All subjects in the longitudinal healthy control cohort and the whole blood RNA-seq cohort gave written informed consent in accordance with the Declaration of Helsinki, the IRB-approved protocols at the Benaroya Research Institute at Virginia Mason (IRB07109), and the VA Puget Sound Health Care System (MIRB#00755). The clinical trial cohorts were approved by independent IRBs at each participating clinical site, as described in the original clinical trial reports (15–17). Participants in each of these trials also provided informed consent prior to participation.

The phenotype, frequency, function, and modulation of T_EX were assessed using complementary assays and cohorts ( Supplementary Figure 1 , Table 1 ). Cross-sectional samples were used from T1D, RA, and renal cancer carcinoma (RCC) patients with age- and sex-matched health controls (HC). Longitudinal samples were analyzed from HC subjects and published clinical trials (15, 17, 18). Whole blood transcriptional analyses were performed from tempus tube collections. For all cellular analyses, peripheral blood mononuclear cells (PBMCs) were isolated from whole blood and cryopreserved until used. Additional transcriptional analyses were performed on sorted populations from PBMCs. When assessing the influence of age on T_EX in HCs, male and female subjects were selected for even representation across all ages. All assays were run and analyzed in a blinded manner, and staining batches included an internal control.

Table 1 lists cohorts used in this study, including demographics, percentage of RA HLA risk carriers and percentage of cytomegalovirus (CMV)-seropositive subjects. The longitudinal T1D cohort in Figure 1A consisted of 66 subjects with recent onset T1D who were placebo arms of Immune Tolerance Network and TrialNet trials (19–22). The longitudinal HC cohort in Figure 1B consisted of 99 HC subjects with no personal or family history of autoimmune disease who were recruited through the Sound Life Project led by the Benaroya Research Institute (BRI) in partnership with the Allen Institute for Immunology. The cross-sectional HC, T1D, and RCC cohorts in Figure 2 were from the BRI Registry and Repository; the HC had no personal history or first-degree relatives with autoimmune disease. The cross-sectional HC cohort in Figure 3F consisted of 30 individuals who had no personal history or first-degree relatives with autoimmune disease who were recruited through the BRI Registry and Repository. The whole blood RNA-seq cohort in Figure 4A is a cross-sectional cohort consisting of 97 seropositive RA subjects and 114 HC subjects matched for age, sex, and race. The RA subjects carried a diagnosis of RA based on the 2010 American College of Rheumatology criteria, were positive for ACPA and were recruited from the Virginia Mason Medical Center and the VA Puget Sound Health Care System. HC subjects had no first-degree relatives with autoimmune disease and were recruited through the BRI Registry and Repository. The clinical trial cohort in Figure 5A was from the teplizumab (anti-CD3) trial in individuals at risk for T1D conducted by the Type 1 Diabetes TrialNet (15) and consisted of 32 subjects. The clinical trial cohort in Figures 5B, C was from the Early AMPLE trial (16) and consisted of 29 individuals with new onset RA.

PBMCs from subjects with T1D, RA, RCC and age/gender-matched HC were stimulated with antibodies against CD3 (1 µg/ml plate-bound, UCHT1) and CD28 (2 µg/ml plate-bound, CD28.2) for 16 hours with and sorted for memory (CD45RO⁺) CD8 T cells that either co-expressed KLRG1 and TIGIT or lacked both markers as a comparison population using the cell sorting panel ( Supplementary Table 1 ). Sytox Green (1:1000, Invitrogen) was added to samples prior to acquisition to differentiate dead cells. For comparisons across cells of differing antigen specificities, cells from HC were enriched for CD8 T cells following the 16-hour stimulation protocol, then stained with Class I pentamer to identify CMV, Epstein-Bar Virus (EBV), and Flu antigens ( Supplementary Table 1 ) as described below.

The indicated populations were sorted directly into SMARTer v3 or SMARTseq v4 lysis reagents (Clontech). Cells were lysed and cDNA was synthesized. After amplification, sequencing libraries were prepared using the Nextera XT DNA Library Preparation Kit (Illumina) according to C1 protocols (Fluidigm). Barcoded libraries were pooled and quantified using a Qubit^® Fluorometer (Life Technologies). Single-read sequencing of the pooled libraries was carried out on a HiSeq2500 sequencer (Illumina) for 74 cycles, using TruSeq v3 or v4, and SBS kits (Illumina). Target read depths were ~5-10 million raw reads per sample.

Supplementary Table 1 lists antibodies used for each flow cytometry panel, including target, fluorophore, clone, and manufacturer. For assessment of cytokines, cells were treated with anti-CD3 (1 µg/ml plate-bound, OKT3) and anti-CD28 (2 µg/ml plate-bound, CD28.2) for 24 hours or Phorbol-Myristate-Acetate (PMA, Sigma) and Ionomycin (I, Sigma) for 6 hours and Brefeldin A (BioLegend) and Monensin (BioLegend) were each added at 1X for the last 4 hours. Dead cells were detected (Zombie NIR Kit, BioLegend), surface markers were added in brilliant stain buffer (BD Biosciences) for 20 minutes at RT and intracellular markers were detected (30 minutes RT) following permeabilization (FoxP3/Transcription factor staining buffer set, eBioscience, 30 minutes at 4°C). For assessment of antigen-specific phenotype, cells were enriched for CD8 T cells using negative selection (CD8⁺ T cell isolation Kit, Miltenyi) and incubated with dasatinib (50 nM, 250 µl/2 million cells, LC Laboratories) for 8-10 minutes at 37°C prior to staining with 25 µL solution containing 1.5 µL of commercially obtained Class I pentamer (ProImmune) for 15 minutes at 37°C, followed by surface marker detection in the T1D antigen-specific panel ( Supplementary Figure 2 ).

PBMCs were labeled (Cell Trace Violet, Invitrogen), stained with surface markers of the cell sorting panel ( Supplementary Table 1 , Supplementary Figure 3 ) and Sytox Green to differentiate dead cells (1:1000, Invitrogen). PBMCs were sorted using a BD Aria II until 3-5 × 10³ CD8 T cells were obtained per condition which co-expressed KLRG1 and TIGIT or lacked both markers. Sorted cells were mixed back into whole PBMCs from the same subject and stimulated with anti-CD3 (1 µg/ml plate-bound, UCHT1). Percent divided were assessed in labeled cells using FlowJo proliferation modeling. Labeled cells were also assessed for changes in KLRG1 and TIGIT expression following stimulation (% Stable = purity of labeled population following stimulation/purity of labeled population at baseline × 100).

RNA isolation, RNA-seq, and pipeline analyses including differential expression (Limma-Voom) and protein-protein networks were performed as described previously (23).

Whole blood libraries from the 211 HC and RA subjects in Figure 4A were TMM normalized and batch corrected for age, percent lymphocytes and percent duplication, a quality metric associated with PC1. SNPs were generated with an Affymetrix Axiom PMRA chip and single nucleotide polymorphisms (SNPs) on chromosome six (containing the HLA region) were selected for study. SNPs exhibiting little variance or frequent missing genotypes were removed from the analysis. The most significant DRB1*04 associated SNP (rs72492350) and an unassociated SNP (Chr6:32183175) were used as phenotypes in separate GSEA analyses (24) with 100-gene modules (25).

A linear mixed-effects model with a random effect for subject was used to calculate intraclass correlation coefficient (ICC), which quantifies the proportion of biomarker variation that is within and between subjects. Age and CMV seropositivity were added as covariates to investigate their association with biomarker frequency. Summary statistics include mean, range, median, and inter-quartile range; 95% confidence intervals are reported where appropriate. Spearman’s correlation coefficients were used for associations, Kolmogorov-Smirnov tests were used for cumulative distribution comparisons, Wilcoxon matched-pairs signed-rank tests were used for paired comparisons, and Mann-Whitney test was used for unpaired comparisons while a Kruskal-Wallis test with Dunn’s correction for multiple tests was used for multiple unpaired comparisons. All P values < 0.05 were considered significant.

We previously reported that co-expression of TIGIT and KLRG1 marked CD8⁺ T cells that had phenotypic and functional features of exhaustion, including an EOMES signature, and that these cells expanded following teplizumab (anti-CD3) therapy in individuals with T1D (12, 13). To address variation of TIGIT⁺KLRG1⁺ CD8 T cells in the absence of therapy, we first investigated the stability of these cells in vivo in a T1D cohort ( Table 1 , Supplementary Figure 1 ) measuring the proportion of TIGIT⁺KLRG1⁺ CD8⁺ T cells at multiple time points over two years ( Figure 1A ). We found that the frequency of TIGIT⁺KLRG1⁺ CD8 T cells varied little within T1D subjects over two years (mean within-subject range 8.2% [95% CI: 6.9-9.5]) but varied greatly between T1D subjects with a mean frequency range of 2.9% to 50.6%. To confirm that this stability is not unique to T1D, we measured the proportion of TIGIT⁺KLRG1⁺ CD8⁺ T cells at multiple time points over two years in a HC cohort ( Table 1 , Figure 1B ). We found that the frequency of TIGIT⁺KLRG1⁺ CD8 T cells also varied little within HC over time (mean within-subject range 6.5% [95% CI: 5.6-7.4]) while the mean frequency ranged from 4.2 to 59.8%. Lack of variation within subjects is supported by high intraclass correlation coefficient (ICC) values (83.7% and 92.0%, respectively), a measure comparing variability within versus across subjects.

A known contributor to increased T cell exhaustion in an individual is age and chronic viral infection (1). In both the HC and T1D cohorts ( Figure 1 ), increasing years of age (effect of 0.32 [95% CI: 0.20, 0.45], P = <0.0001) and CMV seropositivity (effect of 3.67 [95% CI: 2.06, 5.28], P = <0.0001) were significantly associated with TIGIT⁺KLRG1⁺ CD8 T cell frequency in a linear mixed-effects model. However, disease status did not have a significant effect (P = 0.65, fixed effect test) and variance contributed by age and CMV status were significant but not robust (age effect, 0.32, CMV effect, 3.67) suggesting that other factors also contribute to TIIGT⁺KLRG1⁺ variation across subjects.

We also investigated the stability of TIGIT⁺KLRG1⁺ T_EX in vitro using an in vitro assay system designed to track the frequency of TIGIT⁺KLRG1⁺ T_EX cells that maintain co-expression of TIGIT and KLRG1 upon activation. Specifically, memory TIGIT⁺KLRG1⁺ cells were sorted and labelled with a cell trace dye to identify them as TIGIT+KLRG1+ prior to activation. Sorted cells were then mixed with autologous PBMCs before activation with anti-CD3/anti-CD28 antibodies ( Supplementary Figure 3A ). Labelled cells were monitored over time for maintenance of TIGIT and KLRG1 expression. Measuring maintenance of this phenotype, we found that TIGIT⁺KLRG1⁺ CD8 T cells were stable for 8 days following anti-CD3/CD28 activation ( Supplementary Figure 3B ).

We previously reported that co-expression of TIGIT and KLRG1 marked CD8⁺ T cells that had phenotypic and functional features of exhaustion, including an EOMES signature, and that these cells expanded following teplizumab (anti-CD3) therapy in individuals with T1D (12–14). To expand the functional characterization of TIGIT⁺KLRG1⁺ CD8 T cells in the absence of therapy, we compared the transcriptome of TIGIT⁺KLRG1⁺ memory CD8⁺ T cells and TIGIT^-KLRG1^- memory CD8⁺ T cells from HC. Similar to our previous finding in the setting of T1D and immunotherapy (12), TIGIT⁺KLRG1⁺ memory CD8⁺ T cells had increased expression of T cell exhaustion markers including the transcription factor TOX and inhibitory receptors (i.e., LAG-3, CD160, CD244), and reduced expression of cell cycle genes ( Figure 2A , Supplementary Table 2 ). To determine whether TIGIT⁺KLRG1⁺ memory CD8⁺ T cells are similar across disease settings, we compared EOMES module expression in sorted TIGIT⁺KLRG1⁺ memory CD8⁺ T cells and TIGIT^-KLRG1^- memory CD8⁺ T cells from HC, individuals with T1D, and individuals with RCC; RCC was included because cancer is a setting where exhaustion is expected (1). We also included TIGIT⁺KLRG1⁺ CD8⁺ T memory cells sorted from both acute (influenza (FLU) and chronic (CMV) viral-specific T cells identified using pentamer staining ( Supplementary Figure 2 ). Across all disease settings tested, TIGIT⁺KLRG1⁺ memory CD8⁺ T cells differed from memory CD8⁺ T cells lacking TIGIT and KLRG1 expression (K-S test, P = 9.8e-10) ( Figure B ).

We assessed similarity of the TIGIT⁺KLRG1⁺ EOMES signature with other published signatures of T_EX identified across disease settings using cumulative distribution function (CDF) curves, asking whether other published signatures can discriminate TIGIT⁺KLRG1⁺ cells from TIGIT^-KLRG1^- cells. Published T_EX signatures included four murine T_EX subsets (3), common human cancer T_EX signatures (26) and the exhaustion-associated EOMES module that we previously identified in T1D subjects treated with teplizumab (anti-CD3) (12). Given that TIGIT and KLRG1 co-expression were identified in peripheral blood of T1D subjects, the T1D EOMES signature (12) best discriminated transcriptional profiles of TIGIT⁺KLRG1⁺ and TIGIT^-KLRG1^- populations (K-S test, P = 9.8e-10). Terminal T_EX signatures from the mouse and cancer data sets were also more similar to TIGIT⁺KLRG1⁺ cells (K-S test, P = 4.3e-03 and P = 9e-06, respectively). Moreover, we confirmed that EOMES protein expression correlates with co-expression of TIGIT and KLRG1 on memory CD8 T cells from HC using flow cytometry (Spearman test: r = 0.7015) ( Figure 2C ). Together, these data suggest that the TIGIT⁺KLRG1⁺ CD8⁺ T cell population is primarily composed of T_EX and is present in the peripheral blood in healthy individuals, individuals with autoimmune disease, and cancer.

To demonstrate that TIGIT⁺KLRG1⁺ memory CD8⁺ T cells are functionally exhausted and display reduced proliferation and cytokine production, we compared HC TIGIT⁺KLRG1⁺ memory CD8⁺ T cells to the total memory CD8⁺ T cell population which includes all memory CD8⁺ T cell subsets ( Supplementary Figure 4 ). Compared to total memory CD8, TIGIT⁺KLRG1⁺ memory CD8⁺ T cells divided fewer times ( Figure 3A ) and produced lower levels of TNF-α and IFN-γ upon T cell receptor stimulation ( Figure 3B ). This reduced effector function corresponded with phenotypic features of exhausted cells. Markers of effector function (CD127, CD226) were reduced, while inhibitory markers (PD-1, CD160, EOMES) were increased ( Figures 3C, D ). Thus, co-expression of TIGIT and KLRG1 on memory CD8⁺ T cells marks phenotypically and functionally exhausted TIGIT⁺KLRG1⁺ CD8⁺ T cells with reduced effector functions. For simplicity, henceforth, we refer to this population as TIGIT⁺KLRG1⁺ T_EX.

To confirm that TIGIT⁺KLRG1⁺ T_EX are increased in settings previously reported to display increased CD8 T cell exhaustion, we parsed TIGIT⁺KLRG1⁺ T_EX by progressive differentiation states and acute or chronic viral specificity. TIGIT⁺KLRG1⁺ cells were present in all subsets of CD8⁺ T cells with the majority being memory cells ( Figure 3E , Supplementary Figure 2 ). Within TIGIT⁺KLRG1⁺ CD8 T cells, effector memory were the most abundant (61%) with central memory (13%) and CD45RA⁺ effector memory (18%) being next abundant in the same dataset analyzed in Figure 3E . Consistent with an increase of T cell exhaustion in chronic as compared to acute viral infections (1), we found increased frequencies of TIGIT⁺KLRG1⁺ T_EX in CMV- and EBV-specific T cells identified by pentamer reagents as compared to influenza-specific T cells ( Figure 3F ). Thus, TIGIT⁺KLRG1⁺ T_EX can be identified across lineages, but are found primarily in effector cells and settings previously associated with increased T_EX (1).

Due to the TIGIT⁺KLRG1⁺ T_EX stability within and high variation across subjects, we were able to leverage cross-sectional datasets to explore autoimmune-related factors that influence the frequency of TIGIT⁺KLRG1⁺ T_EX. We examined whole blood RNA sequencing (RNA-seq) data from a large cohort of age- and sex-matched HC and RA subjects ( Table 1 ). While we found transcriptional differences between HC and RA, we also observed enrichment in the expression of genes that comprise the EOMES signature previously associated with CD8 T cell exhaustion (12) ( Figure 2 ) when stratifying the combined cohorts by RA HLA autoimmune risk alleles ( Figure 4A ). Specifically, we focused on the HLA DRB1*04 alleles (*0401, 0404, 0405 and 0408) and the closely related DRB1*1001 genes most strongly associated with RA (odds ratios > 4.2) (27), and refer to carriers of these alleles as risk RA HLA and non-carriers as non-risk RA HLA. The HLA distribution for the risk RA HLA subjects is shown in Supplementary Table 3 . Further investigation showed that the enrichment of EOMES modules in the non-risk RA HLA cohort was not due to CMV positivity since CMV-positive subjects were actually underrepresented (46%) in the non-risk RA HLA cohort as compared to the risk RA HLA cohort (52%). Likewise, the EOMES signature does not appear to be secondary to disease as similar enrichment in the non-risk RA HLA cohort was observed when HC were analyzed separately ( Supplementary Table 4 ). Last, complementary SNP association analyses within the HLA-DRB1 locus confirmed decreased RNA-seq EOMES module association with risk RA HLA alleles ( Supplementary Figure 5 ). Together, these findings support the association of an EOMES signature with the lack of RA HLA risk.

To determine whether the composition of the EOMES signatures in carriers of risk and non-risk RA HLA differ, we compared EOMES module expression in TIGIT⁺KLRG1⁺ memory CD8⁺ T cells isolated from HC carriers of risk and non-risk RA HLA. As in Figure 2B , we found the TIGIT⁺KLRG1⁺ cells isolated from both risk and non-risk RA HLA subjects were more similar to each other than their TIGIT^-KLRG1^- counterparts ( Supplementary Figure 6A ). Risk and non-risk RA HLA TIGIT⁺KLRG1⁺ memory CD8 T cells also shared interconnected genes common with genes identified in TIGIT⁺KLRG1⁺ T_EX as visualized in a protein-protein interaction network and were functionally similar ( Supplementary Figure 6B ).

Given the consistency of increased EOMES signature across disease settings and the correlation with TIGIT⁺KLRG1⁺ protein expression ( Figure 2 ), we predicted that the increased EOMES signature in non-risk RA HLA subjects would also be reflected at the protein level. For this experiment, we measured the frequency of EOMES-associated TIGIT⁺KLRG1⁺ T_EX in CMV-negative age- and sex-matched HC and RA subjects selected for high versus low EOMES signature, defined by upper and lower terciles ( Table 1 ). We did not observe differences in the frequency of EOMES-associated TIGIT⁺KLRG1⁺ T_EX between HC and RA subjects; nor were TIGIT⁺KLRG1⁺ T_EX functionally different ( Supplementary Figure 7 ) as assessed by similarly low IFNγ production. However, there was a significant increase in TIGIT⁺KLRG1⁺ T_EX abundance in the non-risk RA HLA subjects as compared with risk RA HLA subjects ( Figure B ). Thus, these data suggest the autoimmune-associated RA HLA genotype or linked genes contributes to variation in the frequency of TIGIT⁺KLRG1⁺ T_EX in a cohort of HC and RA subjects.

DR4 is a common risk allele between RA and T1D (28) and is associated with better outcome in a clinical trial of teplizumab (anti-CD3) therapy in individuals at risk for T1D (15). Given the association of T_EX with better response to therapy in autoimmune disease (12, 13, 15), we explored the relationship between TIGIT⁺KLRG1⁺ T_EX frequency and HLA risk alleles in the setting of immune interventions leveraging recent clinical trials. We first asked whether TIGIT⁺KLRG1⁺ T_EX are selectively modulated in DR4 T1D subjects, examining the teplizumab (anti-CD3) trial in individuals at risk for T1D since DR4 was previously identified as a weak correlate of response (15). We found a significant increase in TIGIT⁺KLRG1⁺ T_EX among DR4 risk subjects (P = 0.0033), but not DR4 non-risk subjects (P = 0.2650) ( Figure 5A ). Note, CMV seropositivity and mean age did not differ between DR4 risk and non-risk subjects. Thus, we link the previous DR4 association with response to a selective increase in TIGIT⁺KLRG1⁺ T_EX in DR4 subjects.

We analyzed CyTOF data from the Early AMPLE trial (ClinicalTrials.gov: NCT02557100), a randomized, head-to-head, single-blind study comparing abatacept (CTLA4Ig) and adalimumab (anti-TNF) in new-onset RA (16). The results from this trial in biologic naïve patients demonstrated a superior response in the abatacept arm that was more pronounced in subjects who carried the shared epitope alleles (HLA DR1, DR4, DR10) (16). Here, we examined TIGIT⁺KLRG1⁺ T_EX in the abatacept-treated group based on risk and non-risk RA HLA as defined in Figure 4 . We did not find an increase in TIGIT⁺KLRG1⁺ T_EX with treatment across all subjects but there was notable heterogeneity. When stratifying by RA HLA risk, we observed a significant increase in the frequency of TIGIT⁺KLRG1⁺ T_EX in risk RA HLA subjects (P = 0.0043), but not non-risk RA HLA subjects (P = 0.1250) following treatment with abatacept ( Figure 5B ). In contrast, there was no change in the frequency of TIGIT⁺KLRG1⁺ T_EX in RA HLA risk subjects after adalimumab treatment in either risk or non-risk RA HLA subjects ( Figure 5C ). Mean age of RA HLA risk groups did not differ in either study. Collectively, these findings suggest that TIGIT⁺KLRG1⁺ T_EX frequency depends, in part, on HLA risk alleles and may be modulated by some immunotherapies.