamoxifen-injected mice, GFP⁺tdTomato⁺ T_reg cells and GFP⁻tdTomato⁺ exT_reg cells are detected as distinct CD4⁺T cell subsets, with negligible GFP⁺ or tdTomato⁺ cells among control non-T cells and CD4⁻ T cells (Extended Data Fig. 1a). We quantified the frequencies of T_reg cells (GFP⁺tdTomato⁺) and exT_reg cells (GFP⁻tdTomato⁺) in the spleen and lymph nodes (LNs) of these mice at 4, 8, 12 and 20 weeks of age after tamoxifen injection (Fig. 1a). At 4 weeks, exT_reg cells accounted for <1% of total CD4⁺T cells in either organ. They gradually accumulated over time, reaching ~7% in spleen and ~4% in LNs at 20 weeks. As previously reported^10,14, we observed significantly lower surface expression of CD25 in exT_reg cells (Fig. 1b) than in T_reg cells.

Fig. 1 |. Deep transcriptomes from mouse exT_reg cells and T_reg cells identify differentially expressed candidate genes.

a, Frequency of mouse exT_reg cells (red) and T_reg cells (blue) among all CD4⁺T cells in spleen and LNs of FoxP3^{eGFP-Cre-ERT2}ROSA26^{CAG-fl-stop-fl-tdTomato} Apoe^−/− mice at 4 (T_reg n = 7; exT_reg n = 5), 8 (T_reg n = 4; exT_reg n = 4), 12 (T_reg n = 5; exT_reg n = 5) and 20 (T_reg n = 6; exT_reg n = 5) weeks after tamoxifen injection. All mice were on regular CD. b, Median fluorescence intensity (MFI) of CD25 expression on exT_reg cells (red circles), T_reg cells (blue circles) and conventional CD4⁺T cells (T_con, black circles) in spleen (n = 4) and LNs (n = 6) from 16-week-old FoxP3^{eGFP-Cre-ERT2}ROSA26^{CAG-fl-stop-fl-tdTomato} Apoe^−/− mice on CD. Results (a and b) are represented as the mean ± s.e.m. Spleen (b) *P = 0.0286; LN (b) **P = 0.0022; two-tailed Mann–Whitney U test. c, PCA of bulk RNA-seq data from sorted mouse exT_reg cells and T_reg cells from spleen and LNs of 20-week-old FoxP3^{eGFP-Cre-ERT2}ROSA26^{CAG-fl-stop-fl-tdTomato} Apoe^−/− mice on CD. Results (a–c) are from independent biological replicates. d, Volcano plot of differentially expressed mouse exT_reg and T_reg cell-classifying genes, identified by the SVM trained on mouse transcripts with human orthologs and filtered for those present in the human scRNA-seq targeted gene panel. y and x axes capped at 20 (P = 10⁻²⁰) and ±4 (log₂FC), respectively. Horizontal line at −log₁₀ (P adjusted) = 1.3 (same as P_adj = 0.05). Statistical analyses were performed using a two-tailed Wald test with Benjamini–Hochberg P-value adjustment. e, Mouse aortas with carotid artery branches from FoxP3^{eGFP-Cre-ERT2}ROSA26^{CAG-fl-stop-fl-tdTomato} Apoe^−/− mice were fixed and imaged using a Leica SP8 multiphoton microscope. GFP⁺ T_reg cells (pseudocolored green) and tdTomato⁺ exT_reg cells (pseudocolored pink) in the adventitia (top). Bottom image is a zoomed-in view of the white box. Blue-green indicates a second harmonic generation microscopy analysis of extracellular matrix. Scale bar, 100 μm. Data are representative of four independent experiments.

Mice of both sexes were maintained on a regular chow diet (CD) and LNs and spleens were collected at 20 weeks after tamoxifen injection. T_reg cells (TCRβ⁺CD4⁺GFP⁺tdTomato⁺) and exT_reg cells (TCRβ⁺C D4⁺GFP⁻tdTomato⁺) were sorted by flow cytometry (sorting scheme as previously described⁷) and sequenced using an optimized bulk RNA-seq method. In the principal component analysis (PCA), principal component (PC) 1 distinguished between T_reg cells and exT_reg cells and PC2 distinguished between the site of origin (LNs or spleens; Fig. 1c). The first two components explained the majority (86%) of the transcriptome variance. The mouse genes were filtered for human orthologs and intersected with a published dataset of 496 human genes that were analyzed by scRNA-seq with CITE-seq²¹. Mouse orthologs were found for 383 of the 496 human genes (Supplementary Table 1). Expression of these 383 genes in the transcriptomes from sorted mouse T_reg cells and exT_reg cells were used to train a support vector machine (SVM) model. T_reg cell versus exT_reg cell genes were ranked by weight and the top 60 classifying genes were identified for each subset (Supplementary Table 2). As expected, the top 60 genes classifying T_reg cells included Foxp3 and Il2ra (encoding CD25; Extended Data Fig. 1b). The top 60 genes classifying exT_reg cells included Tbx21, Gzmk, Prf1, Nkg7, Ifng and Ccl4 (Extended Data Fig. 1b). Most (50 of 60 T_reg cell genes and 51 of 60 exT_reg cell genes) were significantly (P adjusted < 0.05) differentially expressed between mouse T_reg cells and exT_reg cells (Fig. 1d, statistics in Supplementary Table 3). T_reg cells and exT_reg cells were also found in the artery wall (Fig. 1e), but numbers were insufficient for cell sorting.

exT_reg cell-specific genes and surface markers in human CD4⁺ T cells

Mouse exT_reg DEGs (Fig. 1d, up in exT_reg cells) were projected onto the UMAP projections of a published scRNA-seq transcriptome and CITE-seq dataset from 40,821 CD4⁺ T cells from 61 participants in the Coronary Assessment in Virginia (CAVA) cohort²¹. This cohort contains men and women, aged from 40 to 80 years, with coronary artery disease (CAD) quantified by angiography (Gensini scores), where Gensini scores <6 were considered CAD⁻ and scores >30 were considered CAD⁺. A full description of the cohort is published²¹. Statistical significance (P < 0.05) of exT_reg cell gene enrichment (pct.1 and log₂ fold change) in each CD4⁺ T cell cluster versus the other CD4⁺ T clusters, (one-versus-all) was evaluated. Six genes (CST7, NKG7, GZMA, PRF1, TBX21, CCL4) showed high and specific expression in cluster CD4T_7 (encompassing 1,359 cells), which was previously classified as cytotoxic effector memory cells²¹ (Fig. 2a). In the UMAP projections of human CD4⁺ T cells, each of the six exT_reg cell classifier genes was most highly expressed in cells that mapped to CD4T_7 (Extended Data Fig. 2a). The FOXP3-expressing T_reg cells resided in cluster CD4T_17, which was previously identified as the T_reg cluster²¹ (Fig. 2b). The fraction of cells expressing the exT_reg candidate genes was significantly higher in cluster 7 than in T_reg cells for all six genes (Fig. 2c).

Fig. 2 |. Mouse exT_reg cell classifier genes identify human exT_reg cell candidate genes and surface markers in a human scRNA-seq and CITE-seq dataset of CD4⁺T cells.

a, Differentially expressed mouse exT_reg classifier genes were examined among all CD4⁺ T cell clusters from a published single-cell human dataset from 61 men and women (aged 40–80 years). Statistical significance for the enrichment of each gene in one cluster versus all others was assessed. Average log₂ fold change (log₂FC, dot color) and positive cell proportions (pct.1, dot size) for significantly (P < 0.05) enriched genes are shown. Six highly expressed exT_reg genes (CST7, NKG7, GZMA, PRF1, TBX21 and CCL4) enriched in human cluster CD4T_7 (red boxes). b, Feature plot showing expression of FOXP3 gene (blue dots) in the human CD4⁺ T single-cell dataset. T_reg cell cluster CD4T_17, previously identified as T_reg cells, highlighted. c, Frequencies of cells that expressed each of the exT_reg cell signature genes in CD4T_7 (red circles) and CD4T_17 (T_reg cells, blue circles). n = 61. d, Cells expressing optimal exT_reg cell candidate gene combination (CST7 + NKG7 + GZMA + PRF1 + TBX21) are highlighted in red on UMAP embeddings of CD4⁺ T clusters from the scRNA-seq data. All other cells are in gray. e, Differentially expressed surface markers (CITE-seq antibodies) on cells expressing candidate gene combinations. Average log₂FC (dot color) and log₂(pct.1/pct.₂) × avg_log₂FC (dot size) for significant (P < 0.05) differentially expressed protein markers on candidate cells versus CD4T_17 (T_reg cells) are shown. Enriched exT_reg markers CD127, CD16 and CD56 are marked with red boxes. The second-to-last combination is candidate 5 (CST7 + NKG7 + GZMA + PRF1 + TBX21). f, UMAP embeddings of CD4⁺ T clusters. Black outline marks cluster CD4T_7; cells that coexpressed CD16 and CD56 are shown as red dots. g, Representative plots from flow cytometry (FACS) showing CD16⁺CD56⁺CD4⁺T cells (exT_reg cells, red) and CD25⁺CD127^loCD4⁺T cells (T_reg cells, blue) with FMO controls. h, MFI of CD16 (n = 6), CD56 (n = 6), CD127 (n = 3) and FOXP3 (n = 6) expression in exT_reg cells (red circles) and T_reg cells (blue circles). 20% male, 80% female donors; ages 23–64 years. Results (c and h) are represented as the mean ± s.e.m. Each dot (c and h) represents a biological replicate from an independent human donor. Statistical comparisons using a two-tailed Mann–Whitney U test (c and h) and a two-tailed Wilcoxon’s rank-sum test with Benjamini–Hochberg correction for multiple comparisons (a and e). In c, ****P < 0.000000000000001 for CST7, NKG7, GZMA, PRF1 and ****P = 3.86 × 10⁻¹³ for TBX21. In h, *P = 0.0121, **P = 0.0022.

Having learned that cluster 7 likely contains human exT_reg cells, we wished to deconvolve their identities at single-cell resolution. Serial combinations of the six candidate exT_reg markers improved specificity and identified consecutively smaller CD4⁺ T cell populations (Supplementary Table 4). Coexpression of CST7 and NKG7 was found in 9.4% of all CD4⁺ T cells; coexpression of CST7 and GZMA in 8.7%; CST7, NKG7 and GZMA in 7.1%; CST7, NKG7, GZMA and PRF1 in 3.5%; CST7, NKG7, GZMA, PRF1 and TBX21 in 1.1% (candidate 5 in Fig. 2d) and CST7, NKG7, GZMA, PRF1, TBX21 and CCL4 in 0.3% (candidates 1–4 and 6 in Extended Data Fig. 2b).

We considered exT_reg cell candidate 5 (Fig. 2d) optimal, because this signature identified no cells in the T_reg cell cluster CD4T_17 and showed negligible expression in other CD4⁺ T cell clusters.

Next, we studied the cell surface phenotype of the candidate human exT_reg cells, based on the CITE-seq data (51 antibodies). Analysis of differentially expressed surface markers on candidate exT_reg cells versus T_reg cells (Fig. 2e) revealed statistically significant (P < 0.05) overexpression of CD16, CD56 and CD127 (Fig. 2e). Projecting CD56 and CD16 surface expression onto the UMAP revealed that CD56 protein was expressed on the surface of 1,681 cells (4.1% of all CD4⁺ T cells; Extended Data Fig. 2c). CD16 was expressed on fewer cells (109 cells, 0.3% of all CD4⁺ T cells; Extended Data Fig. 2c). Interestingly, almost all CD16⁺ cells coexpressed CD56 (94 of 109 cells; Fig. 2f). This analysis thus identified two markers by which candidate human exT_reg cells could be sorted: CD56 and CD16.

Having identified CD56 and CD16 as surface markers for candidate human exT_reg cells, we conducted flow cytometry on PBMCs and found that some CD4⁺T cells coexpressed CD16 and CD56 protein on their surface (Fig. 2g, with fluorescence-minus-one (FMO) controls; more gating strategy in Extended Data Fig. 2d). T_reg cells were identified as CD25⁺CD127^loCD4⁺T cells (Fig. 2g, with FMOs). Quantitative analysis by flow cytometry showed that exT_reg cells, unlike T_reg cells, expressed high levels of CD16, CD56 and CD127, and low levels of FOXP3 protein (Fig. 2h). We sorted CD16⁺CD56⁺ exT_reg cells and CD25⁺CD127^lo T_reg cells, prepared cDNA and conducted quantitative PCR with reverse transcription (RT–qPCR) for the six exT_reg cell signature genes (CST7, NKG7, GZMA, PRF1, TBX21 and CCL4). Normalized transcript expression levels of all these genes were significantly upregulated in exT_reg cells compared to T_reg cells (Extended Data Fig. 2e).

Deep transcriptomes of human CD16⁺CD56⁺ exT_reg cells

Having established that human exT_reg cells can be isolated using CD16 and CD56, we next sorted CD4⁺CD16⁺CD56⁺ exT_reg cells from seven donors and compared their transcriptomes with CD4⁺CD25⁺CD127^lo T_reg cells (sorting scheme in Extended Data Fig. 3a). PCA revealed that exT_reg cells clustered far from T_reg cells (Fig. 3a). PC1 (56% of variance) was the main driver of this separation (Fig. 3a and Supplementary Table 5). Human exT_reg cells had lost FOXP3 and IL2RA expression (Fig. 3b). They significantly gained cytotoxic and inflammatory genes like NKG7, TBX21, CCL3, CCL4 and CCL5, all of which were not expressed or present at low levels in T_reg cells (Fig. 3b). This pattern of gene expression was also seen in the mouse dataset (Extended Data Fig. 4a).

Fig. 3 |. Deep transcriptomes from sorted human CD3⁺CD4⁺CD16⁺CD56⁺ exT_reg cells contrasted with T_reg cells and NK cells.

a, PCA plot of bulk RNA-seq data from sorted human exT_reg (CD3⁺CD4⁺CD16⁺CD56⁺) and T_reg (CD3⁺CD4⁺CD25⁺CD127^lo) cells. n = 7. Donor details for human bulk RNA-seq in Supplementary Table 13. b,c, Comparative gene signature analysis between human exT_reg cells and T_reg cells (b) or NK (c) cells. Genes were filtered for significant differential expression, consistent in both mouse and human datasets (only human shown here). An external dataset was used for human NK cells: GSE133383 (samples GSM3907331, GSM3907341 and GSM3907351). Lowly expressed genes (<7 raw reads in all samples) in our dataset were filtered out. EdgeR was used to normalize the counts by applying the trimmed mean of M-values method and counts-per-million conversion. Analysis of DEGs was done using DESeq2. Curated lists of significant DEGs (log₂FC ± 1, adjusted P < 0.05) genes are shown on normalized heat maps, scaled by row (z scores). d, Representative FACS plots showing exT_reg cells as CD3⁺CD4⁺CD16⁺CD56⁺ T cells and NK cells as CD3⁻CD4⁻CD16⁺CD56⁺ non-T cells in hPBMCs. e, Histograms showing the fluorescence intensities of conjugated antibodies against TCRαβ, CCR7, CD127 and CD27 on NK cells (gray) and exT_reg cells (red). The scaled y axis was normalized to mode. f, MFI values of the specified markers (n = 12 for each) in NK cells (gray circles) and exT_reg cells (red circles) were plotted as the mean ± s.e.m. 40% male and 60% female donors, aged 20–69 years. Each dot represents a biological replicate from an independent human donor. Statistical comparisons were done using a two-tailed Wald test with Benjamini–Hochberg correction for multiple testing (b and c) and a two-tailed Mann–Whitney U test (f). ****P = 7.396 × 10⁻⁷ in f.

Gene-set enrichment analysis (GSEA) of human bulk exT_reg cell transcriptomes against the human scRNA-seq dataset confirmed their enrichment in CD4 cluster 7 (normalized enrichment score (NES) 1.82, false discovery rate q = 7.8 × 10⁻⁴, Extended Data Fig. 3b). Similarly, mouse exT_reg genes, filtered for human orthologs in the scRNA-seq panel, positively correlated with human exT_reg cell genes from bulk RNA-seq (NES = 1.497, false discovery rate q = 0.033, Extended Data Fig. 4b). This formally showed that the leading exT_reg cell gene signatures are significant and consistent among the three datasets (mouse bulk RNA-seq, human scRNA-seq, human bulk RNA-seq). Pathway analysis on all genes that were significantly upregulated in human exT_reg cells, compared to T_reg cells (Supplementary Table 6) revealed significant (adjusted P value < 0.05) enrichment of cellular processes related to TCR, Notch, cytokine and receptor-mediated signaling, metabolism, cytotoxicity, adhesion and hemostasis (Extended Data Fig. 3c).

In humans, CD16 and CD56 are canonical markers of natural killer (NK) cells. Unlike NK cells, human exT_reg cells expressed CD3 subunits (Fig. 3c,d). We compared our exT_reg cell transcriptomes from both mouse and human bulk RNA-seq datasets with publicly available NK cell data. DEGs were filtered for significance in both species (human in Fig. 3c; mouse in Extended Data Fig. 4c). The TCR-related signaling complex genes CD3D, CD3E and CD3G were all significantly higher in exT_reg cells than in NK cells. TCF7, encoding a transcription factor involved in T cell development and memory differentiation, was also significantly higher in exT_reg cells than in NK cells. CD27, known to be absent on mature human NK cells²³, was significantly expressed in ex T_reg cells. Conversely, KLRD1 (encoding CD94) and FCER1G were significantly higher in NK cells (Fig. 3c). The identity of CD3⁺CD4⁺CD16⁺CD56⁺ human exT_reg cells and CD3⁻CD4⁻CD16⁺CD56⁺ human NK cells was also confirmed by fluorescence-activated cell sorting (FACS; representative plots in Fig. 3d; gating scheme in Extended Data Fig. 3d). exT_reg cells, and not NK cells, expressed TCRαβ, CCR7, CD127 and CD27 proteins on the surface (representative histograms in Fig. 3e; quantification in Fig. 3f). Taken together, we found that the exT_reg cell gene signatures were conserved between mice and humans. Human CD3⁺CD4⁺CD16⁺CD56⁺ exT_reg cells are distinct from both T_reg cells and NK cells.

TCRβ sequencing shows clonal expansion of human exT_reg cells from T_reg cells

To study the provenance of human CD16⁺CD56⁺CD4⁺T cells, we used TCRB sequencing. We reasoned that if CD16⁺CD56⁺ T cells are exT_reg cells, their TCRB CDR3 sequences would overlap with TCRB CDR3 sequences in T_reg cells. Moreover, if exT_reg cells arise by clonal expansion from T_reg cells, we expected to see clonal enrichment. Thus, we sequenced TCRβ from sorted T_reg cells (CD25^hiCD127^lo) and exT_reg cells (CD16⁺CD56⁺) CD4⁺T cells from four healthy donors (sorting strategy was the same as that used for bulk RNA-seq, Extended Data Fig. 3a; number of templates and rearrangements in Supplementary Table 7). For comparison, we included naïve CD4⁺T cells (CD45RA⁺CCR7⁺). Simpson’s clonality was higher in exT_reg cells than in T_reg cells or naïve CD4⁺T cells (Fig. 4a). Morisita’s index, a measure of overlap between samples, showed a significantly higher frequency of shared rearrangements of exT_reg cells with T_reg cells than with naive T cells (Fig. 4b).

Fig. 4 |. Oligoclonal human exT_reg cells are clonally expanded from proliferating T_reg cells.

a, Productive Simpson’s clonality of TCRβ sequences from sorted naïve T cells (black circles), T_reg cells (blue circles) and exT_reg cells (red circles). n = 4. Donor details for human TCR-seq in Supplementary Table 13. b, Frequency of rearrangements shared between exT_reg cells and naïve or between exT_reg cells and T_reg cells was measured using Morisita’s index. n = 4. Results (a and b) are represented as the mean ± s.e.m. c, GLIPH2-analyzed conserved amino acid motifs in TCRβ sequences. Groups with exT_reg TCRs were filtered for statistically significant expansion score (P < 0.05). 178 of 345 expanded exT_reg GLIPH2 groups were shared by T_reg cells. Relative abundance of T_reg and exT_reg TCRs in these 178 T_reg/exT_reg groups was compared and DEG patterns (log₂FC ± 1, two-sided Poisson test P < 0.05) are shown as a volcano plot. Horizontal line at −log₁₀ (P value) = 1.3 (same as P = 0.05). Vertical lines at |log₂FC| = 1. d, Violin plots showing normalized expression levels (transcripts per million) of proliferation genes MKI67, TOP2A, CCNA2, CCNB1, CCNB2, MCM2, MCM3, MCM5 and MCM6 in human bulk transcriptomes from sorted human T_reg cells (blue dots) and exT_reg cells (red dots). n = 7. 33.33% male and 66.67% female donors, aged 21–54 years. e, Frequency of PD-1 (n = 16), GITR (n = 4), LAG3 (n = 16) and TIGIT (n = 16) expressing cells by FACS, percentage of parent (all CD4⁺T cells (black circles), exT_reg cells (red circles), T_reg cells (blue circles); mean ± s.e.m.). 50% male and 50% female donors, aged 20–69 years. Each dot represents a biological replicate from an independent donor. Statistical comparisons were done using one-way analysis of variance (ANOVA) with Dunnett’s multiple-comparison test (a), a two-tailed Mann–Whitney U test (b and d) and a Kruskal–Wallis test, adjusted with Dunn’s multiple-comparison testing (e). In a, exT_reg cells versus naive **P = 0.0012; versus T_reg cells **P = 0.0068. In b, *P = 0.0286. In d, *P = 0.0111 for MKI67, CCNB1; *P = 0.0175 for MCM6; **P = 0.007 for CCNA2 and MCM3; **P = 0.0012 for CCNB2; ***P = 0.0006 for TOP2A, MCM2, MCM5. In e, PD-1 **P = 0.0023, GITR **P = 0.0047, LAG3 ****P = 3.21 × 10⁻⁷, TIGIT ****P = 5.227 × 10⁻⁵ for exT_reg cells versus CD4⁺T cells; PD-1 *P = 0.0329, GITR P = 0.3396, LAG3 ****P = 1.659 × 10⁻⁵, TIGIT P = 0.423 for exT_reg cells versus T_reg cells. NS, not significant.

For a more detailed analysis, we used GLIPH²², a software that identifies conserved amino acid sequence patterns in CDR3 sequences. First, we filtered the GLIPH groups based on their enrichment (Fisher P < 0.05) in our dataset, as compared to a published set of CDR3 sequences from naïve T cells²². CDR3 sequence patterns with exT_reg TCRs were filtered for statistically significant expansion based on the productive frequencies of their TCRs (expansion score P < 0.05). We identified 345 expanded exT_reg cell groups. TCRβ sequences from T_reg cells were present in ~50% (178 of 345) of them (Supplementary Table 8). We compared the relative abundance of T_reg cell and exT_reg cell TCRβ sequences in these groups and found many TCRβ CDR3 patterns that were significantly expanded (log₂FC ± 1, Poisson test P < 0.05) in exT_reg cells compared to T_reg cells (Fig. 4c; GLIPH patterns in Supplementary Table 9).

The clonal expansion observed in exT_reg cells prompted us to reason that T_reg cells may have proliferated and differentiated into oligoclonal exT_reg cells. This was supported by our observation that the SVM model had identified genes related to DNA replication and cell cycle, such as Aurkb, Ccnd2, Mcm4, Mki67 and Top2a, as T_reg cell-classifying genes (Supplementary Table 2). These genes were significantly upregulated in mouse T_reg cells, compared to exT_reg cells (Extended Data Fig. 1b and Supplementary Table 3). To functionally compare T_reg cell versus exT_reg cell proliferation in atherosclerotic mice, we conducted an in vivo BrDU incorporation experiment in FoxP3^{eGFP-Cre-ERT2} ROSA26^{CAG-fl-stop-fl-tdTomato} Apoe^−/− mice (experimental design in Extended Data Fig. 5a). Proliferating T_reg cells and exT_reg cells (gating scheme, controls and representative plots in Extended Data Fig. 5b–d) were monitored based on Ki67 (encoded by Mki67) expression and BrDU incorporation. Frequencies of Ki67⁺BrDU⁺CD4⁺T cells were significantly higher among T_reg cells than exT_reg cells (quantification in Extended Data Fig. 5d), showing that mouse T_reg cells actively proliferate during atherosclerosis.

Next, we examined proliferation-related genes in our human bulk RNA-seq dataset (Fig. 3). We found that human T_reg cells, as compared to exT_reg cells, expressed significantly higher transcript levels of MKI67 (gene encoding Ki67), topoisomerase TOP2A, MCM2, MCM3, MCM5 and MCM6 (members of the replicative helicase that are involved in the formation of pre-replicative complexes) and cyclin A and B genes (which activate DNA replication and mitosis, respectively; Fig. 4d). These data are consistent with the idea that exT_reg cells may arise from T_reg cells by proliferation and clonal expansion.

To assess whether human exT_reg cells retain any phenotypic similarities with T_reg cells, we monitored expression of the T_reg cell markers PD-1, GITR, LAG3 and TIGIT on exT_reg cells using FACS (representative plots and FMO controls in Extended Data Fig. 6, quantification in Fig. 4e). All CD4⁺T cells were used as a negative control. exT_reg cells retained TIGIT and GITR, expressed significantly more PD-1 than T_reg cells, and gained expression of LAG3. We also filtered the transcriptomes of human exT_reg cells by known T_reg genes²⁴ and found 32 such genes that were retained in exT_reg cells (Supplementary Table 10).

Human exT_reg cells are not suppressive but cytotoxic

To test whether human exT_reg cells retained the suppressive capacity of T_reg cells, we conducted a suppression assay to compare the effects of T_reg cells and exT_reg cells on the proliferation of effector T cells (T_eff) cells. Co-culturing T_reg cells with T_eff cells at a 1:1 T_reg:T_eff cell ratio effectively suppressed proliferation of T_eff cells, as measured by carboxyfluorescein succinimidyl ester (CFSE) dilution (Fig. 5a). By contrast, T_eff cell proliferation in the presence of exT_reg cells (Fig. 5a) remained the same as proliferation of T_eff cells alone (Fig. 5a). T_eff cell proliferation was significantly higher when co-cultured with exT_reg cells than with T_reg cells (Fig. 5a) and not different from proliferation of T_eff cells alone.

Fig. 5 |. Human exT_reg cells are not suppressive but are cytotoxic.

a, CFSE-labeled CD3⁺CD4⁺CD25⁻T cells (T_eff) were co-cultured at 1:1 ratio with either CTV-labeled T_reg cells or exT_reg cells, in the presence of anti-CD3/CD28/CD2-coated beads. Left, representative histograms from three independent experiments showing CFSE in T_eff cells alone (black line), in T_eff cells cultured with T_reg cells (blue line) and in T_eff cells cultured with exT_reg cells (red line). Right, proportion of dividing T_eff cells after 5 d of co-culture with T_reg cells (blue circles, n = 5) or ex T_reg cells (red circles, n = 3). 40% male and 60% female donors; aged 23–43 years. b, Heat map of DEGs (P < 0.01 and |log₂FC| > 2, based on a two-tailed Wald test with Benjamini–Hochberg P-value adjustment) between human exT_reg versus T_reg transcriptomes from Fig. 3. Cytotoxic and T_reg signature genes are highlighted. c, FACS analysis of cytotoxic proteins FASLG, perforin and granzyme B (genes are labeled red in b) in human exT_reg cells, compared to bulk CD4⁺T cells. Left, representative histograms of intensities of fluorochrome-conjugated antibodies against FASLG, perforin and granzyme B on exT_reg cells (red) and all CD4⁺T cells (black). Right, MFIs of marker protein expression in exT_reg cells. n = 6. 50% male and 50% female donors, aged 25–37 years. d, Representative contour plots (left) and quantification (right) of basal and anti-CD3-induced degranulation in exT_reg cells (red) and NK cells (gray), as measured by surface mobilization of the degranulation marker CD107a by FACS. hPBMCs were co-cultured with uncoated P815 cells (target alone, open circles) or 5 μg ml⁻¹ anti-CD3-coated P815 cells (target + anti-CD3, filled circles) for 6 h at a 10:1 PBMC:P815 ratio. n = 3. 33.33% male and 66.67% female donors; aged 25–38 years. e, Anti-CD3-loaded P815 cells were co-cultured with CD8 CTLs (black circles, n = 5), T_reg cells (blue circles, n = 5) or exT_reg cells (red circles, n = 4) at a 1:5 P815:effector cells ratio for 16 h. 60% male and 40% female donors, aged 21–45 years. Cytotoxicity was assessed by measuring lactate dehydrogenase amounts in the supernatant. Results (a and c–e) are represented as the mean ± s.e.m. Each dot represents a biological replicate from an independent donor. Statistical comparisons were done using a two-tailed Mann–Whitney U test (a and e) and a two-tailed unpaired t-test (d). *P = 0.0357 (a). ***P = 0.0008 (d, exT_reg cells); P = 0.7707 (d; NKs). **P = 0.0079 (e, T_reg cells versus CTLs); *P = 0.0159 (e; T_reg cells versus exT_reg cells).

To delineate the functional profile of human exT_reg cells, we carefully examined their transcriptomes (Fig. 3) and contrasted them with T_reg cells (heat map of DEGs in Fig. 5b). We found that human exT_reg cells expressed many genes present in the KEGG list of cytotoxic genes. These included GZMB, PRF1 and FASLG, known to mediate effector functions in cytotoxic cells. Additionally, exT_reg cells had lost FOXP3, IL2RA, CTLA4, IL10 and LRRC32 gene expression, supporting their functional loss of suppressive capacity.

We validated intracellular expression of perforin (encoded by PRF1), granzyme B (encoded by GZMB) and FAS ligand (encoded by FASLG) in human exT_reg cells at the protein level by FACS. We detected high expression of perforin and granzyme B, but not FAS ligand on exT_reg cells (representative histograms in Fig. 5c; quantification in Fig. 5c). Although both exT_reg cells and NK cells expressed the cytotoxic marker perforin, only exT_reg cells upregulated the activation-induced CD4⁺T cell marker CD40L, upon stimulation with phorbol 12-myristate 13-acetate (PMA) and ionomycin (Extended Data Fig. 7). In a degranulation assay with P815 target cells, crosslinking with anti-CD3 triggered surface mobilization of the degranulation marker CD107a specifically in exT_reg cells and not in NK cells (representative histograms in Fig. 5d, quantification in Fig. 5d). These functional data further confirm that exT_reg cells are distinct from NK cells. To directly test the cytotoxic potential of exT_reg cells, we used anti-CD3-coated P815 target cells and co-cultured them with sorted human exT_reg cells at a 1:5 target-to-effector ratio. CD8⁺ cytotoxic T lymphocytes (CTLs) were used as positive control. T_reg cells were used as a negative control. exT_reg cells showed cytotoxicity in the same range as bona fide CD8⁺ CTLs (Fig. 5e). As expected, T_reg cells showed no cytotoxicity (Fig. 5e).

These data collectively show that CD16⁺CD56⁺ exT_reg cells do not exhibit T_reg cell-like suppressive properties and instead acquire cytotoxicity, although they retain some T_reg cell genes and markers. Unlike cytotoxic NK cells, exT_reg cells can be activated by TCR-related stimulation and express T cell markers upon activation.

Human exT_reg cells express cytotoxic and inflammatory markers

We recently reported that mouse exT_reg cells, unlike T_reg cells, express tumor necrosis factor (TNF) and IFN-γ⁷. Here we compared the expression of cytotoxic and inflammatory cytokines in human T_reg cells and exT_reg cells in an intracellular cytokine staining (ICS) assay (gating scheme in Extended Data Fig. 8a). About 60% of exT_reg cells coexpressed granzyme B and perforin (Fig. 6a). We confirmed that about 30–40% of exT_reg cells expressed intracellular TNF protein as measured by flow cytometry (Fig. 6a). Basal expression of IFN-γ was low in unstimulated exT_reg cells but strongly upregulated upon activation with PMA and ionomycin (Fig. 6b). Stimulation-induced IFN-γ production was significantly higher in exT_reg cells than in T_reg cells (Fig. 6b).

Fig. 6 |. Human exT_reg cells express cytotoxic proteins, inflammatory cytokines, chemokines and chemokine receptors.

a, Frequencies of GzmB⁺perforin⁺ and TNF⁺ cells in T_reg cells (blue circles) and exT_reg cells (red circles) were assessed in an ICS assay by FACS. n = 6. 50% male and 50% female donors; aged 24–39 years. b, Representative contour plots (left) showing intracellular expression of IFN-γ under unstimulated and PMA–ionomycin-stimulated conditions in exT_reg cells and T_reg cells. Right, frequency of PMA-induced IFN-γ in exT_reg cells (red circles) and T_reg cells (blue circles). n = 5. 40% male and 60% female donors; aged 23–39 years. c, hPBMCs were stained for intracellular expression of CCL3 (MIP-1α), CCL4 (MIP-1β) and CCL5 (RANTES). Top, representative contour FACS plots showing expression among all CD4⁺T cells and in exT_reg cells. Bottom, frequencies of CCL3⁺, CCL4⁺ and CCL5⁺ cells among the parent subset (all CD4⁺T cells (black circles), exT_reg cells (red circles)). n = 6. 50% male and 50% female donors, aged 26–33 years. d, Frequencies of CCR5 (n = 12), CXCR2 (n = 5), CXCR3 (n = 12), CXCR4 (n = 12) and CX3CR1 (n = 13) chemokine receptor-expressing exT_reg cells (red circles) and all CD4⁺T cells (black circles). 40% male and 60% female donors, aged 20–69 years. Results (a–d) were plotted as the mean ± s.e.m. Each dot represents a biological replicate from an independent donor. Statistical comparisons were performed using a two-tailed Mann–Whitney U test (a–d). **P = 0.0022 (a), **P = 0.0079 (b), **P = 0.0022 (c). In d, **P = 0.0068 (CCR5), **P = 0.0079 (CXCR2), ****P = 7.396 × 10⁻⁷ (CXCR3), **P = 0.0023 (CXCR4), ****P = 1.92 × 10⁻⁷ (CX3CR1).

In our gene expression analysis (Fig. 3b), the chemokines CCL3, CCL4 and CCL5 were significantly higher in human exT_reg cells than in T_reg cells. Using FACS, we confirmed that intracellular protein expression of all three chemokines was specifically enriched in exT_reg cells and not detected in bulk CD4⁺T cells (Fig. 6c). About 60% of exT_reg cells expressed CCL3 or CCL4 proteins (Fig. 6c). CCL5 expression was found in 70–80% of exT_reg cells (Fig. 6c). These chemokines attract monocytes²⁵ and promote inflammation.

Chemokine receptors play a major role in T cell trafficking and tissue infiltration. In the pathway analysis of human exT_reg genes, the CXCR4 signaling pathway was significantly enriched (Extended Data Fig. 3c). To confirm this finding, we used flow cytometry to monitor the expression of CXCR4 and other inflammation-related chemokine receptors such as CCR5, CXCR2, CXCR3 and CX3CR1. Expression of all these receptors was significantly higher on ex T_reg cells than on bulk CD4⁺T cells (representative data and FMO controls in Extended Data Fig. 8b, quantification in Fig. 6d). About 20% of exT_reg cells expressed CCR5 or CXCR2. CXCR3, CXCR4 and CX3CR1 could be detected in about 50–60% of exT_reg cells. In atherosclerotic mice, CCR5 expression on T cells is required for T cell homing to aortic lesions¹⁰. CXCR3 is a well-established inflammatory chemokine receptor expressed on T_H1 cells and on activated effector and memory T cells that migrate to inflammatory lesions²⁶. CXCR2 was highly expressed on CD4⁺T cells that infiltrated gray matter tissue in a mouse model of multiple sclerosis²⁷. CX3CR1 has been shown to be expressed on highly polarized cytotoxic dengue-virus specific CD4⁺T cells in humans²⁸. Its ligand CX3CL1 is highly expressed in atherosclerotic lesions^29,30. SDF-1, the ligand for CXCR4, has been detected at higher levels in human atherosclerotic plaques, and not in normal blood vessels³¹.

Inflammatory and cytotoxic genes in exT_reg cells from individuals with coronary artery disease

The CAVA cohort, from which the scRNA-seq data with CITE-seq came, encompasses participants with Gensini scores > 30 (CAD⁺) and < 6 (CAD⁻). First, we intersected the genes significantly upregulated in human exT_reg cells (up in exT_reg cells; Supplementary Table 6) with the genes tested in the human scRNA-seq dataset and found 108 such genes (Fig. 7a and Supplementary Table 11). Next, we tested which of these genes were expressed at significantly higher levels in exT_reg cells from CAD⁺ cases as compared to CAD⁻ controls. In non-diabetic CAD⁺ individuals, we found that CNOT2, CCL4, IL18RAP, KLRG1, KLRC1, LYN, SAMD3 and SYNE1 were significantly upregulated (Fig. 7b) compared to CAD⁻ controls (Fig. 7b). In exT_reg cells from diabetic individuals with CAD (Fig. 7b), CCL5, FGFBP2, ITGA4, ITGAM, KLRB1 and KLRC1 were significantly higher in CAD⁺ than in CAD⁻ individuals (Fig. 7b). KLRC1 was found to be significantly overexpressed in both non-diabetic and diabetic CAD⁺ individuals. Statistical analysis of the differential expression of these 13 genes between exT_reg cells and T_reg cells showed highly significant upregulation in human ex T_reg cells (Fig. 7c).

Fig. 7 |. Inflammatory and cytotoxic human exT_reg genes overexpressed in individuals with coronary artery disease.

a, Genes that were significantly upregulated in human exT_reg cells compared to T_reg cells (2,803 genes) in the human bulk RNA-seq data (Fig. 3) were intersected with the genes present in the published human scRNA-seq panel (Fig. 2). Donor details for human bulk RNA-seq are in Supplementary Table 13 and those for scRNA-seq are in ref. 21. b, Mean expression of genes in CD16⁺CD56⁺ exT_reg cells from the scRNA-seq dataset that were significantly increased in CAD⁺ non-diabetic (brown circles in top, n = 7) or diabetic (red squares in bottom, n = 11) individuals in comparison to control CAD⁻ non-diabetic (black circles, n = 12) individuals. Results are shown as the mean ± s.e.m. Each point represents data from exT_reg cells from an independent donor. Statistical comparisons were done using a two-tailed Mann–Whitney U test. Top, **P = 0.0064 (CNOT2), *P = 0.0414 (CCL4), **P = 0.009 (IL18RAP), *P = 0.0361 (KLRG1), *P = 0.0163 (KLRC1), *P = 0.0371 (LYN), **P = 0.0095 (SAMD3), *P = 0.0395 (SYNE1). Bottom, *P = 0.0126 (CCL5), *P = 0.0354 (FGFBP2), *P = 0.0373 (ITGA4), *P = 0.018 (ITGAM), *P = 0.031 (KLRB1), *P = 0.032 (KLRC1). c, log₂FC and adjusted P values of a DEG analysis (based on a two-tailed Wald test with Benjamini–Hochberg P-value adjustment) between exT_reg cells and T_reg cells for the 13 CAD-relevant exT_reg genes from b.

In summary, using gene signatures from mouse exT_reg cells, we identified human CD3⁺CD4⁺CD16⁺CD56⁺ T cells as human exT_reg cells and validated expression of exT_reg signature genes and surface markers. Bulk RNA-seq showed that exT_reg cells are very different from T_reg cells and NK cells, representing a distinct subset of CD4⁺T cells. Human exT_reg cells express inflammatory and cytotoxic genes and retain some T_reg cell markers. TCR-seq showed shared T_reg/exT_reg cell TCRB clones that were more expanded in exT_reg cells compared to T_reg cells. Such oligoclonal exT_reg cells likely originated from proliferating T_reg cells. Functionally, human exT_reg cells lack suppressive capacity and have acquired cytotoxic properties. They express cytotoxic effectors, inflammatory cytokines, chemokines and chemokine receptors. Expression of multiple inflammatory and cytotoxic genes in exT_reg cells is augmented in individuals with CAD compared to CAD⁻ controls.

Discussion

Our findings define human exT_reg cells as CD3⁺CD4⁺CD56⁺CD16⁺ T cells. These cells express the exT_reg signature genes CST7, NKG7, GZMA, PRF1, TBX21 and CCL4. At the protein level, human exT_reg cells express T cell markers like TCRαβ, CCR7, CD27 and CD127. Inflammatory chemokine receptors including CCR5, CXCR2, CXCR3, CXCR4 and CX3CR1 were enriched on exT_reg cells. Most human exT_reg cells also expressed granzyme B and the monocyte-attracting, pro-inflammatory chemokines CCL3, CCL4 and CCL5. They upregulated IFN-γ production upon stimulation. Unlike T_reg cells, exT_reg cells expressed little FOXP3 and completely failed to suppress T_eff cell proliferation. Instead, they exerted cytotoxic effects. Clonality analysis by TCRβ sequencing revealed that exT_reg cells contain expanded clones sharing TCR CDR3 sequences with T_reg cells.

It is estimated that 80% of systemic T_reg cells come from the thymus (tT_reg cells)^32,33. The remaining 20% are induced from conventional T cells in the periphery (iT_reg cells). Both nT_reg cells and iT_reg cells are mostly in LNs, but some are found in the blood^34,35. Fate-mapping experiments show that nT_reg cells are quite stable under steady-state conditions²⁰, whereas iT_reg cells are known to be unstable^14,15,19. Under conditions of chronic inflammation, such as those that exist in atherosclerosis, exT_reg cells, that is, cells that no longer express FoxP3, are found routinely^4,14,16,17. Zhou et al. showed the development and pathogenicity of exT_reg cells from T_reg cells using Foxp3^GFP × Rosa26^{-loxP-Stop-loxP-YFP} fate-mapping mice crossed with NOD mice¹⁴. We previously showed that exT_reg cells from mice, vaccinated with an apolipoprotein B peptide, produced IFN-γ⁷. Adoptive transfer of such exT_reg cells increased atherosclerotic lesion size⁷.

Because the transcriptome of T_reg cells shares similarity with that of T_H17 cells^36,37, exT_reg cells were suspected to become T_H17 cells. We have previously shown that in women with cardiovascular disease, many APOB-specific T cells express both FoxP3 and RORγt³. This phenotype is consistent with T_reg cell plasticity, which proposes that T_reg cells acquire the transcriptional program of the cells they regulate^38,39. Thus, T_H1-T_reg cells express T-bet, T_H17-T_reg cells express Stat3 and T_H2-T_reg cells express IRF4 (refs. 40–42). However, in atherosclerosis, T_reg cells lose FoxP3 expression. Based on flow cytometry, T_reg cells in atherosclerosis have been reported to become T_H17-like cells³, T_H1-like cells^10,11 or T_FH-like cells¹⁷. Our present lineage tracking, gene and protein expression and functional data show that, in both mice and humans, exT_reg cells are cytotoxic CD4⁺ T cells. The relationship between these different flavors of exT_reg cells remains to be explored.

Recent work has shown a loss of tolerance to self in mice⁴³ and humans⁴⁴ with atherosclerosis. Many T_reg cells express TCRs specific for self-epitopes⁴⁵. Thus, it is plausible that the conversion of T_reg cells to exT_reg cells through proliferation and clonal expansion may be part of the mechanism for this loss of tolerance to self. The function of exT_reg cells includes cytotoxicity. In fact, CTLs have been reported in plaques⁴⁶. exT_reg cells are also pro-inflammatory by attracting monocytes. Thus, exT_reg cells are expected to exacerbate atherosclerosis by attracting more monocytes, some of which can differentiate to macrophages⁹.

Transcriptomic analysis, ICS and functional assays all show that exT_reg cells have a highly activated cytotoxic phenotype. However, exT_reg cells retain expression of the T_reg cell markers PD-1, GITR and TIGIT, and expression of some T_reg cell genes. TCRβ sequencing shows that exT_reg cells have a limited repertoire diversity and show clear clonal expansion of TCR sequence patterns also found in T_reg cells. Experimentally, we show that T_reg cells proliferate in mice with atherosclerosis. Considering these findings together, we hypothesize that the development of exT_reg cells in chronic inflammation may be driven by repetitive TCR stimulation. Repetitive TCR stimulation is known to lead to a T_emra-like phenotype⁴⁷. Through this process, exT_reg cells become cytotoxic and pro-inflammatory T_eff cells.

It was surprising to find CD16 and CD56 as cognate surface markers for human exT_reg cells. CD56, also known as NCAM (neural cell adhesion molecule) and CD16, an Fc receptor (FcγRIIIa), are typically coexpressed on NK cells⁴⁸. To contrast human exT_reg cells and NK cells, we compared their transcriptomes and found genes that were significantly differentially expressed in both mice and humans. We used FACS analysis to confirm the expression of T cell markers on CD3⁺CD4⁺CD16⁺CD56⁺ human exT_reg cells, which were not detectable on CD3⁻CD4⁻CD16⁺CD56⁺ NK cells. Functionally, exT_reg cells, unlike NK cells, degranulated upon TCR engagement.

T_reg cells represent a unique, preexisting, non-naïve population of CD4⁺T cells that have already been exposed to antigen. Unlike naïve T cells, T_reg cells can quickly respond to antigen exposure. In cancer models, T_reg cells play a role in shielding the tumor from the immune system. However, inhibiting indoleamine 2,3-dioxygenase drives some of these T_reg cells to provide CD40L-dependent help to dendritic cells for cross-presentation, thus licensing CD8⁺T cells to become CTLs^49,50. The reprogrammed T_reg cells were induced by vaccinating with antigen and a TLR9 ligand⁵⁰. They arise from an Eos (Ikzf4)-labile population of T_reg cells⁴⁹. Unlike the cytotoxic exT_reg cells identified here, the T_reg cells reprogrammed by indoleamine 2,3-dioxygenase inhibition still express FoxP3 at levels similar to T_reg cells⁴⁹. Thus, these cells may represent early exT_reg cells, driven by adjuvant-induced inflammatory signals including IL-6 (ref. 50). In experiments using CFSE-labeled T_reg cells, reprogramming occurred at a time when proliferation was still minimal (that is, before proliferation) and required the IL-6 receptor on the T_reg cells⁴⁹. The relationship between these early exT_reg cells and the cytotoxic exT_reg cells described here remains to be determined.

A limitation of this study is that the mouse exT_reg cell signatures are from sorted spleen and LN cells, while the human exT_reg cell transcriptomic data are from peripheral blood. This discrepancy is due to practical constraints. The yield of PBMCs in mice is insufficient for sorting T_reg cells and exT_reg cells for preparation of high-quality libraries. Conversely, LN or spleen biopsy samples are not available from the CAVA cohort (from where the human scRNA-seq dataset came). Future, more detailed studies, beyond the scope of this work, may yield more transcriptomic data for phenotypic and functional analysis.

In conclusion, we discovered and defined, phenotypically and functionally, human exT_reg cells. Although human exT_reg cells retain some T_reg cell markers, their main gene signature is cytotoxic. Functionally, human exT_reg cells kill target cells as efficiently as CTLs. exT_reg cells are pro-inflammatory in that they express IFN-γ and monocyte-attracting chemokines. The conversion of T_reg cells to exT_reg cells may be part of the recently described breaking of tolerance to self in atherosclerosis.

Online content

Any methods, additional references, Nature Portfolio reporting summaries, source data, extended data, supplementary information, acknowledgements, peer review information; details of author contributions and competing interests; and statements of data and code availability are available at https://doi.org/10.1038/s41590-023-01589-9.

Methods

Human samples

Healthy volunteers were recruited by the clinical core at the La Jolla Institute for Immunology (LJI). All participants received financial compensation according to guidelines approved by LJI’s Institutional Review Board. Written informed consents were obtained from all enrolled participants. Donors self-reported ethnicity and race details, and tested negative for hepatitis B, hepatitis C and HIV. None of the donors had any ongoing infection. They had no known conditions of cancer, diabetes, heart or kidney or liver disease. Donors were neither pregnant nor nursing. De-identified blood or PBMC samples for the study were made available by LJI’s clinical core. Ethical approval for the study was provided by the Institutional Review Board of LJI (IRB protocol nos. IB-248–0821 and VD-057).

Animal experiments

All animal experiments in this study used fate-mapping lineage-tracker mice (FoxP3^{eGFP-Cre-ERT2}ROSA26^{CAG-fl-stop-fl-tdTomato} Apoe^−/− in a B6 background) from both sexes, as previously described⁷. All mouse experiments were approved by the LJI Animal Care and Use Committee (protocol no. AP00001019). The housing conditions for these mice were as follows: lights on at 6:00 and off at 18:00, ambient temperature between 20 °C and 24 °C and humidity between 30% and 70% (mean, 40%; outdoor environment can affect humidity inside).

Seven- to eight-week-old lineage-tracker mice were injected with tamoxifen for 5 consecutive days (intraperitoneally; 75 mg per kg body weight). For time-course experiments, spleens and LNs from four to seven mice (exact number indicated in figure legends) were harvested at 4, 8, 12 or 20 weeks after tamoxifen injection. Fixed samples of mouse carotid arteries from four 20-week-old tamoxifen-injected mice were imaged using a Leica SP8 confocal microscope with a HC Fluotar ×25/0.95 water objective. CD25 expression was quantified in single-cell suspensions from spleens (n = 4) and LNs (n = 6) of 16-week-old tamoxifen-injected mice. Bulk RNA-seq was performed on spleens (n = 5) and LNs (n = 4) from 20-week-old tamoxifen-injected mice. Animals were on a regular CD. For analysis of Foxp3/GFP and tdTomato/RFP within non-T cells, TCRβ⁺CD4⁻ cells and TCRβ⁺CD4⁺ cells, three 8-week-old female lineage-tracker mice were injected with tamoxifen twice for 5 d each, at week 1 and week 6. These mice were fed a western diet (WD; 42% kcal from fat, 0.2% cholesterol) for 12 weeks and then spleens (n = 3) and LNs (n = 3) were harvested for flow cytometry.

BrDU proliferation assay

Seven 8-week-old female Foxp3^{eGFP-Cre-ERT2}ROSA26^{fl-STOP-fl-tdTomato} Apoe^−/− (B6) mice were injected with tamoxifen twice for 5 d each, at week 1 and week 6, then fed a WD for 12 weeks. For six mice, BrDU (0.8 mg ml⁻¹) was incorporated in the drinking water for the last 9 d of WD feeding. Spleen cells were harvested and processed into single-cell suspensions. After staining for viability and surface markers for 30 min at 4 °C, cells were fixed and permeabilized for 20 min at 24 °C with the Foxp3-staining kit (eBioscience). Cells were then treated with DNase I solution for 1 h at 37 °C. Cells were washed and stained for intracellular markers for 30 min at 24 °C. Samples were analyzed on an LSR II (BD Biosciences).

Cell isolation

Mouse spleens and LNs were harvested, crushed and filtered through a sterile 100-μm filter in 1× PBS. Red blood cells and platelets from spleen cells were discarded using RBC lysis buffer (Invitrogen) for 5 min followed by centrifugation at 250g for 10 min.

hPBMCs were isolated by density gradient centrifugation. Briefly, undiluted blood was transferred onto Ficoll-PaquePLUS (Cytiva) in SepMate tubes (StemCell) and centrifuged at 1,200g for 10 min. PBMC ring was harvested, washed with 1× PBS (Gibco) and centrifuged at 800g for 10 min. Red blood cells and platelets were discarded using RBC lysis buffer (Invitrogen) for 5 min followed by centrifugation at 250g for 10 min.

Flow cytometry

Mouse splenocytes or lymphocytes and hPBMCs were incubated with fixable viability dye (Ghost Dye Violet 510; Tonbo Biosciences or zombie yellow; BioLegend) for 30 min at 4 °C, washed and antibodies against surface markers were added for 30 min at 4 °C, according to the analysis mentioned in the figure legends. ICS was performed using either the eBioscience IC Fixation Buffer (only cytokines or chemokine panels) or the eBioscience Foxp3/Transcription factor staining buffer set (when cytokines were co-stained with FOXP3) for 20 min at 24 °C. Briefly, cells were permeabilized, then incubated with antibodies against intracellular markers (as mentioned in the figure legends) for 45 min at 24 °C and analyzed on an LSR Fortessa or LSR II (BD Biosciences).

For suppressive assay readout, cells were incubated with anti-CD3 or anti-CD4 for 20 min at 4 °C. DAPI was added right before analyzing on a BD FACSCanto (BD Biosciences).

For ICS assay, hPBMCs were stimulated for 6 h with PMA and ionomycin (1×; Cell Stimulation Cocktail, eBioscience). For the last 4 h, protein transport inhibitors brefeldin and monensin (1×; Protein Transport Inhibitor Cocktail, eBioscience) were added. After staining for viability and surface markers, cells were fixed, permeabilized and stained for intracellular markers. Samples were analyzed on an LSR II (BD Biosciences). All flow cytometry data were analyzed using FlowJo version 10.8.1.

Antibodies used are listed in Supplementary Table 12.

Cell sorting

Mouse cells from spleens and LNs were first incubated with a fixable viability dye for 30 min at 4 °C, washed and antibodies against surface markers were added for 30 min at 4 °C. Cells were sorted with a FACSAria II (BD Biosciences). Mouse T_reg cells were defined as TCRβ⁺CD4⁺GFP⁺TdTomato⁺ and exT_reg cells as TCRβ⁺CD4⁺ GFP⁻TdTomato⁺. hPBMCs were first enriched for CD4⁺ T cells. Cells were incubated with purified anti-CD14, anti-CD19 and anti-CD8a for 20 min at 4 °C. Then, Dynabeads coated with goat anti-Mouse IgG (Invitrogen) were added to the cells for 10 min on a turning wheel at 4 °C. Using a magnet, the supernatant containing CD4-enriched T cells was collected. Enriched cells were incubated for 30 min at 4 °C with antibodies against surface markers. Dead cells were stained using DAPI right before sorting on a FACSAria II or FACSAria Fusion (BD Biosciences). After gating on morphology, singlets, live cells (DAPI⁻) and CD8a⁻CD14⁻CD19⁻ T_reg cells were defined as CD3⁺CD4⁺CD25^hiCD127^lo, exT_reg cells as CD3⁺CD4⁺CD56⁺CD16⁺ and naive T cells as CD3⁺CD4⁺ CD45RA⁺CCR7⁺.

Antibodies used are listed in Supplementary Table 12.

Quantitative real-time PCR

Human T_reg cells and exT_reg cells were directly sorted into 750 μl TRIzol LS (Thermo Fisher) and 20 μl of low-input lysis buffer (0.1% Triton X-100 (vol/vol), 1 U μl⁻¹ RNase inhibitor, 2.5 mM dNTP), respectively. For TRIzol samples, RNA was extracted with miRNAeasy micro kit (Qiagen). Total RNA was quantified using a nanodrop spectrophotometer. TRIzol and low-input samples were reverse transcribed into cDNA using SuperScript II reverse transcriptase (Thermo Fisher). Real-time PCR reactions were performed according to the RT2 SYBR green gene expression assay protocol (Qiagen). RT² SYBR Green qPCR master mix and premade RT² qPCR Primer Assays (Qiagen) for human CST7 (GeneGlobe ID PPH05560E-200), NKG7 (GeneGlobe ID PPH07745A-200), GZMA (GeneGlobe ID PPH00314F-200), PRF1 (GeneGlobe ID PPH07126A-200), TBX21 (GeneGlobe ID PPH00396A-200), CCL4 (GeneGlobe ID PPH00563B-200) and ACTB (GeneGlobe ID PPH00073G-200) were used. Details can be found at https://geneglobe.qiagen.com/us/product-groups/rt2-qpcr-primer-assays. Results were calculated by applying the 1/^ΔCT method using ACTB as housekeeping gene.

Bulk RNA-seq

Mouse T_reg cells and exT_reg cells were sorted into TRIzol LS (Thermo Fisher). Human T_reg cells (200,000–500,000) were sorted into 750 μl of TRIzol LS (Thermo Fisher), while human exT_reg cells (1,000 cells) were directly sorted in 20 μl of low-input lysis buffer as described previously⁵¹. For TRIzol samples, RNA was extracted using the miRNeasy micro kit (Qiagen). RNA quality was measured by using a 2100 TapeStation (Agilent Technologies). Samples with high-quality RNA (RNA integrity number > 8.0) were used for the next steps. For both TRIzol LS and low-input samples, double-stranded-cDNA was prepared using the SuperScript II reverse transcriptase (Thermo Fisher) according to the manufacturer’s instructions. cDNA was amplified using 15 cycles and eluted in 24 μl. Around 100 ng of resulting cDNA was processed using the Illumina DNA Prep Kit (20018704, Illumina) following the manufacturer’s instructions. Samples were normalized based on DNA concentration, pooled and loaded onto an Illumina NovaSeq 6000, and sequenced with 50 base-pair paired-end reads (PR50). Post-mapping quality-control checks were used to exclude poor-quality samples. Sequencing quality control was performed with FastQC v0.11.9 and MultiQC version v1.12. RNA-seq reads were trimmed using Ilumina’s DRAGEN FASTQ toolkit version 1.0.0. The STAR (v2.7.1 with default parameters)⁵² aligner was used to map the transcriptomes of the human and mouse bulk RNA-seq data to GENCODE GRCh38.p13 and GENCODE GRCm39, respectively. Raw read gene counts were obtained using STAR aligner with the ‘--quantMode GeneCounts’ option, which was used for the differential gene expression analysis. Quality of read counts was assessed before normalization, batch-effect correction and performing differential expression analysis using DESeq2 (v1.34)⁵³. To eliminate lowly expressed genes, a cutoff (≤100 raw counts across samples) was enforced to remove genes with low read counts. Normalized counts data from DESeq2 were used to make z-score heat maps and perform the differential gene expression analysis. Variance stabilizing transformed (VST) counts were used to make the PCA plot and the spearman correlation plot.

Donor information is listed in Supplementary Table 13. Mouse and human bulk RNA-seq data are available on the NCBI Gene Expression Omnibus (GEO) under accession no. GSE217010.

Support vector machine

We used the SVM classifier from Scikit-learn (v1.1)⁵⁴ machine learning library to classify data points. Packages such as Numpy (v1.23.2), Matplotlib (v3.5.2), Pandas (v 1.4.3) and Seaborn (v0.11.2) were used along with Scikit-learn to aid preprocessing. The mouse bulk RNA-seq data on sorted exT_reg cells and T_reg cells from spleen and LNs were used for SVM classification. For binary classification purposes, samples were labeled as either exT_reg cells or T_reg cells. Information on tissue type was disregarded. In total, 383 overlapping mouse–human orthologous genes were used in the single-cell datasets. The linear kernel was used to make predictions for which data were split into 90% train and 10% test. Fivefold cross-validation resulted in an accuracy of ~98%. The feature weights were extracted from the linear classifier to interpret the weights assigned to the genes. The weights defined the classifying power of each gene. The top 60 genes with positive weights (classifying exT_reg cells) and the top 60 genes with negative weights (classifying T_reg cells) were identified.

Single-cell RNA-seq analysis

A published²¹ single-cell dataset with CITE-seq on hPBMCs was used. This dataset is from 61 men and women (aged 40–80 years) undergoing cardiac catheterization at the University of Virginia Health System. All participants provided written informed consent before enrollment, and the study was approved by the Human Institutional Review Board (no. 15328) at the University of Virginia. Peripheral blood for sequencing experiments was obtained from these participants before catheterization. We performed differential gene expression analysis using Seurat v4.0.6 (FindAllMarkers and FindMarkers function), which uses the Wilcoxon rank-sum test. We compared these cells against all the CD4⁺ T cells and the T_reg cells (cluster 17 from ref. 21) to identify DEGs and surface markers. Human scRNA-seq data are available at the NCBI GEO (accession no. GSE190570).

Plots

GGplot2 v3.3.5 and ComplexHeatmaps v2.12.1 were used to make bar plots and heat maps. Feature plots and UMAP plots were generated using Seurat’s FeaturePlot and DimPlot functions. Dot plots were generated using the package Ggpubr v0.4.0.

Enrichment analysis

For pathways analysis, gene lists have been inputted to the Enrichr online tool (https://maayanlab.cloud/Enrichr/). GSEA⁵⁵ (v4.2.3) was used to identify cell clusters enriched for T cell-specific gene sets. We interrogated the pseudobulk of the single-cell transcriptome by performing GSEA against the exT_reg cell and T_reg cell gene signatures to identify clusters enriched for these cell types. We used the gene signatures of T_reg cells and exT_reg cells from mouse and human bulk RNA-seq data to validate and characterize cells expressing these genes in the single-cell dataset.

Bulk TCRβ sequencing

Human T_reg cells, exT_reg cells and naive T cells were sorted in HEPES buffer (1× PBS with 2% FBS and 0.025 M HEPES, pH7.3) and gDNA was extracted using QIAamp DNA Micro Kit (Qiagen). TCRβ data were sequenced with the immunoSEQ assay (Adaptive Biotech). Processing of raw Illumina sequence reads, filtering, demultiplexing, clustering and mapping of CDR3 sequences and annotation of VDJ genes using IMGT (ImMunoGeneTics) database sequences were performed by Adaptive Biotechnologies. Final sequence data were made available for download and analysis with their immunoSEQ Analyzer. Details of all productive TCR sequences are provided in Supplementary Table 14. TCR data were analyzed using GLIPH (v.2)^22,56. Further downstream analysis was done in R using packages dplyr v1.0.9 and stats v4.1.1. GLIPH clusters were filtered for Fisher’s score < 0.05. Clonally expanded and enriched motifs were identified by P-value statistics using the poisson. test function to compare summed contribution scores of samples from each cluster (summed template frequency of each cell type by cluster). The poisson. test function performs an exact test of a simple null hypothesis about the ratio between two rate parameters. The log₂ fold change values and P values of the enriched motifs were plotted as a volcano plot to identify expanded TCRs.

Donor information is listed in Supplementary Table 13.

Suppressive assay

Sorted CD16⁺CD56⁺ T cells and T_reg cells were labeled with CellTrace Violet (Thermo Fisher) and tested for suppressive activity at a ratio of T_reg cells/CD16⁺CD56⁺:Teff of 1:1 by a co-culture with CFSE labeled (Thermo Fisher) T_eff cells (12,500 cells per well) stimulated with the T_reg cell suppression inspector (Miltenyi) at a ratio of cells:beads of 1:2 in a 50 μl final volume of complete RPMI medium (RPMI 1640 medium + L-Glutamine, 1× penicillin–streptomycin, 10 mM HEPES buffer pH 7.3, 1 mM sodium pyruvate and 1× MEM non-essential amino acid) containing 5% AB Serum (GeminiBio). After 5 d of co-culture, proliferation was analyzed by dilution of CFSE in T_eff cells (L/D⁻CD3⁺CD4⁺).

Degranulation assay

hPBMCs were co-cultured with P815 target cells at a 10:1 (PBMC:target) ratio for 6 h at 37 °C, 5% CO₂. Target cells were either pre-loaded with 5 μg ml⁻¹ human anti-CD3 (OKT3 clone) or left uncoated. Fluorochrome-conjugated anti-CD107a was added to each well (1:100 dilution, vol/vol) at the start of the incubation period. After 2 h, protein transport inhibitor cocktail was added at a 1× concentration. After the incubation, cells were harvested, washed and stained with fluorescently labeled antibodies for surface markers. CD107a antibody was not added again at this stage. CD16⁺CD56⁺ exT_reg cells were gated from singlets, viable, CD14⁻, CD19⁻ CD3⁺CD4⁺T cells, while CD16⁺CD56⁺ mature NK cells were identified from CD3⁻CD4⁻ non-T cells.

Cytotoxic assay

P815 cells were incubated for 30 min with anti-CD3 (OKT3 clone; 5 μg ml⁻¹). After a washing step, cells were resuspended in complete RPMI medium containing 5% AB Serum and 50 μM β-mercaptoethanol. Sorted CD8⁺ CTLs, T_reg cells and CD16⁺CD56⁺ T cells were co-cultured with anti-CD3-loaded P815 at a ratio of 1:5 of P815:effector cells (7,500 P815 cells per well) in a final volume of 50 μl for 16 h at 37 °C and 5% CO₂. Cytotoxicity was measured in the supernatant using CyQUANT LDH Cytotoxicity Assay Kit following the provider’s instructions (Invitrogen).

Statistics

Data analysis and statistical comparisons were done using GraphPad Prism version 9.3.1 and R version 4.0.1. We used a two-tailed Wald test with Benjamini–Hochberg P-value adjustment for comparing differential expression of genes in the human and mouse bulk RNA-seq transcriptomes. For analysis of gene and surface marker enrichment within CD4⁺ T cell clusters in the human scRNA-seq data, we used a two-tailed Wilcoxon’s rank-sum test with Benjamini–Hochberg correction for multiple comparisons. We used a two-sided Poisson test for the identification of enriched CDR3 motifs in the human TCR-seq dataset. Two-sample comparisons were done with a two-tailed Mann–Whitney U test or with a two-tailed unpaired student’s t-test. We used one-way ANOVA with Dunnett’s multiple-comparison test or a Kruskal–Wallis test with Dunn’s multiple-comparison testing for analyses involving more than two samples. A two-way ANOVA with Dunnett’s multiple-comparison test was performed when two independent variables were involved. We used a two-tailed Fisher’s exact test with Benjamini–Hochberg adjusted P values for pathway enrichment analysis. All statistical tests, sample sizes and error bar descriptions of graphs are detailed in the legends of respective figures. Statistical tests for supplementary tables are provided in the column headers. Exact P values are reported in the figure legends and Supplementary Information.

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.

Extended Data

Extended Data Fig. 1 |. Experimental controls for lineage-tracker atherosclerotic mouse model and differentially expressed genes in mouse Treg cells vs exTregs.

(a,b) Eight week-old female Foxp3^{eGFP-Cre-ERT2} ROSA26^{fl-STOP-fl-tdTomato} Apoe^−/− mice were injected with Tamoxifen twice for 5 days each, at week 1 and 6, then fed Western diet (WD) for 12 weeks. a) gating strategy, b) representative plots and quantification of exTreg and Treg cells among CD4⁺T cells (black circles) in lymph nodes (LNs) and spleen, harvested after 12 weeks of WD from 3 independent mice. Non-T cells (orange open squares) and CD4⁻T cells (green open circles) are negative controls. Frequencies of exTregs and Treg cells among the parent subsets were plotted as mean ± SEM. Statistical comparisons were done using 2-way ANOVA with Dunnett’s multiple comparison test. **** p < 0.0001. c) Volcano plot representing significantly differentially expressed genes between mouse Treg cells and exTregs from 20-week old FoxP3^{eGFP-Cre-ERT2} ROSA26^{CAG-fl-stop-fl-tdTomato} Apoe^−/− mice (lymph nodes and spleen pooled). Left, up in Treg cells (blue). Right, up in exTregs (red). Horizontal dotted line is at -log₁₀ (p adjusted) = 1.3 (p_adj = 0.05). The top 60 exTreg and 60 Treg classifying genes from the SVM model are annotated. Canonical Treg genes Il2ra and Foxp3 are shown in black boxes. Statistical analyses of DE genes using two-tailed Wald test with Benjamini-Hochberg correction for multiple comparisons.

Extended Data Fig. 2 |. Expression of exTreg candidate genes in scRNAseq data and validation by qRT-PCR from sorted human cells.

(a) Feature maps showing the gene expression of the single gene markers CST7, NKG7, GZMA, PRF1, TBX21 and CCL4 in the human single-cell dataset for all CD4 T cell clusters. (b) Combinations 1–4 and 6 of exTreg candidate genes are highlighted in red on UMAP embeddings of CD4 T clusters from the scRNA-Seq. (c) UMAP embeddings of CD4 T clusters. Black outline marks cluster CD4T_7; cells that express either CD56 (left) or CD16 (right) are shown as red dots. (d) Gating strategy to identify exTreg and Treg cells in human PBMCs. Dump channel: CD14, CD19. (e) Gene expression analysis of CST7, NKG7, GZMA, PRF1, TBX21 and CCL4 in sorted human Treg cells (blue circles) and exTregs (red circles) by qRT-PCR. Gene-specific Ct values were normalized (ΔCt) based on actin (ACTB). Relative expression was calculated by the 1/ΔCt method. n = 7. 33.33% male, 66.67% female donors; age: 21–54 yrs. Data shown as mean ± SEM. Each dot represents a biological replicate from an independent donor. Statistical comparisons by two-tailed Mann Whitney U test. **p = 0.0012,***p = 0.0006.

Extended Data Fig. 3 |. Human bulk RNAseq.

(a) Gating strategy used to sort human exTregs and Treg cells to perform bulk RNA-seq. (b) gene set enrichment analysis (GSEA) of bulk RNA-seq transcriptomes of sorted human exTreg cells against CD4T_7 (left) and all other clusters (right). Normalized enrichment score (NES) and FDR q values are indicated. (c) Significantly (adjusted p < 0.05) enriched pathways in human exTreg cells, based on genes expressed at significantly higher levels in human exTreg than in Treg cells. Analysis by Bioplanet2019 from the EnrichR suite. Dotted line indicates adjusted p = 0.05 (-log₁₀ padj=1.3). Statistical comparisons with two-tailed Fisher’s exact test and Benjamini- Hochberg adjustment of p-values. (d) Gating strategy to identify exTregs and NK cells in human PBMCs. Dump channel: CD14, CD19.

Extended Data Fig. 4 |. Mouse bulk RNAseq.

(a) Comparative gene signature analysis between mouse exTregs and Treg cells. Genes were filtered for significant differential expression in mouse and human dataset. Gene expression shown here is from FoxP3^{eGFP-Cre-ERT2} ROSA26^{CAG-fl-stop-fl-tdTomato} Apoe^−/− mice. Low-expressed genes (<7 raw reads in all samples) in our dataset were filtered out. Technical replicates were averaged, biological replicates shown as columns. Analysis of differentially expressed (DE) genes was done using DESeq2. Curated list of significant DE (log₂FC ± 1, adjusted p < 0.05) genes are shown on normalized heatmaps, scaled by row (z scores). (b) Gene set enrichment analysis (GSEA) of mouse exTreg genes from bulk RNA-seq transcriptomes against human exTreg (left) and Treg cells (right) from the human bulk RNA-seq data set. Mouse orthologs of human genes, filtered for those present in the human scRNA-Seq targeted gene panel, were used to calculate enrichment for mouse bulk RNA-seq dataset. (c) Comparative gene signature analysis between mouse exTreg and NK cells. An external dataset was used for mouse NK cells (3 samples): GSE122597, GSE116177, and GSE52043. EdgeR was used to normalize the counts by applying the trimmed mean of M-values (TMM) method and counts per million (CPM) conversion. All other data processing and filtering steps were same as in a. Curated list of significant DE (log₂FC ± 1, adjusted p < 0.05) genes are shown on normalized heatmaps, scaled by row (z scores). Statistical analyses of DE genes (a,c) using two-tailed Wald test with Benjamini-Hochberg correction for p-value adjustment. All data from independent biological replicates.

Extended Data Fig. 5 |. Assessment of proliferation in mouse Treg cells and exTregs.

(a) Eight week-old female Foxp3^{eGFP-Cre-ERT2} ROSA26^{fl-STOP-fl-tdTomato} Apoe^−/− mice were injected with Tamoxifen twice for 5 days each, at week 1 and 6, then fed Western diet (WD) for 12 weeks. BrDU (0.8 mg/mL) was incorporated in the drinking water for the last 9 days of WD (n = 6). (b) Gating scheme for CD4⁺T cells. (c) Ki67 FMO control. (d) Representative plots and quantification of proliferating Treg cells (blue circles, %Ki67⁺BrDU⁺CD4⁺Foxp3⁺RFP⁺) and exTregs (red circles, %Ki67⁺BrDU⁺CD4⁺Foxp3⁻RFP⁺) in the spleen (n = 6), as identified by anti-BrDU and anti-Ki-67 Abs. Data shown as mean ± SEM. Each animal is an independent biological replicate. Gates were set by FMO for Ki67 and by no BrdU controls for BrdU. Background from “No BrDU” control was subtracted for normalization. The percentage of proliferating cells was compared by two-tailed Mann-Whitney U test, **p = 0.0087.

Extended Data Fig. 6 |. Treg marker expression on human exTregs.

Representative contour FACS plots showing the expressions of PD-1, GITR, LAG3 and TIGIT in exTreg cells (left). Corresponding FMO controls were used to set the gates. Right, contour plots showing the expression of these markers in all CD4⁺T cells and in Treg cells from the same donor.

Extended Data Fig. 7 |. Cytotoxic and T cell activation marker expression on stimulated human exTreg vs NK cells.

Contour plots show intracellular expression of CD40L (X-axis) and Perforin (Y-axis) in exTregs and NK cells from unstimulated and PMA+ionomycin stimulated PBMCs. Data from three independent donors. 33.33% male, 66.67% female donors, age: 25–43 yrs.

Extended Data Fig. 8 |. Gating strategy and representative plots.

(a) Gating strategy used to analyze granzyme B, perforin and TNF in Treg cells and exTregs. (b) Contour plots show surface expression of chemokine receptors CCR5, CXCR2, CXCR3, CXCR4 and CX3CR1 on exTreg cells (left) and their corresponding expression in all CD4⁺T cells (right). Individual FMO controls were used to set the gate for expression of each receptor.

Data availability

RNA-seq data have been uploaded to the NCBI GEO and are accessible under accession numbers GSE217010 (mouse and human bulk RNA-seq) and GSE190570 (human scRNA-seq data). Human TCR-seq data were generated and processed by Adaptive Biotechnologies. Details of productive TCR sequences, accessed through their immunoSEQ Analyzer portal, are provided in Supplementary Table 14. Source data are provided with this paper.