Introduction

The ability to induce immune memory is the hallmark of the adaptive immune system and the basis for one of the most successful medical interventions, the use of vaccines to induce protective immunity. A critical dimension of immune memory is the expansion and differentiation of antigen-specific memory T cells. T cell memory is best functionally defined as the ability to generate an enhanced response relative to the first encounter upon re-encountering an antigen. The key desirable property of immune memory is functional durability, frequently implying a compromise between the expansion of antigen-specific memory T cells while allowing for space for T cell responses against new antigens.

One major confounding factor in memory T cell generation and homeostasis is age1,2. However, the immense heterogeneity of memory cells represents a challenge to define age-related mechanisms that influence the durability of functional memory T cells3. Cells that have encountered antigens differentiate into multiple memory subsets, mainly defined by their homing pattern, such as tissue-resident or re-circulating central and effector memory T cells. Single cell sequencing analyses have shown that phenotypic subsetting cannot capture the full degree of transcriptional and epigenetic heterogeneity4,5,6.

A second level of complexity that complicates assessing the impact of age on memory T cell longevity comes from re-encountering the antigen. Each phenotypically defined T cell subset includes a composition of T cells with a range of antigen experiences, ranging from infrequent iterative boosting by re-infection or reactivation of latent viruses to continuous exposure to antigens in chronically active viral infections. Studies on memory T cell aging, even if done at the single cell level, therefore only provide estimates of averages and not age-associated trajectories, such as loss of an antigen-specific memory T cell or a change in functional states, including cellular senescence or exhaustion.

Murine studies suggested that the conditions of antigen re-encounter, particularly the timing, determine the fate of peripheral antigen-specific memory T cells7. When antigen re-stimulation events are spaced at least 4 weeks apart, memory T cells are extraordinarily resilient to aging, persist and maintain their expansion and cytokine production potential. However, when re-stimulation intervals are short, memory T cells rapidly become dysfunctional and lose their ability to expand effectively.

To examine the influence of age in humans on the repertoire and functional state of memory T cells in a more controlled approach, we focused on antigen-specific memory T cells against a given pathogen. We investigated the heterogeneity and aging traits of human peripheral, re-circulating memory CD8+ T cells specific for multiple Epstein-Barr virus (EBV) antigens. This approach allowed us to examine aging trajectories of memory T cells with differing antigen specificities that were primed simultaneously early in life.

The T cell memory response against EBV infection is a suitable model system as it leverages important characteristics of T cell heterogeneity and aging: (i) EBV infects > 90% of people globally with primary EBV infections occurring predominantly before 25 years-of-age8. After resolution of the primary EBV infection, EBV viral DNA is retained as a latent, chronic infection in memory B cells over lifetime9. (ii) Lymphocytes, especially CD8+ T cells, are vital to virus control and resolution of acute EBV infections. A subset of EBV-reactive T cells is retained as memory T cells over lifetime10,11. (iii) Although primed in parallel, CD8+ T cells specific for antigens expressed during the lytic and latent stage exhibit different trajectories in the first years after infection, both in terms of frequencies and expression of CD45 isoforms12. (iv) Similar to other herpes viruses, such as varicella zoster virus (VZV)13, EBV reactivates under immunocompromised conditions14 including those associated with normative aging15. Molecularly, EBV viral DNA can be detected in the peripheral blood of adults after the age of 50 years16,17. While age-related reactivation remains subclinical in most older adults, EBV-related malignancies (including lymphoproliferative diseases such as diffuse large B cell lymphoma, Hodgkin lymphoma, Burkitt lymphoma and nasopharyngeal cancer) are more frequently diagnosed with advancing age18. (v) EBV-memory T cells are numerically largely stable into older age. Data on whether their functional traits and phenotype are sensitive to age-related dysfunction are conflicting11,19,20. (vi) Aging of the humoral response against EBV is less important as EBV serostatus does not change with age8.

To determine how antigen-specific CD8+ T cells age, we employ barcoded-enabled antigen tracing of memory CD8+ T cells specific for different EBV proteins. We find that depending on the antigen recognized, cells have preferred and different differentiation states in young adults and undergo divergent fate trajectories with advancing age. While some aging signatures are shared across antigen specificities, many of the aging changes are non-uniform or differ in their extent. Specifically, larger age-related changes are seen for some clones recognizing latent EBV proteins suggesting that these specificities account for age-related T cell memory defects.

Results

Differentiation states of EBV-specific CD8+ T cells in young adults

Hislop et al.12 examined the evolution of a memory response after EBV infection and concluded that the antigen-specific CD8+ T cell populations seen during primary infection and in the first one to ten years of asymptomatic viral persistence are different12. The primary response was dominated by specificities to lytic epitopes which contracted more upon resolution of the acute infection than the initially less abundant latent epitope specificities. Importantly, the nature of the antigen also determined the evolution of phenotypic changes in these populations with a greater shift to a CD45RA-positive phenotype in T cells specific to lytic antigens. To identify aging trajectories of CD8+ memory T cells, we decided to control for the type of recognized antigen and first perform a cross-sectional, broad phenotypic characterization of CD8+ T cells specific to different EBV epitopes loaded onto tetramers in a cohort of young adults (< 40 years, Fig. 1A). We included 8 HLA-A*02:01+ individuals who had detectable EBV-specific T cells, presumably due to a clinically silent infection as they frequently occur in childhood. We performed multiplex spectral flow cytometry with 4 EBV peptide-loaded tetramers against antigens expressed in the EBV lytic (BMLF1, BRLF1) or latent life cycle (LMP2, EBNA3C), as well as 20 cell surface markers to identify classical T cell subsets, phenotypes, and cell states. We first performed a high dimensional cluster analysis of tetramer-negative, bulk CD8+ T cells to generate an unbiased reference map recapitulating CD8+ T cell subsets such as multiple clusters of naïve and stem-like memory cells (SCM), central memory (CM), effector memory (EM) and terminal effector memory cell with CD45RA expression (TEMRA) (Fig. 1B). To assess the phenotype and heterogeneity of EBV-specific cells (Fig. 1C), we next mapped EBV-tetramer-positive cells onto the bulk reference map and found marked differences in phenotypic states across antigen-specific T cells but also within each antigen specificity (Fig. 1D). Consistent with the observation by Hislop et al.12, T cells specific for the lytic antigens BMLF1 and BRLF1 on average were more differentiated and included end-differentiated TEMRA cells, while those against the latent antigens, LMP2 and EBNA3C, largely lacked TEMRA cells and contained more EM cells (Fig. 1E). However, there was considerable inter-individual variation (Supplementary Fig. 1A). Similarly, flow cytometry gating for classical surface markers signifying advanced differentiation (KLRG1+ CD28 and CX3CR1+ TIGIT+) yielded a higher proportion of end-differentiated cells in EBV lytic-specific T cells than in EBV latent-specific T cells (Supplementary Fig. 1B–D). To statistically compare the distribution of antigen-specific cells in these high-dimensional clustering data, we performed linear regression on probability vectors summarizing the cluster-specific cell frequencies. We found that T cells against lytic antigens were more similar to each other, while those against latent antigens, especially LMP2, differed significantly (Fig. 1F). This phenotypic divergence between EBV-specific T cells was also underscored by the principal component analysis (PCA) based on T cell cluster distribution data (Fig. 1G). Since the T cell differentiation stage is tightly associated with the expression of differentiation-associated transcription factors (TFs), we perform flow cytometry assessments with EBV tetramers and intracellular staining. Indeed, we found that EBV lytic antigen-specific T cells expressed higher T-BET and RUNX3 levels and contained more cytotoxic effector molecules, perforin and granzyme B (Fig. 1H and Supplementary Fig. 1E). However, other TFs that correlated with T cell differentiation, such as TOX/TOX2, did not show clear differences between lytic and latent-antigen specific T cells, suggesting a more complex picture beyond the divergence of EBV lytic and latent antigen-specific T cells.

Fig. 1: EBV-specific CD8+ T cells differ in their preferred, phenotypically defined differentiation states dependent on the recognized antigen.
figure 1

A Schematic of experimental design: CD8+ T cells from peripheral blood of young adults (< 40 years) were stained with the panel of antibodies shown in Supplementary Data 3 and HLA-A*02:01 tetramers loaded with indicated EBV peptides and then subjected to spectral flow cytometry. B UMAP reference plots and feature plots of indicated cell surface marker showing bulk CD8+ T cell subsets. Functional subset assignment was based on classical marker profiles including CD45RA/RO, CCR7, and CD28. C Representative flow cytometry scatter plots of tetramer staining and box plot of tetramer-positive cell frequencies (n = 7). Fluorophores for tetramers were PE-Cy5 (BMLF1), BV480 (EBNA3C), APC (BRLF1), and PE (LMP2). D Reference mapping of tetramer-positive cells projected on UMAP plots from bulk CD8+ T cells. E Bubble chart of tetramer-positive cell distributions among bulk CD8+ T cell-defined clusters defined in (B). The bubble size and color reflect the median proportion of antigen-specific T cells in a given cluster. F Comparison of probability vectors reflecting the subset distributions of tetramer-positive cells. Significance levels between the indicated pairwise comparisons are shown as bubble size and color. G Principal component analysis (PCA) of cell frequency distributions. H Protein levels of indicated transcription factors or cytotoxic molecules in tetramer-positive cells as assessed by spectral flow cytometry (n = 8). Data show the median with 95% confidence intervals (C), median (E), or mean ± SEM in (H). All datapoints represent distinct biological replicates. Data were compared by two-sided permutation tests (F), or two-tailed, unpaired t-tests comparing all lytic (BMLF1, BRLF1) versus all latent (LMP2, EBNA3C) antigen-specific T cell populations (H). Source data are provided as a Source Data file.

To support our targeted flow cytometry assessments, we leveraged a published single-cell sequencing dataset containing antigen-specific CD8+ T cells of predominantly young adults21. We extracted data for EBV-specific T cells against peptides from 8 different EBV epitopes (lytic antigens BMLF1, BRLF1, BZLF1, and latent antigens LMP1, LMP2/2 A, EBNA3A, Supplementary Fig. 2A). Cluster analysis based on targeted transcriptome and surface proteome data (Supplementary Fig. 2B–F) confirmed the preference of EBV lytic antigen-specific T cells to acquire an end-differentiated TEMRA phenotype, while latent antigen-specific T cells lacked a TEMRA contribution and contained more CM and EM cells. Differences contrasted not only lytic versus latent EBV-specific T cells but also indicated distinctions between each antigen specificity (Supplementary Fig. 2D, E). As such, these data suggest that each antigen-specific T cell population harbors unique phenotypic and transcriptional traits despite undergoing T cell priming simultaneously early in life.

Epigenetic profiling distinguishes antigen-specific signatures

ATAC-seq analysis of the open chromatin landscape is a powerful tool to profile the epigenetic state even of rare, antigen-specific T cells22,23. We performed ATAC-seq on fresh, FACS-purified EBV-specific CD8+ T cells against BMLF1, BRLF1, or LMP2 as well as sorted bulk populations of classical CM (CD45RA CCR7+), EM (CD45RA CCR7), or TEMRA (CD45RA CD28) cells (Supplementary Fig. 3A–C). We included naïve CD8+ T cell data from our previously published ATAC-seq dataset as a reference point24. Progressive differentiation as reflected by classical bulk subsets correlated with an increase in PC1, accounting for 50% of the variance (Fig. 2A and Supplementary Fig. 3D). Lytic BMLF1-specific cells showed TEMRA-like preference, and latent LMP2-specific cells CM-like clustering. Although being a lytic EBV antigen, BRLF1-specific cells showed broad distribution across PC1, suggesting a more amendable epigenetic fate and a less committed cell state. K-means clustering of ChromVar scores signifying inferred TF accessibility in ATAC-seq peaks25 (Supplementary Fig. 3E, F) as well as K-means clustering on differential peaks (Fig. 2B) confirmed the phenotypic dominance for BMLF1 as TEMRA- and EM-like and LMP2 as CM-like. Again, BRLF1-specific cell profiles were least constrained and showed signatures of both CM-like cells as well as TEMRA- and EM-like features. TF motif enrichment analysis suggested concerted actions of key TFs to promote T cell differentiation states and features in each K-means cluster (Fig. 2B and Supplementary Fig. 4A). For each cluster, we identified genes associated with nearby promoter and intragenic peaks and determined their putative biological contributions. Cluster 1 peaks were shared between T cells specific to lytic EBV antigen BMLF1 and BRLF1 cells; functional attributes of their associated genes included cytotoxicity, cytokine signaling, and TCR signaling as well as NK cell-like and innate immune features (Fig. 2C). Consistent with our flow cytometry analyses (Fig. 1), chromatin at the perforin (PRF1), granzyme B (GZMB) and CX3CR1 loci was more open in lytic EBV antigen-specific T cells. ZEB2, a master TF of CD8+ T cell terminal differentiation as well as NK maturation26,27,28, was the gene with the most differentially open peaks in K-means cluster 1 and may orchestrate many of these traits. K-means cluster 3 was most specific to lytic BMLF1-specific cells and may reflect higher TCR signaling and cytokine production, including IFNG (Supplementary Fig. 4B). Cluster 4 (CM-like signature) and cluster 5 (stem-like signature) were dominant for latent LMP2-specific cells and was enriched for pathways fundamental to T cell homeostasis including LEF1, Wnt29 and mTOR30 signaling, as well as features of T-helper differentiation (Fig. 2D and Supplementary Fig. 4C). Wnt signaling confers stem-like features in T cells suggesting that LMP2-specific cells are not only less differentiated but have a higher self-renewal capacity and homeostatic fitness. K-means cluster 7 was most similar between lytic BRLF1-specific and EM/TEMRA bulk cells and emphasized enhanced cytotoxicity, cytokine production and innate immune features via a different set of genes, including the top gene HELIOS (IKZF2) that fine-tunes T cell chromatin remodeling, differentiation and function31,32,33 (Supplementary Fig. 4D). Collectively, these data extend our observation on phenotypic and transcriptomic heterogeneity and show that T cell populations specific for distinct EBV antigens also differ at the epigenetic level. This finding underscores the critical need to investigate memory T cell fate decisions and aging trajectories at an antigen-specific and single-cell level.

Fig. 2: EBV-specific T cells of differing antigen specificities are distinguished in their chromatin accessibility landscapes.
figure 2

A PCA of ATAC-seq data of CD8+ bulk subsets and tetramer-positive, EBV-specific CD8+ T cells. PCA is based on the 5000 most variable accessible sites across all groups. B K-means clustering of sites differentially accessible in any binary comparison across all groups. Peaks of each cluster were analyzed for transcription factor motif enrichment via HOMER. Key transcription factor families of enriched motifs are shown at the right margin. C, D Genes corresponding to peaks within each K-means cluster were identified via ChIPseeker. Only promoter and intergenic peak-associated genes were used to depict the top 5 peak-associated genes (as defined having the highest number of differential peaks in the corresponding K-means cluster) as well as for pathway and annotation enrichment. Selected ATAC-seq peak tracks are shown for K-means cluster 1 (C) and 5 (D). One-sided Fisher’s exact tests with Benjamini-Hochberg correction were used for enrichment analyses. Source data are provided as a Source Data file.

Distinct differentiation and aging signatures

To examine memory T cell aging while accounting for antigen specificity and heterogeneity, we first compared the phenotypic adaptations of EBV-specific CD8+ T cells in young (< 40 years) and older adults (> 65 years) by tetramer staining and multiplex spectral flow cytometry. As expected, naïve and stem-like memory cells in the bulk CD8+ T cell population were markedly reduced with age, while end-differentiated TEMRA cells accumulated (Fig. 3A and Supplementary Fig. 5A–C). As previously described11,34, none of the EBV-specific T cell populations changed significantly in overall frequencies (Fig. 3B). Interestingly, only a subset of antigen-specific T cells, especially against latent EBNA3C, changed their phenotype. Unbiased reference mapping demonstrated that latent EBNA3C-specific T cells gained a TEMRA population with advanced age, while the latent LMP2-specific T cells largely retained their dominant EM phenotype across age (Fig. 3C–E and Supplementary Fig. 5D). Lytic BMLF1- and BRLF1-specific cells stayed consistently more differentiated and TEMRA-like in young and older adults. Statistical analyses by comparing probability vectors of each antigen specificity between young and older samples reached significance for only EBNA3C (Fig. 3E). Targeted flow cytometry gating on classical end-differentiation markers, TFs and effector molecules underscored the age-related acquisition of end-differentiation traits selectively in latent EBNA3C-specific cells (Fig. 3F and Supplementary Fig. 5E). Collectively, these data suggest that only a subset of T cells are sensitive to phenotypic adaptation with age, while other populations are capable of effectively maintaining their phenotype and avoiding broad age-related adaptation.

Fig. 3: Antigen-specificity correlates with age-related phenotypic changes.
figure 3

CD8+ T cells from peripheral blood of 9 older (O, > 65 years, mean age of 71.9 years) adults were analyzed by spectral flow cytometry as described in Fig. 1 and contrasted to those of 8 young (Y) adults from Fig. 1 and Supplementary Fig. 1. A UMAP reference plot of bulk CD8+ T cell subsets based on integrated data from Y and O adults, as well as their distributions shown separately for Y and O. B Frequencies of EBV tetramer-positive cells for different antigens. C Reference mapping results depicting side-to-side comparison of the phenotypic distribution of tetramer-positive cells in 8 Y and 9 O adults. D Bubble plot comparing distributions of tetramer-positive cells among CD8+ T cell clusters defined in (A). The bubble size and color reflect the proportion of antigen-specific T cells in a given cluster. E Probability vectors describing the subset distributions of tetramer-positive cells in Y and O adults were compared. Significance levels between the indicated pairwise comparisons are shown as bubble size and color. F Protein levels of indicated molecules in tetramer-positive cells of 8 Y and 9 O adults as assessed by spectral flow cytometry. Data show the median with 95% confidence intervals (B), median (D) or mean ± SEM (F). All datapoints represent distinct biological replicates. Data were compared by two-way ANOVA with Šídák’s multiple comparisons test (B, F) or by two-sided permutation tests (E). Source data are provided as a Source Data file.

BEAM-T captures the heterogeneity of antigen-specific cells

To gain a comprehensive and unbiased snapshot of aging trajectories of antigen-specific T cells, we performed single-cell sequencing of CD8+ T cells from 7 young (< 40 years) and 7 older adults (>65 years) with 4 modalities. We leveraged a novel method, Barcode Enabled Antigen Mapping of T cells (BEAM-T), to capture antigen specificity in addition to unbiased gene expression data, as well as surface proteome staining and TCR sequences. For each HLA-A*02:01-positive individual, we extended our analysis to additional antigen-specific populations and collected BEAM-T-positive T cells specific for 4 lytic EBV antigens (BMLF1, BRLF1, BMRF1, BALF1), 4 latent EBV antigens (LMP1, LMP2, EBNA1, EBNA3C) and 2 antigens (IE62, IE63) from the related varicella zoster virus (VZV; Fig. 4A). CD8+ bulk T cells were collected for comparison. We obtained data on 29,304 antigen-specific T cells that passed the quality control criteria. BALF1 and IE63 yielded too few cells for downstream analyses. Antigen-specific T cells were clustered into 17 clusters representing 8 broad, classical T cell subsets (Supplementary Fig. 6A, B) and including rare T cell types with NKT-like and MAIT phenotypes (Supplementary Fig. 6B, C). We found previously identified TCR CDR3 sequences against 5 EBV antigens, supporting the validity and potential of our approach (Supplementary Fig. 6D).

Fig. 4: Transcriptional profiles and clonal diversity of EBV-specific T cells in young adults depend on antigen-specificity.
figure 4

A Schematic of single-cell RNA-seq studies of antigen-specific CD8+ T cells from 7 young (Y, < 40 years, mean age of 31.9 years) and 7 older (O, > 65 years, mean age of 71.1 years) adults. BEAM-T (Barcode Enabled Antigen Mapping of T cells) was employed. Antigen-specific T cells were live, single cells that were CD8+ and stained positive for the BEAM-T conjugate (which included a PE fluorophore, see the Methods for additional information). Total CD8+ T cell subpopulations that were negative for the BEAM-T conjugate were included as well. B Bubble plot of antigen-specific T cell proportions among clusters defined in Supplementary Fig. 6A. The bubble size and color reflect the proportion of antigen-specific T cells in a given cluster. C Probability vectors describing the subset distribution of antigen-specific T cells were compared. Significance levels between the indicated pairwise comparisons are shown as bubble size and color. D PCA based on cell subset frequencies. E K-means clustering of differentially expressed genes in any binary pseudobulk comparison of antigen-specific memory T cell populations. Pathway and annotation enrichment in each K-means cluster and selected genes are shown on the right. F T cell receptor (TCR) diversity and clonality are shown as the Inverse Simpson index and Gini index, respectively. 7 Y and 7 O adults were assessed, but only cells with TCRα and TCRβ sequences in single-cell sequencing data and samples with ≥20 cells were used for the calculation. G Cumulative TCR clone numbers depicting the numbers of distinct antigen-reactive TCR chains ordered by decreasing clonal sizes plotted against the cumulative space they occupy. Line graphs with different colors illustrate the clonal size distributions for different antigen specificities. All datapoints represent distinct biological replicates. Data show the median (B) or mean ± SEM (F). Data were compared by two-sided permutation tests (D). Source data are provided as a Source Data file.

First, we assessed the heterogeneity of antigen-specific T cells in young adults. All populations except for latent LMP2 contained a large proportion of stem-like cells (Supplementary Fig. 6E). In line with our flow cytometry data, antigen-specific populations were heterogenous in their phenotype, but each antigen-specific memory T cell population exhibited their own unique subset distribution (Fig. 4B) that, with few exceptions, was significantly different to other antigen-specific populations (Fig. 4C, D and Supplementary Fig. 6F, G). K-means clustering of differentially expressed transcripts followed by pathway and annotation enrichment analyses corroborated the distinct feature of each antigen-specific population (Fig. 4E). For example, unlike all other populations, latent LMP2-specific cells were devoid of NK-like and cytotoxic genes, while these cells showed the highest expression of co-activating receptors CD27 and CD28.

When assessing TCRα/β sequences in our single cell-sequencing data, we found that diversity (Inverse Simpson index and Shannon index) as well as clonality (Gini index) of antigen-specific memory cells varied across a broad range that did not correlate with the classification into lytic and latent antigen (Fig. 4F and Supplementary Fig. 6H) and was irrespective of their dominant phenotype. This contrasts with the bulk level, where the end-differentiated TEMRA phenotype is typically associated with high clonality and lower T cell diversity35. In fact, the lowest diversity was seen for latent LMP2 cells that were devoid of TEMRA and had predominantly acquired an EM phenotype (Fig. 4F, G). Conversely, cells specific for VZV-IE62, an antigen expressed in the lytic and latent VZV life cycles, had the highest TCR diversity. These data show that antigen-specificity, more than differentiation state, influences repertoire diversity in young individuals.

Antigen-specific stem-like T cells are lost with aging

Next, we interrogated aging trajectories with consideration of three levels: phenotype, transcriptional profiles, and clonal diversity. The most striking age-related phenotypic change was the loss of the stem-like compartment, including putative naïve and stem-like memory cells across most antigen-specific T cells with age (Fig. 5A), which accounted for a clear separation of young and older samples in PCA based on subset frequencies (Fig. 5B and Supplementary Fig. 7A–C). A PCA focused on only the memory T cell compartment showed age-related differentiation state shifts that varied with the antigen. Specifically, latent antigen-specific T cell populations that had a low TEMRA contribution in young adults (especially EBNA1 and EBNA3C), accumulated this end-differentiated cell type in older age (Fig. 5C and Supplementary Fig. 7C), resulting in significantly different probability vectors (Fig. 5D). Lytic EBV-specific memory T cells were phenotypically more stable than those against latent antigens, with the exception of BMRF1, that frequently were MAIT-like cells in young adults that were lost with age. Latent LMP2-specific cells were an exception as they maintained their phenotypic state surprisingly well over the lifetime.

Fig. 5: Aging results in a loss of the T cell reserve of undifferentiated antigen-specific T cells across most antigens.
figure 5

Single cell-sequencing of antigen-specific CD8+ T cells from peripheral blood of 7 young (Y) and 7 older (O) adults as described in Fig. 4 and Supplementary Fig. 6. A Proportions of stem-like cells, as defined in Supplementary Fig. 6A, in different antigen-specific T cell populations. 7 Y and 7 O adults were assessed, but only samples with ≥ 20 cells were considered for this analysis. B PCA based on subset frequency distributions of antigen-specific cells across all clusters, including stem-like cells and memory cells (left) or in only memory cells after exclusion of antigen-specific cells expressing a stem-like phenotype based on transcriptome or cell surface markers (Supplementary Fig. 6A, B) (right). C Bubble plot depicting proportion of antigen-specific memory cells among T cell subset clusters defined in Supplementary Fig. 6A. D Comparison of probability vectors describing T cell subset distributions of different antigen-specific T cells. Y versus O comparisons were made only within each antigen. Significance levels between the indicated pairwise comparisons are shown as bubble size and color. Data show the mean ± SEM (A) or median (C). Data were compared by two-way ANOVA with Šídák’s multiple comparisons test (A) or by two-sided permutation tests (D). Source data are provided as a Source Data file.

Transcriptional signatures distinguish T cell aging trajectories

Antigen-dependent adaptations in T cell differentiation states with age were even more apparent when we assessed transcriptional trajectories. We compared memory T cells of young and older adults for each antigen-specific population by pseudobulk differential expression analyses and found highly variable amounts of differentially expressed genes (DEGs), between 45 and 701 genes, depending on the antigen (Supplementary Fig. 7D). Latent EBV-specific T cells typically exhibited higher numbers of DEGs, consistent with their higher sensitivity towards phenotypic change with age. However, although LMP2-specific cells appeared stable in their cell differentiation stage, these cells had the most transcriptional changes, including many unique DEGs (Fig. 6A), suggesting qualitative changes not reflected by T cell differentiation. Consistent with the notion of population-specific aging trajectories, we find minimal overlap of DEGs across cells with different specificities (Fig. 6B). Sets of DEGs tended to be detected in cells specific to similar types of antigens, such as the nuclear antigens EBNA1 and EBNA3C, but even there the overlaps were small.

Fig. 6: Gain of lymphocyte innateness in memory T cells is a common age-associated transcriptional signature.
figure 6

Pseudobulk differential expression of single-cell sequencing data from young (Y) and older (O) adults within each antigen specificity were compared, and differentially expressed genes (DEGs) were identified (see also Supplementary Fig. 7D). Only antigen-specific memory T cells were selected, and stem-like cells (see Fig. 5A) were excluded from consecutive analyses. A, B Number of unique (A) and shared (B) DEGs from O versus Y comparisons. C Heatmap and K-means clustering of DEGs identified in Supplementary Fig. 7D. Selected genes in each K-means cluster are listed on the right. D Pathway and annotation enrichment analysis of DEGs in K-means cluster 1. One-sided Fisher’s exact tests with Benjamini-Hochberg correction were used for statistical analyses. For enrichment analyses of other K-means clusters, see Supplementary Fig. 7E. No enrichments were found for K-means clusters 5 and 6. E Gene set enrichment analyses in single-cell sequencing data based on pseudobulk expression analyses of antigen-specific memory T cells comparing Y and O groups. For statistical analyses, see “Methods”. The lymphocyte innateness gene set was initially collated by Gutierrez-Arcelus et al.36. NES, normalized enrichment score.

To uncover transcriptional patterns common or specific to antigen-specific T cell populations, we performed K-means clustering on all DEGs (Fig. 6C). Of the 7 clusters, clusters 1 (C1) and 2 (C2) were particularly informative as they largely represented common aging signatures. C1 genes increased with age and included cytotoxicity genes (GZMB, PRF1, NKG7), advanced differentiation markers (KLRG1, LILRB1/CD85J) and genes typically expressed in NK cells (KLRD1, TYROBP). Pathway and annotation enrichment on C1 genes substantiated that a gain in cytotoxicity, NK cell-like features, as well as cytokine and TCR signaling, were common traits of antigen-specific memory T cell aging (Fig. 6D). C2 genes were often reduced with age and included TCF1, LEF1, ID3, CCR7 that are part of essential pathways promoting T cell longevity, self-renewal capacity, and lymph node homing properties, although enrichment for these pathways did not reach significance (Supplementary Fig. 7E). Antigen-specific T cells against VZV-IE62 showed similar age-related transcriptional patterns suggesting conserved features of antigen-specific memory T cell aging, at least across chronic viral infection models. Gene expression of LMP2-specific T cells was a notable exception, showing no change in C1 or C2 genes, but instead favored C3, C4 and C5 clusters. Interrogation of the respective DEGs and enriched pathways suggested an age-related increase in stress-response and survival capacity (BCL2, DDIT3, DDIT4) as well as pro-inflammatory NFκB signaling (RELB, NFKBIA) (Supplementary Fig. 7E). However, these cells also lost features of T cell longevity (MYC, RUNX1, CD27, SELL/CD62L), suggesting complex adaptations in memory T cell fitness. C6 and C7 genes were particularly different in BMRF1-specific T cells.

Loss of essential T cell fidelity genes (C2) and gain in NK cell traits (C1) suggested a rewiring of adaptive and innate lymphocyte immunity programs. Specifically, cytotoxic features have been previously correlated with lymphocyte innateness, while TCF7, MYC and other key TFs have been correlated with T cell adaptiveness36. To test this hypothesis, we performed gene set enrichment analyses (GSEA) for gene modules previously defined as lymphocyte adaptiveness or innateness by examining the trajectory from CD4+ over CD8+, MAIT, iNKT, and γδ T cells to NK cells36. We found a highly significant, positive enrichment for innateness genes in 7 out of 8 antigen-specificities in older adults (Fig. 6E), indicating that a gain in innateness is a common signature of memory T cell aging. Conversely, the adaptiveness signature was lost with aging in 6 out of the 8 antigen-specific T cell populations (Supplementary Fig. 8A), indicating a transcriptional, age-sensitive switch towards innate T cell functions such as those seen in γδ T cells and NK cells. BRLF1-specific T cells consistently differed from other specificities in these enrichment analyses. In contrast, GSEA for other, classical states of cellular dysfunction, such as T cell exhaustion or cellular senescence, were not significantly enriched in EBV-specific T cells of older adults (Supplementary Fig. 8B–D).

Repertoire contraction versus diversification with age

Aging reduces T cell diversity of the CD8+ T cell compartment predominantly due to a loss of CD8+ T cells with a naïve or stem-like phenotype, while increasing clonality37. Consistent with this observation in bulk T cells, the stem-like compartment of antigen-specific T cells that existed in young adults was essentially lost with age accounting for a contraction in diversity (Fig. 7A). In contrast, TCR diversity of antigen-specific T cells expressing memory markers was largely unchanged with advancing age for most specificities (Fig. 7B). Clonality as determined by the Gini index was increased in 2 (EBNA1, VZV-IE62) of the 6 assessed antigen-specific populations, while unchanged or even reduced with age for the other antigens (Fig. 7C). Similarly, the fraction of large T cell clones was only increased for these same 2 specificities (Fig. 7D). For a more intuitive illustration of T cell diversity, we plotted clones ordered by size rank against the cumulative space they occupy. Increased clonality with age, i.e., fewer clones occupying a given space, was seen for the responses to VZV-IE62, EBNA1, and EBNA3C (Fig. 7E). In contrast, a repertoire diversification was seen for the antigens BMLF1, BRLF1, and LMP1, possibly indicating a recruitment of stem-like cells into the memory T cell compartment over adult lifetime. Taken together, aging depleted the source of antigen-specific stem-like cells while the repertoire diversity of antigen-specific memory cells did not change much or even increased, depending on the antigen.

Fig. 7: Memory T cell receptor diversity of antigen-specific T cells is maintained with aging.
figure 7

A TCR diversity and clonality are shown as Shannon, Inverse Simpson, and Gini indices in antigen-specific CD8+ T cells mapping to the stem-like clusters (see Supplementary Fig. 6A). Only cells with TCRα and TCRβ sequences in single-cell sequencing data were used for the calculation. Due to low cell numbers, indices were calculated on the combined populations of all antigen-specific T cells with a stem-like phenotype (n = 7 Y and n = 7 O). B T cell diversity indices as in (A) but calculated on antigen-specific T cells with a memory phenotype. Data from older (O) individuals were contrasted to younger (Y) adults’ data described in Fig. 4F and Supplementary Fig. 6H. 7 Y and 7 O adults were assessed, but only samples with ≥20 cells were considered for this analysis. C Gini index as in (A) to estimate T cell clonality in antigen-specific T cells with a memory phenotype. Data from O individuals were contrasted to Y data described in Fig. 4F. 7 Y and 7 O adults were assessed, but only samples with ≥ 20 cells were considered for this analysis. D Clonal size distributions for each antigen specificity are shown as stacked bars for Y and O adults. E Numbers of distinct antigen-specific TCR chains ordered by decreasing clonal sizes plotted against the cumulative space they occupy. The number of distinct antigen-specific TCR chains from Y adults are as shown in Fig. 4G and are contrasted to results from O adults (dotted lines). All datapoints represent distinct biological replicates. Data show the mean ± SEM (AC). Data were compared by two-tailed, unpaired ttests (A), or two-way ANOVA with Šídák’s multiple comparisons test (B, C). Source data are provided as a Source Data file.

The TCR sequence contributes to T cell phenotypes and aging trajectories

Finally, we investigated factors affecting the phenotypic diversity and aging trajectories of antigen-specific T cells. We explored the contribution of the TCR and TCR avidity considering that the downstream TCR signaling strength can shape T cell differentiation and functionality38,39,40. TCR V segment usage was highly diverse for each antigen specificity, with little age-related changes (Supplementary Fig. 9A, B). Motif enrichment in TCRβ CDR3 regions identified shared motifs within antigen-specific T cell populations, however, these motifs were typically found in a small proportion of clones and tended to be shared by young and older individuals (Supplementary Fig. 9C and Supplementary Data 1).

To test whether identical TCRs drive identical T cell phenotypes, we identified clones with identical TCR amino acid sequences shared by 2 independent individuals in our dataset. We identified 14 such clones that we projected on the UMAP to assess their phenotype (Fig. 8A and Supplementary Fig. 10A). Identical clones displayed a preferred T cell phenotype with similar albeit non-identical features. Most dissimilarity was observed in identical clones that were shared between a young and an older individual (Fig. 8A), suggesting that age-related changes contribute to phenotypic shifts independent of the TCR sequence. Considering that individual clones can acquire a range of phenotypes, we performed TCR clonotype tracking across T cell subsets for each antigen (Fig. 8B). Indeed all assessed antigen-specific T cell populations included clones that were found across multiple T cell subsets, suggesting a phenotypic diversification despite the same TCR sequence, in line with reports that a single naïve T cell can generate multiple phenotypic states41,42. This suggests additional factors beyond TCR sequence that dictate T cell phenotypes, for example, TCR avidity.

Fig. 8: TCR sequence and avidity contributes to T cell phenotypes and aging trajectories.
figure 8

A UMAP projection of 4 expanded T cell clones that have identical TCR amino acid sequences in 2 individuals, marked in pink or blue. B Bubble plots of TCR clone sharing across T cell subsets for each antigen and age. TCR similarity scores (TRSS) are shown as both color gradient and bubble size. C Violin plot of TCR avidity estimates for each single antigen-specific memory cell based on BEAM-T binding and normalized to CD3E expression. Violins show summary data across all 7Y or 7 O individuals. Only antigen-specific memory T cells were selected, and stem-like cells were excluded from this and consecutive analyses. D Bubble plots of pairwise statistical analyses comparing TCR avidities across antigens in Y or O individuals. The bubble color represents the difference in avidity between 2 antigens, and the bubble size reflects the significance levels. E Correlation of TCR avidity and Gini clonality index per antigen and age, summarized for all 7Y or 7 O individuals. F Correlation of TCR avidity and innateness or adaptiveness gene scores per antigen and age, summarized for all 7Y or 7 O individuals. G Correlation coefficients of avidity and adaptiveness or innateness gene scores for each individual (n = 6 Y and n = 7 O) across antigens per individual, or in combined data using a random effects model. Data show the median (C, E, F) or as median with 95% confidence interval (G). Data were compared by two-sided permutation tests, or linear regression analyses and two-tailed F tests (E, F). For statistical analyses of (G), see “Methods”. Source data are provided as a Source Data file.

We computed a TCR avidity score based on the binding of a T cell to the antigen peptide/BEAM-T conjugate, normalized to CD3E expression to account for TCR density per single cell. We found that TCR avidity differed between antigen-specific T cell populations, with LMP2-specific cells showing the overall highest avidity (Fig. 8C, D). Avidity appeared to decrease with age in an antigen-dependent fashion. Avidity was positively correlated to T cell clonality as assessed by the Gini index (Fig. 8E)43,44. To assess the contribution of TCR avidity to aging trajectories, we computed gene scores for the aging-sensitive adaptiveness and lymphocyte innateness gene sets described earlier (see Fig. 6). Gene scores followed the expected trends across cell types, antigens, and age (Supplementary Fig. 10B–E) but were not correlated with the Gini index estimating T cell clonality (Supplementary Fig. 10F). Correlation analyses between TCR avidity and adaptiveness gene scores controlling for antigen specificity or individuals consistently showed a positive correlation suggesting that higher T cell avidity favors adaptive, stem-like T cell traits (Fig. 8F, G). A tendency towards an inverse correlation between TCR avidity and innateness gene score was detected as well. Considering the loss of stem-like signatures in many antigen-specific T cells and the concomitant reduction of T cell avidity with age, it is conceivable that high avidity antigen-specific T cells are lost with age, possibly through dysfunction such as exhaustion45, while those with lower avidity, innate-like character and elevated cytotoxic potential persist into older ages.

Discussion

Here, we describe that CD8+ T cells specific for different EBV-derived peptides differ in the changes that they undergo with aging. These differences correlate with differentiation states that T cells have developed with the initial infection and that are preserved into young adulthood (Fig. 9). Globally, T cells that are specific for lytic EBV antigens are already more biased towards effector end-differentiation than those specific for latent antigens in young adults. Conversely, T cells specific for latent antigens undergo more progressive changes with age as determined by phenotypic analyses, expression of transcription factors, and transcriptomic signatures. Heterogeneity of age-associated changes in EBV-specific T cells existed even within the sets of lytic and latent antigens that we tested, indicating that this distinction between lytic and latent antigens is an oversimplification. While a common aging trajectory of most T cell specificities was a loss in the expression of genes related to adaptive immunity and a gain in innate, NK cell-related immunity, age-associated changes in the transcriptome were not widely shared between T cells specific for peptides derived from different EBV proteins, and each antigen-specificity had its own dominant aging signature.

Fig. 9: Model summarizing aging trajectories of EBV antigen-specific CD8+ T cells.
figure 9

T cells primed early in life acquire highly heterogeneous phenotypes depending on the antigenic exposure. These memory T cells undergo phenotypic and transcriptional changes with age, although to varying extents and depending on their initial phenotypes. Age-related changes commonly occurring in most investigated antigen-specific T cell populations are limited to a loss in stem-like cells and/or traits, and a gain in NK-like, cytotoxic features. Most age-related changes selectively occur in certain T cell populations and include a gain in TEMRA cells, loss in MAIT cells, changes in NFkB and survival signaling, and changes in TCR avidity, diversity, and/or clonality.

How immune memory is maintained over the lifetime is central to healthy aging and has important implications for vaccine strategies. While significant progress has been made in the last decade in describing age-associated changes in the global T memory cell compartment, one of the major limitations of current datasets is that they do not allow conclusions at the level of antigen-specific T cells. Studies at the population level of CD8+ T cells have described shifts to more differentiated cells, i.e., effector, TEMRA, and T cells expressing activation or exhaustion markers46. While phenotypic studies alone cannot completely capture the heterogeneity of CD8+ memory T cells3, omics approaches, including single-cell transcriptome studies, have significantly expanded our insights into human immune aging in recent years. Datasets on PBMCs of > 100 healthy individuals have identified an age-associated CD8+ T cell subpopulation that was characterized by the expression of granzyme K47,48. A CD8+ memory T cell population marked by NKG2C+ GZMB, however, decreased with age49. While single-cell 5’-RNA-seq allows to identify TCR sequences and therefore assigns transcriptome signatures to clonal specificities, conclusions from cross-sectional studies on the durability of antigen-specific T cells are difficult given the immense TCR diversity of even antigen-specific T cells across individuals. Longitudinal studies are usually limited to up to 2 years after vaccination24,50. The currently longest longitudinal study by Sun et al.51 only focused on changes in the TCR repertoire over a timespan of ~ 9 years using the Baltimore Longitudinal Study of Aging.

Elegant murine studies provided surprising insights into how antigen-specific memory T cells evolve with repeated re-stimulation over lifetime. Soeren et al.7 used iterative stimulations and adoptive transfers in mice and showed that mouse memory T cells remained fully functional over 51 immunizations without evidence of senescence or exhaustion, across a time span of > 10 years, far exceeding a mouse’s lifespan. Functionality was sustained despite chromatin remodeling at exhaustion markers but was dependent on sufficient rest between stimulation events.

To probe the evolution of human antigen-specific memory T cells over the lifetime and determine whether they develop genetic hallmarks of cellular senescence or exhaustion, we examined epitope-specific CD8+ T cells in EBV infection. The EBV system allows to compare the age-dependent evolution of CD8+ T cells specific for multiple lytic and latent antigens that were primed simultaneously early in life. The primary immune response in acute mononucleosis constitutes a dramatic expansion of CD8+ T cells specific to lytic proteins that can make up 50% of all CD8+ T cells9. Cells responding to latent epitopes are present at lower frequencies, contributing up to 5% of CD8+ T cells. After the acute infection, most effector T cells are culled, and memory T cells specific to lytic and latent antigens develop at a broadly similar immunodominance hierarchy to those seen in effector T cells. Importantly, Hislop et al.12 found that by the time the primary EBV infection resolves, CD8+ memory T cells have developed distinct differentiation states dependent on their epitope specificity. On average, responses to lytic antigens have a more progressed effector phenotype. In our young study population who were in their early adulthood up to 40 years old, we observed that EBV-specific T cells still differed in their differentiation states based on their antigen specificity These states loosely correlated with the classification into lytic and latent proteins, however, transcriptome and chromatin structure data yielded a more complex picture with each antigen specificity having a unique signature. Given the data by Hislop et al.12 that these differences in T cell differentiation states are established early during infection, they are obviously maintained over more than a decade and gave us the opportunity to examine whether they have different aging trajectories, e.g., whether more frequent and more differentiated T cells are more susceptible to develop features of exhaustion or cellular senescence. This was not the case, and transcriptome changes were more pronounced for less differentiated specificities, including an increased expression of granzyme B, in contrast to recently reported data on cells specific for the influenza matrix protein52. Granzyme K, recently described as a common signature of aging48,49,53, was widely expressed in our EBV-specific cells, irrespective of age.

Several variables could influence the evolution of the antigen-specific T cell over time. The observed aging trajectories encompass the aging of the host as well as the durability of the antigen-specific memory T cells, which may be influenced by the early priming event. Moreover, memory T cells may be subject to re-stimulation depending on when and how often EBV antigens are re-expressed. Lytic antigens are highly re-expressed upon viral reactivation. In contrast to other herpes viruses, such as varicella zoster virus and herpes simplex, healthy, non-immunocompromised individuals do not experience clear clinical symptoms that could be attributed to viral reactivation even at older age. However, subclincial reactivation cannot be excluded. In respect to latent antigens, we selected antigens that are expressed at different stages of latency. In healthy individuals, EBV is present in the nucleus as an episome tethered by EBNA1, and only non-coding RNAs such as EBERs and BARTs are expressed54. LMP1 and LMP2 are expressed at latency type II characteristic of selected malignancies, such a Hodgkin’s disease. EBNA3A is expressed in latency type III as seen in lymphomas of immunocompromised individuals. We could not find a clear correlation of the different aging trajectories with the differences in latency expression types. For example, the response to LMP1 and LMP2 was quite different, although these antigens are co-expressed. Even the response to the three lytic antigens was quite different, with BMLF1 prompting highly differentiated and BRLF1 encompassing heterogeneous differentiation stages.

The T cell response to the related herpes cytomegalovirus virus (CMV) and the lytic CMV protein pp65 shows some parallels to the EBV BMLF1 response, with the generation of end-differentiated TEMRA cells early in the infection and already at a younger age55. Similarly, overall CMV-specific T cell numbers do not change over lifetime while their phenotype adapts with advanced age and becomes exceedingly more NK-like11.

Analogous to our study, van de Sandt et al.52 examined how epitope-specific T cells evolve across the human lifespan by examining the phenotype and transcriptome of CD8+ T cells directed at the prominent influenza matrix M158–66 epitope over the lifetime. Unlike total CD8+ T cells, influenza-specific cells did not display a TEMRA phenotype with age, suggesting that influenza M1 stimulation does not trigger terminal differentiation per se. The transcriptome of older influenza-specific T cells correlated with less differentiated cell states, lacking a gain in effector function but with expression of cytotoxic molecules. Similar to our study, there was no evidence for cellular senescence or exhaustion, although cells lost activation and proliferative capacity. Unlike to our findings, these age-associated changes in transcriptomes were correlated with a shift in the TCR repertoire, suggesting that they were not due to T cell aging but recruitment of a new set of specificities into the memory pool.

The memory T cell compartment is highly dynamic. Individual cells have limited lifespans, certainly considerably shorter than the duration of immunological memory. Memory persistence is therefore conferred by a population of cells that are individually short-lived56. This scenario should favor a selection for fitness in the repertoire of EBV-specific T cells that could diverge for different epitopes. Indeed, previous observations suggested that lytic antigen-specific effector T cells contract more than those specific for latent antigens in the early stages after infection12. We observed that the frequencies of all EBV antigen specificities are maintained into older age. TCR diversity was variable between different specificities, with BMRF1- and LMP1-reactive T cells having only a few clonally expanded memory cells. Previously reported longitudinal studies up to 5 years have reported persistence of individual clonotypes57,58,59. In our cross-sectional comparison, we see a lack of TCR singletons expressing a stem-like phenotype with age, in line with a general loss of bulk naïve CD8+ T cells60. Either these singletons are truly lost or they are expanded and recruited into the memory compartment – a phenomenon previously reported after infectious mononucleosis in young adults61 and for VZV-specific CD4+ T cells after vaccination in older adults62. Indeed, in analyses restricted to phenotypically defined memory T cells, we observe less clonality and more diversification of the TCR repertoire for some EBV epitope-specific T cells with age.

Our data support the notion that controlling for T cell heterogeneity, including antigen specificity, is needed to define age-associated changes that are clinically relevant. Understanding whether and how age affects the diversity, function, and durability of memory T cells has implications for designing effective vaccination strategies in older adults, who are exceedingly more vulnerable to infections1,2, as well as for optimizing T cell therapies in cancer patients. Specifically, understanding fidelity of EBV-specific T cells is relevant as EBV is a major oncogenic stimulus for malignancies, many of which are more often diagnosed in older adults18. In EBV-related nasopharyngeal carcinoma or Hodgkin’s lymphoma, adoptive transfer of EBV-specific T cells is successfully pursued as cellular immunotherapy63,64,65. In this approach, autologous or allogeneic T cells are stimulated ex vivo with EBV antigens LMP1, LMP2, and/or EBNA1, expanded, and infused into the cancer patient. Total EBV-specific T cell numbers and their phenotype show some association with stable disease66 and extended overall survival64, but further optimization of the transferred EBV-specific T cells will be beneficial to reduce variability and improve efficacy for this cell therapy approach.

Methods

Human study population

Peripheral blood mononuclear cells (PBMCs) from leukoreduction system chambers of 38 blood or platelet donors were purchased from the Mayo Clinic Blood Donor Center, Department of Laboratory Medicine and Pathology - Component Laboratory. Samples were de-identified except for age and sex information. Samples from male and female individuals were used. Young adults were 21 to 40 years old, and older adults were 65 years or older. In addition, PBMCs from 27 healthy volunteers without acute or active chronic disease, and no history of cancer or autoimmune disease, were recruited from a healthy aging registry at Stanford University. Chronic diseases were permitted if controlled by medication. Information on samples used in key experiments is summarized in Supplementary Data 2. Studies involving human participants were approved by the Stanford University Institutional Review Board and the Mayo Clinic Institutional Review Board, and all participants gave written informed consent.

Primary human cells

PBMCs were purified by density centrifugation with Lymphoprep (StemCell Technologies, #07861). Alternatively, T cells were isolated with the EasySep Human T Cell Enrichment Kit (StemCell Technologies, #19051). Cells were either directly used or frozen in aliquots in Bambanker freezing medium (Lymphotec Wako Chemicals, #302-14681). For ATAC-seq, fresh cells were used; for flow cytometric phenotyping analyses, fresh or frozen cells; for single cell sequencing, frozen cells that were rested before experiments were used. Total CD8+ T cells were isolated using the EasySep Human CD8+ T Cell Isolation Kit (StemCell Technologies, #17953). Screening for HLA-A*02-positive individuals was performed with an anti-human HLA-A2 antibody (BioLegend, # 343326). For single cell sequencing, total CD8+ T cells were purified as above, resuspended in RPMI-1640 medium (Sigma Aldrich, #R8758) supplemented with 5% filtered human AB serum (Sigma Aldrich, #H4522), 100 units /mL penicillin, and 100 units /mL streptomycin (Sigma Aldrich, #P0781) and rested in a 37 °C humidified incubator (5% CO2, atmospheric O2) for 16 h. Cells were collected and washed before subjecting them to BEAM-T and antibody staining.

Flow cytometry

For phenotyping of EBV antigen-specific CD8+ T cells, purified CD8+ T cells were first stained with tetramers against EBV antigens (BMLF1, BRLF1, LMP2, EBNA3C, see Supplementary Data 3). About 5 million cells were stained with 5 μL of each tetramer (containing 0.2 μg/μl MHC monomer conjugated to 0.05 μg streptavidin and fluorophore) and PBS containing 2% fetal bovine serum (FBS) (GeminiBio, #900-108) in a total volume of 50 μL. After 45 min of staining at 4 °C, diluted antibodies against cell surface proteins (see Supplementary Data 3) and a viability dye (LIVE/DEAD Fixable Blue, Thermo Fisher Scientific, #L23105) were added. Cells were stained for an additional 30 min at 4 °C. For the cell surface panel, cells were washed with PBS containing 2% FBS and fixed with Cytofix fixation buffer (BD Biosciences, #554655) for 30 min at 4 °C, washed and subjected to spectral flow cytometry. For the intracellular transcription factor panel, cells were washed with PBS containing 2% FBS and fixed with the eBioscience Foxp3 / Transcription Factor Staining Buffer Set (Thermo Fisher Scientific, #00-5523-00) according to the manufacturer’s instructions. Briefly, cells were fixed for 30 min at 4 °C, washed and permeabilized. Antibodies against intracellular proteins (see Supplementary Data 3) were diluted in permeabilization buffer, and staining was performed for 45 min at room temperature in the dark. After washing, cells were subjected to spectral flow cytometry. A 5-laser Cytek Aurora (Cytek Biosciences) running SpectroFlo v3 was used to collect data. UltraComp eBeads compensation beads (Thermo Fisher Scientific, #01-2222-42) were used for single staining controls. At least 1 million cells were collected, and at least 100 tetramer-positive cells were used for analyses. A representative gating strategy is shown in Fig. 1C.

Flow cytometry analyses

FlowJo v10 was used to analyze data. Live, single CD3+ CD8+ T cells were analyzed for tetramer-staining or classical phenotypic subset markers: naïve (CD45RA+ CCR7+ CD28+), central memory (CD45RA CCR7+ CD28+), effector memory (CD45RA CCR7) and TEMRA cells (CD45RA+ CCR7 CD28). Proteins of interest were analyzed in tetramer-specific cells and each subset. Representative gating schemes are shown in Supplementary Fig. 1C–E.

For clustering and reference mapping analyses, samples were randomly down-sampled to 50,000 CD8+ T cells per individual by the FlowJo plugin DownSampleV3. Channel values of these cells and all tetramer-positive cells were exported from FlowJo and imported into Seurat67 v4.2.1 for in-depth analyses. The raw counts were merged and transformed using centralized log ratio (CLR) across cells using the NormalizeData function. Data were scaled, and principal component analysis (PCA) was performed using the ScaleData and RunPCA functions, respectively. Samples from different experiments were integrated by diagonalized canonical-correlation analysis (CCA) that identifies anchors to act as guides for dataset integration using the function FindIntegrationAnchors and IntegrateData67. The integrated matrix underwent scaling and PCA calculation, generating the Uniform Manifold Approximation Projection (UMAP) along with a uniform manifold optimization (uwot) model for reference mapping via the RunUMAP function. Utilizing the PCA, the graph-based Louvain clustering was performed via the FindNeighbours and FindClusters function from 0.2 to 1 to identify an optimal cluster resolution. Data visualization was performed via Seurat.

To map tetramer-positive cells to the reference dataset, tetramer-positive samples were pre-processed as described above, including CLR transformation, identification of variable features using NormalizeData and FindVariableFeatures. Subsequently, tetramer-positive cells were aligned with the reference to establish anchors, facilitating the transfer of cell type labels from the reference to the samples using the FindTransferAnchors and TransferData functions, respectively67. We summarized the overall distribution of tetramer-positive cells across reference clusters. These proportions were utilized to compute the principal component for proper representation of the samples in a lower dimension.

ATAC-seq and analysis

CD8+ tetramer-positive cells for BMLF1 (GLCTLVAML), BRLF1 (YVLDHLIVV), and LMP2 (CLGGLLTMV) antigens as well as CD8+ populations of central memory (CD45RA CCR7+), effector memory (CD45RA CCR7), and TEMRA cells (CD45RA+ CD28) were collected via FACS in 3 batches. A gating scheme is shown in Supplementary Fig. 11. Representative plots are shown in Supplementary Fig. 3B. We did not include EBNA3C tetramer-specific T cells in these experiments. ATAC-seq24,68 was performed on FACS-collected, pelleted T cells that were washed with RSB buffer containing 10 mM Tris-HCl (pH 7.4) 10 mM NaCl, and 3 mM MgCl2. Cells were lysed in RSB buffer containing 0.1% NP40 and 0.1% Tween-20. The obtained chromatin pellet was resuspended in a transposase reaction mix containing Nextera Tn5 transposase (Illumina, #FC-121-1030) and incubated at 37 °C for 30 min. The transposed DNA was purified via the MinElute kit (Qiagen, #28004) and PCR-amplified (NEBNext High-Fidelity 2x PCR Master Mix, New England Biolabs, #M0541), quantified and sequenced on an Illumina HiSeq4000 instrument. We included naïve CD8+ T cell data from our previously published ATAC-seq dataset24. We aligned ATAC-seq reads to the hg38 reference genome and normalized data without batch correction22,23. Peaks that were present in at least 6 samples were included. Peaks with low read counts across sample groups were removed using the filterByExpr function69. PCA was calculated using regularized log-transformed counts from the top 5000 variable peaks using the plotPCA function in DESeq2. Peak counts were normalized using conditional quantile normalization (CQN), taking into consideration GC bias and length of the peaks for each sample. The normalized data were modeled for mean-variance trend for each gene, incorporating the sample weight adjustment using the voomWithQualityWeights function in limma70. Differential peaks (adjusted P-value < 0.05) were calculated across all populations by fitting each contrast to the linear model and employing empirical Bayes moderation. Differential peaks were clustered using K-means clustering with the default parameters. Clustered differential peaks were used for TF binding site (TFBS) prediction using HOMER71. TFBS prediction was also performed on all peaks using the chromVAR package25. Peaks were annotated using ChIPseeker with a region range of 1 kb across the transcription start site (TSS) and using TxDb.Hsapiens.UCSC.hg38.knownGene annotation72. Genes that were annotated as promoter or intragenic peaks were selected for enrichment analyses using EnrichR73, focusing on Gene Ontologies, Reactome, BioPlanet, KEGG, and ChEA gene sets. The averaged bigwig files were used to visualize peak profiles using the Integrated Genome Viewer (IGV)74.

Barcode Enabled Antigen Mapping (BEAM)-T and single cell sequencing

MHC monomers (Chromium Human MHC Class I A*02:01 Monomer Kit, 10x genomics, #1000542) were loaded with EBV peptides (GenScript) and incubated with the PE-containing BEAM conjugate (Chromium Single Cell 5’ BEAM Core Kit PE, 10x genomics, #1000539) according to the manufacturer’s instructions. Peptides for BMLF1 (GLCTLVAML), BRLF1 (YVLDHLIVV), BMRF1 (TLDYKPLSV), BALF4 (FLDKGTYTL), EBNA1 (FMVFLQTHI), EBNA3C (LLDFVRFMGV), LMP1 (YLQQNWWTL), LMP2 (CLGGLLTMV), VZV-IE62 (ALWALPHAA), and VZV-IE63 (RLVEDINRV) were custom-synthesized by GenScript. Loading of each peptide onto the MHC monomer was confirmed prior to use. A negative control peptide was provided within the kit and included in all analyses.

Magnetic-activated cell sorting (MACS)-purified CD8+ T cells from peripheral blood were subjected to BEAM-T staining according to the manufacturer’s instructions. Four to five individuals per experiment (3 experiments total) were processed. Briefly, about 3-4 million CD8+ T cells were blocked with Fc receptor blocking solution (Human TruStain FcX, BioLegend, #422302) for 10 min on ice and stained with BEAM-T assemblies for 15 min on ice, followed by cell surface marker antibodies plus unique hash-tagged CITE antibodies for 30 min on ice. Cells were washed 3 times with PBS containing 2% FBS before 7-AAD viability staining solution (BioLegend, #420404) was added, and cells were subjected to FACS (BD FACSAria 4-laser digital flow cytometer with FACSDiva v8 software). Cell sorting was performed by the Mayo Clinic Microscopy and Cell Analysis Core Flow Cytometry Lab. A gating scheme is shown in Supplementary Fig. 12. Live, single CD3+ CD8+ PE+ cells representing BEAM-T+ antigen-specific cells and PE cells representing total CD8+ T cells were sorted. After FACS, cells were counted and equal cell numbers of PE+ and PE cells of each individual were pooled. The pooled sample was then stained with diluted TotalSeq-C cell surface marker antibodies (BioLegend, Supplementary Data 3) in PBS containing 2% FBS for 30 min at 4 °C. Cells were washed twice with PBS containing 2% FBS, and cell viability was determined with Trypan Blue. Cell viability was consistently > 90%. Cells were subjected to high-throughput single cell capture via the Chromium X controller (10x genomics), Chromium Next GEM Single Cell 5’ HT Kit v2 (Dual Index, 10x genomics, #PN-1000377) and gel beads (10x genomics, #1000376). About 40,000 cells were targeted for cell capture for each side of the HT chip. Four libraries were generated from each captured sample: 5’ gene expression, feature barcodes (for TotalSeq-C antibodies including hashtag antibodies and cell surface marker, 5’ Feature Barcode Kit, 10x genomics, #PN-1000541), TCR sequences (Chromium Single Cell V(D)J Amplification kit, 10x genomics, #PN-1000252) and BEAM-T antigen specificity. All captured samples and libraries were sequenced in one flow cell using a 100 cycle kit (paired-end) on a NovaSeq 6000 S4 (Illumina) targeting 20,000 read pairs/cell for gene expression, 5000 read pairs/cell for feature barcodes, 5000 read pairs/cell for TCR sequences, and 5000 read pairs/cell for BEAM-T. Cell capture and library construction was performed by the Mayo Clinic Medical Genome Facility - Genome Analysis Core, and sequencing was performed by the San Diego Nathan Shock Center Heterogeneity of Aging Core.

Single-cell sequence analyses

Quality assessment of the fastq reads were conducted using FastQC. Subsequently, the four modalities (gene expression, GEX; CITE antibodies, ADT; T cell receptor sequences [V(D)J], and antigen specificity by BEAM-T, Ag) were processed using 10x Genomics Cell Ranger v7.1.0 multi for each experiment75. GEX was aligned to the Cell Ranger-provided GRCh38-2020-A reference, and V(D)J reads were aligned to the Cell Ranger-provided vdj-GRCh38-alts-ensembl-7.1.0 reference. For ADT and Ag, unique feature barcodes were utilized to extract the specific unique molecular identifier (UMI) associated with both modalities. Sample demultiplexing was carried using hash-tagged CITE antibodies and 10x Genomics cellranger v.7.1.0 multi75. We noticed that about half of the good-quality cells that were identified in CellRanger could not be demultiplexed via antibody-based demultiplexing due to the low UMIs of hash-tagged CITE antibodies. To overcome this challenge, variant-based demultiplexing based on GEX was implemented for each sample using souporcell76. Comparison of the hash-tag-based demultiplexing and variant-based demultiplexing revealed very high accuracy (> 0.98), sensitivity (> 0.97) and specificity (> 0.99) across all experiments.

Poor-quality cells were excluded based on the following criteria: cells with < 200 or > 5000 expressed genes, > 10% UMIs mapping to mitochondria-encoded transcripts and > 10,000 ADT UMIs. In addition, DoubletFinder was used to remove cell doublets. To differentiate antigen-specific cells from the bulk CD8+ T cell population, Ag UMIs were modeled on beta distribution by 10x Genomics Cell Ranger v7.1.0 multi75 to calculate the probability of antigen specificity (known as antigen specificity score) in comparison to the negative control peptide provided by 10x Genomics. Cells with antigen specificity scores of > 50 were classified as putative antigen-specific CD8+ T cells. Cells with scores of < 50 and with low Ag (BEAM-T) UMIs (< 40 UMIs for experimental batch 1 and < 75 UMIs for experimental batches 2 and 3) were classified as bulk CD8+ T cells. To annotate the identity of antigens for each cell, the antigen with the highest antigen specificity score was selected. To account for possible cross-reactivity, also antigens that fall within an antigen specificity score range of 10 from the maximal antigen specificity score were selected. Cells specific to BALF4 and VZV-IE63 were too infrequent and, thus, these specificities were excluded from further analysis. The bulk T cell population and antigen-specific samples were processed similarly for GEX and ADT. GEX samples were individually normalized using SCTransform v277, while CLR transformation was applied for ADT counts. Subsequently, samples from different experimental batches were integrated using diagonalized CCA to identify mutual nearest neighbors that acted as anchors for integration, employing the functions FindTransferAnchors and IntegrateData. Integrated data from GEX and ADT were scaled and PCA was calculated to integrate the modalities using weighted nearest neighbors (WNN) through the FindMultiModalNeighbors function. Based on the WNN, we generated the combined UMAP using the RunUMAP function. Further clustering was performed using the smart local moving (SLM) algorithm67. For data visualization, we addressed dropout in GEX and ADT via Markov affinity-based graph imputation of cell (MAGIC) that denoises the matrix and imputes values across similar cells via data diffusion78. Cell types were annotated for each cluster based on ADT cell surface markers and GEX gene markers identified using the Seurat built-in FindAllMarkers function. Gene signature scores were calculated using the AddModuleScore_Ucell function79 and adaptiveness or innateness gene sets36. Visualization was performed via Seurat67.

Pseudobulk differential expression and K-means clustering

Raw counts of antigen-specific samples were aggregated by summing the UMIs from their respective single cells utilizing Seurat’s built-in function AggregateExpression67. Antigen-specific samples with fewer than 20 cells were not considered. In addition, low-expressing genes with counts fewer than 10 within the group were filtered out using the filterByExpr function69. Raw counts were then regularized log (rlog) transformed for computing principal components, followed by batch effect removal using plotPCA and removeBatchEffect function, respectively69,70. Differentially expressed genes (DEGs) for each antigen-specific T cell comparison between older and young individuals, as well as between antigen-specific T cells of young individuals, were calculated. To identify DEGs, the mean-variance trend was modeled on the log-transformed counts per million (logCPM) value of each gene, incorporating the sample weight adjustment via voomWithQualityWeights in limma70. DEGs were calculated by fitting contrasts to the model and employing empirical Bayes moderation with at least a log2 fold change +/– 0.5 and adjusted P-value < 0.05. DEGs were clustered using the k-means function with default parameters. Enrichment analyses were performed using EnrichR73, focusing on Gene Ontologies, Reactome, BioPlanet, KEGG and ChEA gene sets. Gene set enrichment analysis (GSEA) for adaptiveness and innateness genes36, exhaustion-associated genes (GSE9650, GSE41867), and cellular senescence-associated genes (GO:0090398, Reactome R-HSA-2559583, SenSig80, SenMayo81) was performed by FGSEA v1.20.082. UpSet plots were generated with Intervene83 and heatmaps with Morpheus (https://software.broadinstitute.org/morpheus).

Analysis of TetTCR-SeqHD data

We downloaded publicly available high-dimensional tetramer-associated TCR sequences (TetTCR-SeqHD)21. TetTCR-SeqHD data include targeted GEX, ADT, TCR and antigen-specificity against 282 antigens. We extracted EBV-specific CD8+ T cells with specificities against BMLF1, BRLF1, BZLF1, LMP1, LMP2, LMP2A, EBNA3A. We processed data as described for self-collected samples, elaborated on in the preceding section.

T cell receptor analyses

We assessed whether previously identified TCR CDR3 sequences were found in our data by extracting publicly deposited HLA-A*02 TCR CDR3 sequences from VDJdb84: 4157 entries were found for BMLF1 of which 363 were unique TCRα CDR3 and 1182 were unique TCRβ CDR3 regions, as well as 1155 BRLF1 entries (144 unique TCRα CDR3, 394 unique TCRβ CDR3 regions), 51 EBNA3C entries (21 unique TCRα CDR3, 22 unique TCRβ CDR3 regions), 43 LMP1 entries (11 unique TCRα CDR3, 17 unique TCRβ CDR3 regions), and 530 LMP2 entries (41 unique TCRα CDR3, 95 unique TCRβ CDR3 regions). No entries were found for BMRF1, EBNA1, or VZV-IE62. We then calculated the proportion of unique CDR3 regions found in VDJdb that were also found in our data.

To calculate TCR diversity and clonality, we extracted nucleotide sequences of complete TCRα and TCRβ chains (including CDR1, CDR2, and CDR3). We adopted the Inverse Simpson index and Shannon index to calculate TCR diversity, the Gini index was used to calculate TCR clonality. Only conditions with at least 20 cells were included. The Inverse Simpson index (D) was calculated as

$$D={\left(\sum {p}_{i}^{2}\right)}^{-1}$$
(1)

with \({p}_{i}\) being the proportion of the i-th clones. The Shannon index (H) was calculated as

$$H=-\sum \left({p}_{i}{\log }_{2}{p}_{i}\right)$$
(2)

The Gini index (G) was calculated as

$$G=\frac{{\sum }_{i=1}^{n}\left(2i-n-1\right){x}_{i}\,}{n{\sum }_{i=1}^{n}{x}_{i}}$$
(3)

where i is the rank of values in ascending order of clones, n is the number of clones and xi is a clone in the i-th rank. To assess clonal sharing, we calculated the TCR Repertoire Similarity Score (TRSS) between cell types for each antigen across age groups as described before85. We also analyzed TCR V-gene segment usage of TCRα and β chains for each antigen across age groups. These analyses were based on the CellRanger output. In addition, we performed principal component analysis (PCA) on these data to explore antigen- and age-related patterns in V-gene usage. To identified amino acid motifs enriched in TCRβ CDR3 sequences for each antigen and age, we used GLIPH286.

We identified shared TCR sequences that were identical in their amino acid sequence (CDR1, CDR2, and CDR3 regions of both TCRα and β chains) and shared across 2 individuals who were processed in 2 separate batches. Only expanded clones in both individuals with at least 2 cells per clone were considered, resulting in the identification of 14 shared clones. We assessed their phenotype by visualizing these cells on the UMAP projection.

Statistical analyses

For statistical association of cluster proportions with antigen specificity and age, we characterized each sample as a probability vector, whose components represent the relative cell frequency in each of the clusters, and the cardinality represents the number of distinct clusters. To examine the association between the probability vector and the antigen specificity, we employed linear regression analysis between the individual component of the probability vector, i.e., the relative cell frequency in a cluster, and antigen by treating the probability component as outcome and antigen specificity as the independent variable of interest. We adjusted gender in the regression analysis. The association with the entire probability vector is summarized by the sum of squares of all z scores corresponding to the regression coefficient of the antigen specificity from the linear regression analyses and tested via a permutation test, permuting the antigen specificity. A similar method was used to assess the association of age within each antigen specificity with gender adjustment.

To assess correlations between TCR avidity and gene expression scores, we estimated the mean difference in average log-transformed avidity between two different antigens for each individual. Next, we combined the estimated difference across seven young individuals and seven older individuals separately, using a random effects model with an individual-specific random intercept. The combined between-antigen difference in avidity is summarized by a point estimator and a 95% confidence interval. We further compared the combined difference in young individuals with that in older individuals using the Wald test. We calculated the average log-transformed avidity and average gene score for innateness or adaptiveness gene sets corresponding to each antigen for each individual. Next, we calculated the correlation coefficient between average avidity and average gene score across eight antigens. Lastly, we combined the Fisher-transformed correlation coefficients across seven young individuals and seven older individuals, separately, using a random effects model with an individual-specific random intercept. The combined correlation coefficient is summarized by a point estimator and a 95% confidence interval. We also compared the combined correlation coefficient between young and older individuals.

All other statistical analyses were performed via Prism software 10 (GraphPad). Data are presented as median or as mean with error bars indicating standard error of the mean (SEM), unless otherwise indicated. Unpaired, two-tailed Student’s t tests were used when comparing two groups. One-way ANOVA with Dunnett’s multiple comparisons test, or two-way ANOVA with Šídák’s multiple comparisons test, was used for multigroup comparisons. A P-value less than 0.05 was considered statistically significant.

Reporting summary

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