Age-independent and targetable transcription factor networks regulating CD8⁺ T cell senescence in aging humans

SUMMARY

The age-related decline in immunity is accompanied by the accumulation of senescent CD8⁺ T cells. Using senescent cell isolation coupled with multi-omics profiling, we reveal the transition to senescence to be controlled by chromatin state-specific transcription factor (TF) networks in younger and older donors independent of age. These TF networks mediate widespread enhancer remodeling, repressing cell identity genes while upregulating inflammatory and secretory pathways. Inhibition or downregulation of AP1, KLF5, or RUNX2 modulates the transcriptional output and partially restores the blunted response to stimulation of senescent CD8⁺ T cells. Senescent CD8⁺ T cell gene signatures also predict responsiveness to chimeric antigen receptor (CAR)-T cell therapy in diffuse large B cell lymphomas. Overall, our study defines the gene-regulatory mechanisms underlying human CD8⁺ T cell senescence, highlights TF network perturbation as a viable strategy to manipulate the senescence state, and identifies senescent CD8⁺ T cell gene signatures as prognostic tools for immunotherapy outcome.

In brief

Turano et al. reveal the transcription factor networks driving CD8⁺ T cell senescence in healthy aging humans. Inhibiting key transcription factors modulates the senescence program and partially restores responsiveness to TCR stimulation. They also show that CD8⁺ T cell senescence gene signatures predict response to CAR-T cell therapy in B cell lymphomas.

Graphical Abstract

INTRODUCTION

Aging increases susceptibility to infection and disease in older populations. Recent evidence has established that cellular senescence, a dysfunctional cell state characterized by stable proliferation arrest and a pleiotropic secretome (the senescence-associated secretory phenotype [SASP]), is a critical contributor to the aging process.^1,2 Stresses that perturb cell identity activate the senescence program to signal the presence of tissue damage and coordinate subsequent repair.^3,4 Through the SASP, senescent cells recruit innate and adaptive immune cell types for their removal, thereby limiting the deleterious effects caused by the prolonged presence of senescent cells in tissues.^5,6 However, the immune system’s capacity to remove senescent cells from tissues decreases with age. This overall decline in immune function, known as immunosenescence, is defined by impaired thymopoiesis, disrupted CD4⁺-to-CD8⁺ T cell ratios, and chronic low-level inflammation and contributes to the increased morbidity and mortality in older adults.^7,8 Whether the immune system itself undergoes senescence with age remains poorly understood. Defining the molecular mechanisms that drive senescence of the immune system could reveal opportunities to develop therapies to restore immune function in the older population to increase healthspan.

In the CD8⁺ T cell compartment, the age-associated decline in immune function is linked to the expansion of highly differentiated subsets that acquire dysfunctional markers, including PD1 (PDCD1), KLRG1, and CD57^9–13; decreased proliferation capacity; and nonspecific cytotoxic capacity.¹⁴ Using a senescent cell isolation method based on the detection of senescence-associated β-galactosidase (SA-β-gal) activity, we revealed a drastic increase in CD8⁺ T cells with features of senescence, including high levels of the cell cycle inhibitors CDKN1A (p21) and CDKN2A (p16), an inflammatory SASP, and telomeric DNA damage, across all differentiation states and encompassing several described CD8⁺ T cell dysfunction markers in the peripheral blood of older individuals.¹⁵ Despite their proliferation arrest, recent epigenetic studies have shown that senescent cells are dynamic.^16–18 This nature of the senescence state underlies its heterogeneity but also exposes liabilities in senescent cells that can potentially facilitate their reprogramming to a functional identity or promote their elimination.^19,20 Although recent work on the epigenetic regulation of CD8⁺ T cells during aging has established a correlation between dedicated usage of transcription factors (TFs) to mediate differentiation, aging, and dysfunctional state-specific gene expression programs,^21–24 the gene-regulatory mechanisms that mediate the transition to senescence in CD8⁺ T cells during aging in humans are not known.

In this study, we combined senescent cell isolation with multi-omics profiling to define the gene-regulatory mechanisms driving CD8⁺ T cell senescence in a cohort of younger and older donors. This approach identified senescence as the major driver of epigenome and gene expression changes in CD8⁺ T cells, with a minor contribution of age. We establish enhancer remodeling as a key mechanism driving senescence-associated gene expression in CD8⁺ T cells, define and modulate the TF networks controlling the senescence program, reveal the transcriptional basis underlying the refractoriness of senescent CD8⁺ T cells to T cell receptor (TCR) stimulation and identify CD8⁺ T cell senescence gene signatures with diagnostic and prognostic potential for chimeric antigen receptor (CAR)-T cell therapy and systemic lupus erythematosus (SLE).

RESULTS

Transcriptional landscape of senescent CD8⁺ T cells across age and differentiation

We previously showed varying basal levels of SA-β-gal activity across peripheral blood mononuclear cell subtypes.¹⁵ To accurately quantify the frequency of CD8⁺ T cells with high SA-β-gal activity, we calibrated SA-β-gal-low and -high gates within the CD8⁺ T cell lineage using donors that exhibited a clear bimodal SA-β-gal activity distribution, acting as internal controls for all donors (Figure 1A). Flow cytometry profiling of 25 younger (ages 23–34) and 14 older (ages 49–75) donors revealed a marked age-dependent increase in circulating CD8⁺ T cells with high SA-β-gal activity (Figure 1B), in line with our previous study.¹⁵ This was accompanied by an age-associated expansion of differentiated T effector memory cells (T_EM) and T effector memory cells re-expressing CD45RA (T_EMRA) subsets and a decline in T_naive CD8⁺ T cells, as defined by CCR7 and CD45RA expression^25–27 (Figure 1C). Quantification of CD8⁺ T cell differentiation states within the SA-β-gal-low and -high CD8⁺ T cell compartments demonstrated that the SA-β-gal-low population is predominantly, but not exclusively, composed of naive T cells in younger donors, whereas in older individuals, it encompassed more differentiated subsets (Figure 1D). In contrast, the SA-β-gal-high population consisted mainly of T_EM and T_EMRA cells in both age groups but also included a subset of naive T cells (Figures 1C, 1D, and S1A), suggesting that senescence develops across all CD8⁺ T cell differentiation states. To further characterize the distribution of senescent cells within differentiation states, we applied the flow self-organizing map algorithm (FlowSOM)²⁸ to CD8⁺ T cells from both age groups and projected the resulting maps using uniform manifold approximation (UMAP). This approach recapitulated canonical T cell differentiation states and revealed that, although SA-β-gal activity is high in the T_EM and T_EMRA metaclusters, these subsetsalso contained cells with lower SA-β-gal activity (Figures 1E and 1F). Conversely, the T_naive metacluster mainly contained SA-β-gal-low cells but also some with high SA-β-gal activity, corroborating the notion that senescence occurs across all differentiation states (Figures 1E and 1F). To further investigate the biological basis of these observations, we leveraged our previous approach¹⁵ and isolated CD8⁺ T cells with low and high SA-β-gal activity from donors who exhibited a clear bimodal distribution. We collected non-senescent and senescent CD8⁺ T cells from a cohort of 14 healthy donors (7 donors, age 20–27 years; 7 donors, age 55+ years) and performed bulk RNA sequencing. Consistent with our previous results,¹⁵ principal-component analysis (PCA) and hierarchical clustering confirmed that the transcriptome of CD8⁺ T cells from younger and older donors was partitioned based primarily on SA-β-gal content with a small contribution from donor variability and a negligible contribution from age (Figures 1G and 1H). Further characterization of their transcriptomes utilizing a self-organizing map (SOM) algorithm²⁹ revealed generally consistent expression portraits of non-senescent and senescent CD8⁺ T cells, with some donor-specific differences (Figures 1I and S1B). For instance, genes in clusters K and M, which enriched for inflammatory pathways, exhibited consistent expression levels across age and senescence status (Figures S1C–S1E). In contrast, small but detectable age- and senescence status-specific levels were detected for epithelial-to-mesenchymal transition (EMT)- and coagulation-associated genes in clusters A and H (Figures S1C–S1E). Thus, non-senescent and senescent CD8⁺ T cells exhibit coherent expression portraits across donors independent of age. Differential expression analysis identified 4,702 differentially expressed genes (DEGs) and generated two modules specific to non-senescent (blue) and senescent (turquoise) CD8⁺ T cells, irrespective of age (Figures 1J and 1K). Pathway analysis revealed extensive representation of senescence-associated pathways on the SA-β-gal-high-specific turquoise module, including cell cycle regulation, the P53 pathway, inflammatory signaling, and secretory pathways. In contrast, the blue module, which characterizes non-senescent CD8⁺ T cells, was enriched for proliferation-associated MYC targets (Figure 1L). We identified a small number of age-specific DEGs in non-senescent (126) and senescent (62) CD8⁺ T cells whose expression pattern exhibited more pronounced donor-specific differences than senescence-specific DEGs (Figures S1F–S1H). These genes were overrepresented in a small subset of pathways, including EMT, hypoxia, and glycolysis, indicating that age-specific changes to CD8⁺ T cell expression are specialized. In contrast, senescence-specific DEGs exhibited substantial diversity of biological function (Figure S1I). Our analysis of transcriptomes of CD8⁺ T cells from young and older donors highlights the acquisition of a senescence state as the major driver of gene expression changes with age.

Figure 1. Distribution and transcriptional landscape of senescent CD8⁺ T cells.

(A) Representative flow cytometry histogram and gating of CD8⁺ T cell SA-β-gal expression in a younger (blue) and older (green) donor.

(B) Quantification of the percentage of SA-β-gal-high CD8⁺ T cells in younger (n = 25) and older (n = 14) donors; mean ± SD. Unpaired two-tailed Welch’s t test. **p < 0.01).

(C) Flow cytometry dot plot and gating strategy of the indicated CD8⁺ T cell differentiation states.

(D) Distribution (percent, ±SD) of CD8⁺ T cell differentiation states in total CD8⁺ T cells, SA-β-gal-low, and SA-β-gal-high for both younger (n = 25) and older (n = 14) donors.

(E) UMAP of CD8⁺ T cells from a subset of younger (n = 5) and older (n = 8) donors overlayed by age group (left) and by self-organizing map (SOM) metacluster label (right). Also shown are CD45RA and CCR7 expression profiles of the indicated metaclusters (bottom left).

(F) UMAP embedding of CD8⁺ T cells, colored by scaled SA-β-galGal expression.

(G) PCA of transcriptomes (reads on exons) of SA-β-gal-low and SA-β-gal-high CD8⁺ T cells isolated from younger and older individuals.

(H) Correlation heatmap (r) between the CD8⁺ T cell transcriptomes of the indicated donors. The scale bar indicates the Pearson’s correlation coefficient.

(I) Averaged SOMs of transcriptomes of SA-β-gal-low and SA-β-gal-high CD8⁺ T cells from younger and older individuals. The scale bar indicates the log2 fold-change in expression of the SOM metagenes.

(J and K) Individual (J) and averaged (K) heatmaps of color-coded modules of DEGs in the indicated CD8⁺ T cells from younger (Y) and older (O) donors. The scale bar indicates the row Z score of the regularized log-transformed counts.

(L) Functional overrepresentation analysis map showing significant associations of the Molecular Signatures Database (MSigDB) hallmark gene sets for expression modules shown in (J).

SA-β-gal activity levels define the accessible chromatin of CD8⁺ T cells

To gain insights into the epigenetic mechanisms driving CD8⁺ T cell senescence, we performed the assay for transposase-accessible chromatin followed by sequencing (ATAC-seq)³⁰ on sorted SA-β-gal-low and SA-β-gal-high CD8⁺ T cells from 4 younger and 4 older donors. Consistent with the results from our transcriptome analysis, the accessible chromatin of CD8⁺ T cells was primarily defined by SA-β-gal levels, with a small contribution of inter-individual variation and a negligible effect of age, as revealed by PCA of total reads on peaks (Figure 2A). We identified 8,605 differentially accessible regions (DARs), which were partitioned into two distinct modules, I and II, that defined non-senescent and senescent CD8⁺ T cells. The patterns of DARs were generally consistent across individuals, although we noticed a reduction of signal intensity on the averaged DARs of both SA-β-gal-low and SA-β-gal-high CD8⁺ T cells of older donors (Figures 2B and 2C). Annotation and integration of DARs with gene expression retrieved 1,827 DEGs (924 specific to module I, 726 to module II, and 177 in common) linked to the nearest DARs (Figure 2D), accounting for ~39% of the DEGs in senescent CD8⁺ T cells, suggesting the existence of additional regulatory layers to the transcriptional output. Most DARs from both modules localized to intronic and intergenic regions within 40 kb from the transcription start site (TSS) of DEGs (Figures 2E, 2F, S2A, and S2B). Most DEGs were primarily linked to a single DAR (Figures 2G and S2C), indicating that a single regulatory chromatin binding event is sufficient to drive changes to their expression.³¹ The directionality and magnitude of chromatin accessibility changes generally reflected those of their linked DEGs, consistent with previous reports.^16,31 For instance, decreased accessibility of DARs in module II in senescent CD8⁺ T cells correlated with repression of metabolism-related DEGs in SA-β-gal-low CD8⁺ T cells (Figures S2D and S2E). In contrast, most DEGs linked to open DARs in module I in senescent CD8⁺ T cells were highly expressed and represented senescence pathways, including the SASP, and cell cycle control (Figures 2H and 2I). Collectively, our results demonstrate that the acquisition of senescence is the primary driver of chromatin accessibility changes in circulating CD8⁺ T cells.

Figure 2. Senescence, but not age, drives changes in chromatin accessibility of CD8⁺ T cells.

(A) PCA of chromatin accessibility of SA-β-gal-low (blue) and SA-β-gal-high (green) CD8⁺ T cells isolated from four younger and older individuals.

(B and C) Individual (B) and averaged (C) heatmaps of modules of DARs in SA-β-gal-low and SA-β-gal-high CD8⁺ T cells from younger and older donors. The scale bar indicates the row Z score of the regularized log-transformed counts.

(D) Intersections of senescence-specific DEGs linked to each DAR module identified in (B).

(E) Upset plot showing the frequency distribution of opening DARs (module I) with the indicated genomic features.

(F) Frequency distribution showing the occurrence of opening DARs relative to the TSS of the nearest annotated gene.

(G) Histogram plot showing the number of DEGs linked to opening DARs.

(H) Correlation between chromatin accessibility and expression of DEGs linked to opening DARs. Individual DEGs from the blue and turquoise modules from(C) are shown as color-coded circles.

(I) Functional overrepresentation analysis map showing significant associations of the MSigDB hallmark gene sets for the indicated DEGs linked to opening DARs (as in Figure 1L).

Enhancer remodeling dictates the transcriptional output of senescent CD8⁺ T cells

Previous studies have shown a critical role of enhancers in the timely execution of senescence programs.^17,32,33 To characterize enhancer dynamics of SA-β-gal-low and SA-β-gal-high CD8⁺ T cells, we performed cleavage under targets & tagmentation (CUT&Tag)³⁴ to map poised enhancers (monomethylated lysine 4 of histone H3; H3K4me1), Polycomb-repressed chromatin (trimethylated lysine 27 of histone H3; H3K27me3), and active chromatin (acetylated lysine 27 of histone H3; H3K27ac) genome wide. We then combined histone modification and chromatin accessibility datasets and performed a chromatin state transition analysis,³⁵ modeling the acquisition of senescence as a transition from the SA-β-gal-low to the SA-β-gal-high state. Consistent with our published results,^16,36 the mappable genomes of SA-β-gal-low and SA-β-gal-high CD8⁺ T cells were largely (<80%) devoid of the epigenetic features profiled, with only ~16% of the genome attributed to these (Figure S3A). Intriguingly, and in contrast to our previous results on fibroblasts,^16,17 senescent CD8⁺ T cells were characterized by inactivation of enhancers to the unmarked, poised, and Polycomb-repressed states (Figures 3A, 3B, and S3B). Accordingly, enhancer inactivation was linked to the reduced transcriptional output of nearby genes and was significantly correlated with the repression of SA-β-gal-low-specific DEGs. In contrast, enhancer activation robustly correlated with expression of senescence-specific DEGs in SA-β-gal-high CD8⁺ T cells (Figures 3C and S3C). For instance, induction of the pro-inflammatory mediator TBKBP1³⁷ and TBX21, a TF involved in T cell exhaustion³⁸ and senescence (see below), is linked to enhancer activation from the poised and Polycomb-repressed states, respectively (Figure 3D). In contrast, repression of LAYN, a mediator of antitumor immunity,³⁹ is associated with H3K27me3 deposition across its entire gene body (Figure S3D) and highlights the possibility of reduced antitumor surveillance capacity of senescent CD8⁺ T cells.

Figure 3. Enhancer remodeling drives the transition to senescence in CD8⁺ T cells.

(A) Arc plot summarizing the evolution of the most prominent chromatin states in senescent CD8⁺ T cells. The edge width is proportional to the number of 200 bp bins undergoing the indicated chromatin state transition.

(B) Genome-wide signal intensity heatmaps of H3K4me1, H3K27ac, ATAC-seq, and H3K27me3 at active enhancers (top) and unmarked chromatin (bottom) in SA-β-gal-low and SA-β-gal-high CD8⁺ T cells. The scale bars indicate the normalized signal.

(C) Dot plot of the correspondence analysis linking chromatin state transitions to DEG modules from Figure 1J. Significance of the association was calculated using a chi-square test.

(D) Genome browser snapshot of the TBKBP1 and TBX21 loci, showing signal tracks for RNA-seq, ATAC-seq, H3K4me1, H3K27ac, and H3K27me3 in SA-β-gal-low and SA-β-gal-high CD8⁺ T cells.

(A–D) CUT&Tag data were generated from SA-β-gal-low and SA-β-gal-high CD8⁺ T cells from three younger donors.

(E–G) Averaged heatmaps of color-coded modules (E and G) and PCA projection (F) of DEGs identified for the indicated comparisons (top of heatmaps and PCA plot). The scale bar (E and G) indicates the row Z score of the regularized log-transformed counts.

(H) Functional overrepresentation analysis map showing significant associations of the MSigDB hallmark gene sets for the indicated DEG modules from (G).

(I) Upset plot showing the unique and shared DARs for the indicated CD8⁺ T cell states.

(J and K) As in (E) and (G) for the identified DARs of the indicated comparisons. Data for naive and stimulated CD8⁺ T cells are from GEO: GSE212347.

Although the number of actively proliferating CD8⁺ T cells in healthy human donors is estimated to be low,⁴⁰ the gene-regulatory landscapes of SA-β-gal-low and SA-β-gal-high CD8⁺ T cells could reflect their proliferation status after stimulation. To address this possibility, we compared gene expression, chromatin accessibility, and H3K27 acetylation between SA-β-gal-low and SA-β-gal-high CD8⁺ T cells, along with previously characterized naive CD8⁺ T cells (defined as TIM3/PD1-negative) with and without anti-CD3/anti-CD28 stimulation for 72 h.⁴¹ Clustering and PCA of DEGs relative to SA-β-gal-low CD8⁺ T cells (n = 9,811 DEGs) revealed that TCR stimulation accounted for most of the variability of gene expression. Consistent with our quantification of CD8⁺ T cell differentiation states across SA-β-gal status (Figures 1C–1E), the transcriptome of SA-β-gal-low CD8⁺ T cells was closely associated with that of unstimulated naive CD8⁺ T cells, irrespective of donor age. In contrast, SA-β-gal-high CD8⁺ T cells shared similar expression patterns with stimulated CD8⁺ T cells (see blue and turquoise modules in Figure 3E), explaining their proximity in the PCA projection (Figure 3F). Although these results suggest that senescent and activated CD8⁺ T cells are transcriptionally similar, a direct comparison of their transcriptomes revealed 6,715 DEGs with largely opposite expression patterns. While stimulated CD8⁺ T cells upregulated expression of proliferation-associated genes (blue and black modules, Figures 3G, S3H, and S3I), senescent CD8⁺ T cells expressed genes including CD16, a cell surface receptor found on senescent-like CD8 T cells,⁴² and those involved in secretory pathways (turquoise module, Figures 3G, S3H, and S3I). These analyses demonstrate that senescent CD8⁺ T cells represent a distinct cell state. Differential accessibility and enhancer activation analyses, as measured by H3K27 acetylation, corroborated this conclusion, with TCR stimulation generating the highest number of de novo accessible (~3.5 × 10⁴) and K27ac (~2.1 × 10⁴) peaks, followed by SA-β-gal-high CD8⁺ T cells (~2.4 × 10⁴ and ~1.3 × 10⁴, respectively) (Figures 3I and S3E). Clustering of cell state-specific DARs (n = 45,736) and differential K27ac peaks (n = 27,032) confirmed the similarities between naive and SA-β-gal-low CD8⁺ T cells as well as between activated and SA-β-gal-high CD8⁺ T cells (Figures 3J and S3F). However, a direct comparison of SA-β-gal-high and stimulated CD8⁺ T cell chromatin accessibility and enhancer activation datasets identified 31,482 DARs and 17,491 differential K27ac peaks, respectively, with largely cell state-specific signals patterns (Figures 3K and S3G), Collectively, our comprehensive chromatin and comparative cell state analyses reveal local enhancer remodeling as a critical epigenetic mechanism driving the senescence program and provide compelling evidence that CD8⁺ T cell senescence is a unique cell state.

Local activation of the TF network underpins the transition to senescence in CD8⁺ T cells

TF networks orchestrate cell fate transitions by controlling the timing and magnitude of the gene expression output.^43,44 To define the function and dynamics of the TF networks that mediate the senescence program in CD8⁺ T cells, we searched for TFs exhibiting differential expression between SA-β-gal-low and SA-β-gal-high CD8⁺ T cells from younger and older donors. We identified 455 differentially regulated TFs, representing ~40% of all detectable TFs in CD8⁺ T cells, which were partitioned based on SA-β-gal level (Figure 4A). The high number of differentially regulated TFs suggested a large rearrangement to their gene regulatory network as these cells transitioned to senescence. To address this possibility, we quantified TF binding instances on DAR modules (Figure 2) of SA-β-gal-low and SA-β-gal-high CD8⁺ T cells of younger and older donors. The KLF family of TFs emerged as the most pervasively bound TFs to closing and opening chromatin in both age groups, irrespective of senescence status, suggesting a pioneering role of KLF TFs in CD8⁺ T cells⁴⁵ (Figures 4B–4E and S4A–S4D). Interestingly, members of the AP1 family of TFs were prebound to senescence-specific open chromatin irrespective of senescence status, indicating that, along KLFs and additional TFs, AP1 TFs act as bookmarking agents for the timely execution of the senescence program¹⁷ in CD8⁺ T cells. We subsequently integrated differential TF binding activity at opening and closing chromatin with their respective transcriptional dynamics, which revealed a highly dynamic TF network that relied on both constitutively and differentially expressed TFs (Figures 4F and S4E). The binding activity of AP1 (FOSL2 and FOS) TFs, as well as TBX1 and EOMES, was highest in the open chromatin of SA-β-gal-high CD8⁺ T cells (Figure 4F, top insets), while the binding of proliferation-associated TFs, E2F6 and MYC, was reduced (Figure 4F, bottom insets). Similarly, consistently higher activity of KLFs and SP and REL TFs was observed at closing chromatin of senescent CD8⁺ T cells (Figure S4E, bottom insets). Mapping of the TF co-binding interactions revealed specific TF networks at opening and closing chromatin of senescent CD8⁺ T cells. AP1, KLF, and SP TFs mediated interactions with most other TFs at the open chromatin of SA-β-gal-high CD8⁺ T cells in both younger and older donors (Figures 4G and 4H and see below), while KLF TFs dominated the TF interactome at closing chromatin (Figures S4F and S4G). To visualize the structure and dynamicity of the TF interactome, we constructed TF network models by leveraging the co-binding maps and expression status of TFs at opening and closing chromatin. The CD8⁺ T cell TF networks were composed of both constitutively and differentially expressed TFs and displayed a hierarchical structure composed of top, core, and bottom layers. As the co-binding interactions flowed from top to bottom, there was an increasing number of nodes with increasing dynamicity, which bound progressively fewer regions (Figures 5 and S5). Interestingly, while the structure of the TF networks in CD8⁺ T cells was remarkably conserved in younger and older donors, these networks displayed unique compositions based on their association with opening or closing chromatin. The opening chromatin TF network exhibited a rather complex top layer composed of 15 nodes, with the AP1, RUNX, ETS, and TBX TFs occupying most regions and connecting to most nodes in the core and bottom layers (Figure 5), while KLF TFs controlled most interactions with the remaining nodes in the network at closing chromatin (Figure S5). Interestingly, a node composed of KLF, SP, AP1, and TCF played a critical role in open chromatin by acting as a relay point between the top and bottom layers (Figure 5), highlighting the versatility of TF node usage in different chromatin contexts.^46,47 Although these networks display some age-specific differences in TF usage and node composition (Figure 4F, middle inset; Figures 5 and S5, compare core and bottom layers), our comprehensive analyses of TF binding demonstrate that CD8⁺ T cells of both younger and older individuals rely on the activation of chromatin state-specific TF networks of remarkably similar composition to implement senescence-associated gene expression.

Figure 4. TF network dynamics of senescent CD8⁺ T cells.

(A) Averaged expression heatmaps of differentially expressed TFs in the corresponding DEG modules of SA-β-gal-low and SA-β-gal-high CD8⁺ T cells. The scale bar indicates the row Z score of the regularized log-transformed counts.

(B–E) Rank plots showing the summed binding instances per TF in the senescence-specific open chromatin of SA-β-gal-low (B and C) and SA-β-gal-high (D and E) of CD8⁺ T cells.

(F) Heatmap of differential TF binding activity at senescence-specific opening chromatin of SA-β-gal-low and SA-β-gal-high CD8⁺ T cells. The left annotation heatmaps show the number of bound instances per TF (cyan to red) and their gene expression (TXN) category. Insets: chromatin binding activity of age- and cell state-specific TFs. The scale bar indicates the row Z score of the activity of each TF motif in the dataset.

(G and H) TF co-binding interaction matrices in the senescence-specific open chromatin of CD8⁺ T cells from younger (G) and older (H) donors. The corresponding q values of the interactions are projected onto the clustering and represented in a color scale defined by their significance using a hypergeometric distribution test.

Figure 5. Distinct hierarchical TF networks control the senescence program of CD8⁺ T cells.

Shown are TF networks for the senescence-specific open chromatin module in CD8⁺ T cells from younger (A) and older donors (B). TFs (nodes) are represented as circles and oriented edges as arrows. Nodes represent strongly connected components to facilitate visualization. The fill color of the node’s inner circle is based on the normalized dynamicity (prestimulation) of TFs. The fill color of the outer ring indicates whether the TF is constitutively expressed (black) or belongs to a transcriptomic module as described in Figure 1D. The node’s size is proportional to the bound regions by a given TF(s). Each network has three layers: (1) the top layer with no incoming edges, (2) the core layer with incoming and outgoing edges, and (3) the bottom layer with no outgoing edges. The top-layer most prominent TFs and the highly connected KLF/SP SCC are labeled. The scale bars in the inset denote the prestimulation and directional overlap of the TF interactions. Bold labels indicate TFs targeted with pharmacological inhibitors (Figures 6 and S6).

Rational targeting of the TF network to modulate senescence-associated gene expression in CD8⁺ T cells

The hierarchical structure of TF networks provides an opportunity for their targeted perturbation to manipulate cell states.¹⁶ Using this logic, we purified SA-β-gal-low and -high CD8⁺ T cells, subsequently cultured them in the presence of interleukin-2 (IL-2) for 3 days, and determined their transcriptomic profiles. We identified 4,146 DEGs, which split into two modules (gray and orange) in an SA-βGal level-specific manner, which enriched for MYC targets and multiple senescence-associated pathways, respectively (Figures S6A and S6B), consistent with results from freshly purified SA-β-gal-low and -high CD8⁺ T cells (Figure 1). Interestingly, IL-2 treatment induced 1,742 specific DEGs and shared a core of 2,406 DEGs with freshly purified SA-β-gal-low and -high CD8⁺ T cells, which exhibited 2,296 specific DEGs, indicating a differential response to IL-2 stimulation by non-senescent and senescent CD8⁺ T cells (Figure S6C). Having established this baseline, we perturbed three critical nodes—KLF5, AP1 (FOS), and RUNX2—in the TF networks of both opening and closing chromatin (Figures 5 and S5) using the pharmacological inhibitors SR18662, T5224, and CADD522,^48–50 respectively, in SA-β-gal-high CD8⁺ T cells. Quantification of the number of binding instances per TF and annotation to their nearest DEGs in both opening and closing chromatin revealed unique and shared binding instances and DEGs associated with each TF, indicating that these TFs act both individually and as a network (Figures 6A, 6B, S6D, and S6E). Analysis of bulk transcriptomes of control and inhibitor-treated SA-β-gal-high CD8⁺ T cells confirmed individual and combinatorial gene regulation by the targeted TFs. For example, genes involved in the interferon γ (IFNγ) response in module 1 linked to open chromatin were repressed in cells treated with either SR18662 or CADD522 but only modestly affected by T5224, indicating that these genes are co-regulated primarily by KLF5 and RUNX2 (Figures 6C and 6D). Similar results were found for DEGs linked to closing chromatin (Figures S6F and S6G). In addition, TF inhibitor treatment resulted in both up- and downregulation of target DEG modules, confirming that TFs act to fine-tune the level of expression of their target genes^17,51 (Figures 6C and S6F). This conclusion is supported by an information flow analysis of TF binding and gene expression modules, which showed that only a fraction of binding sites per TF was involved in regulating the perturbed DEG modules (Figures 6E and S6H). To ensure the specific downregulation of these TFs, we knocked down expression of the AP1 subunit JUN, as well as KLF5 and RUNX2 using two distinct small interfering RNA (siRNAs), each, in senescent SA-β-gal-high CD8⁺ T cells and performed RNA sequencing (RNA-seq). After validating the senescent transcriptome in control, non-targeting siRNA (siC)-treated SA-β-gal-high CD8⁺ T cells (Figures S6I–S6K) and efficient downregulation of the targeted TFs by RT-qPCR (Figure 6F), we identified their DEG targets by leveraging the AP1, KLF5, and RUNX2 footprints from opening and closing chromatin of SA-β-gal-high CD8⁺ T cells (Figures 4 and 5). Consistent with the results from inhibitor-treated SA-β-gal-high CD8⁺ T cells, we found unique and shared target DEGsper TF linked to opening and closing chromatin (Figures 6G and S6L) whose expression profiles were modulated by the siRNA-mediated downregulation of the targeted TFs, corroborating the fine-tuning function of TFs in the expression output of their target genes. For instance, the expression of opening chromatin-associated DEGs in module ii involved in complement and allograft rejection was enhanced by treatment with JUN-targeting siRNAs, whereas the downregulation of both KLF5 and RUNX2 led to their repression (Figures 6H and 6I). Similarly, closing chromatin-linked DEGs in module b involved in apoptosis were derepressed to varying degrees following knockdown of JUN, KLF5, or RUNX2 (Figures S6M and S6N). Taken together, results from pharmacological and RNAi approaches highlight that perturbation of the TF networks is a feasible strategy for manipulating CD8⁺ T cell senescence.

Figure 6. Perturbation of the senescence CD8⁺ T cell TF network modulates the output of the senescence program.

(A) Upset plot showing the individual and co-binding instances of KLF5, AP1, and RUNX2 in the senescence-specific open chromatin of CD8⁺ T cells cultured in the presence of 20 U/ml IL-2.

(B) Intersections of gene targets for KLF5, AP1, and RUNX2 in open chromatin of senescent CD8⁺ T cells.

(C) Heatmap showing the effect of KLF5, AP1, and RUNX2 inhibitors (SR18662, T5224, and CADD522, respectively [30 μM]) on average expression (n = 3 donors) of the direct, senescence-specific DEG TF targets linked to open chromatin in SA-β-gal-high CD8⁺ T cells. The scale bar indicates the row Z score of the regularized log-transformed counts.

(D) Functional overrepresentation analysis showing significant associations of the MSigDB hallmark gene sets for each expression module shown in (C).

(E) River plots showing the proportion of KLF5, AP1, and RUNX2 bound sites linked to the expression of their respective gene targets in open chromatin of senescent CD8⁺ T cells.

(F) RT-qPCR of JUN, KLF5, and RUNX2 using two independent siRNAs per TF relative to a non-targeting siRNA (siC) in CD8⁺ T cells from two donors; mean plus range.

(G) As in (B) but in SA-β-gal-high CD8⁺ T cells targeted with JUN (AP1), KLF5, and RUNX2 siRNAs.

(H) Heatmap showing the effect of siRNA-mediated knockdown of JUN, KLF1, and RUNX2 on average expression (n = 3 donors) of the direct, senescence-specific DEG TF targets linked to open chromatin in SA-β-gal-high CD8⁺ T cells. The scale bar indicates the row Z score of the regularized log-transformed counts.

(I) As in (D) but for DEG modules shown in (H).

CD8⁺ T cell senescence is a robust barrier to TCR signaling-mediated proliferation and predicts the response to CAR-T cell therapy and active SLE

We have shown that SA-β-gal-high CD8⁺ T cells are present across all CD8⁺ T cell differentiation states (Figures 1A–1F and S1A) and exhibit features of senescence, including dysfunctional telomeres, elevated levels of p16, and proliferation defects.¹⁵ These results, combined with the epigenomic analyses presented here, suggest that SA-β-gal-high CD8⁺ T cells have transitioned into a dysfunctional state, which could have a substantial impact on the response to therapeutic treatments and disease progression. To test this possibility, we evaluated the effect of senescence burden on CD8⁺ T cell expansion under conditions used to generate CAR-T cell products by long-term culturing of bulk CD8⁺ T cells with low (1.93%–28.8%, n = 11 donors) and high (35.6%–81.8%, n = 7 donors) senescence content. As expected, CD8⁺ T cells with higher initial SA-β-gal activity plateaued on day ~30 after undergoing 8.8 ± 3.9 population doublings, whereas CD8⁺ T cells with low senescence content continued to proliferate until day ~50, achieving 11.36 ± 5 population doublings (Figure 7A). Reasoning that targeting critical nodes of the senescence TF network of CD8⁺ T cells could restore their proliferation capacity under conditions of TCR stimulation, we isolated SA-β-gal-high CD8⁺ T cells, electroporated them to deliver control and TF-targeting siRNAs (JUN, KLF5, and RUNX2), stimulated them with soluble anti-CD3 and anti-CD28 antibodies for 7 days, and monitored their proliferation relative to siC-treated SA-β-gal-low CD8⁺ T cells. As expected, ~30% of siC-treated SA-β-gal-low CD8⁺ T cells proliferated over the 7-day period, while siC-treated SA-β-gal-high CD8⁺ T cells remained essentially arrested (Figures 7B, top two lanes, and 7C). While knockdown of JUN and KLF5 did not rescue the proliferation defects of SA-β-gal-high CD8⁺ T cells, downregulation of RUNX2 resulted in a modest, albeit not significant, increase in proliferation (~1%; Figures 7B and 7C). To identify the potential roadblocks preventing the restoration of proliferation capacity in TF network node-targeted SA-β-gal-high CD8⁺ T cells, we determined the transcriptomes of siC- and siTF-treated SA-β-gal-high CD8⁺ T cells with and without TCR stimulation for 16 h and compared them to siC-treated SA-β-gal-low CD8⁺ cells. We identified 8,214 TCR stimulation-specific DEGs, which were partitioned into 5 modules (V–Z). In siC-treated SA-β-gal-low CD8⁺ T cells, TCR stimulation resulted in robust expression of genes involved in cell cycle progression, IL-2, signaling, and metabolism (Figures 7D and 7E, module X). Interestingly, while siC-treated SA-β-gal-high CD8⁺ T cells were also capable of upregulating these sets of genes, they did so to a substantially lower magnitude (Figure 7D, inset heatmap). Furthermore, SA-β-gal-high CD8⁺ T cells constitutively expressed SASP- and P53 pathway-associated genes, irrespective of stimulation (Figures 7D and 7E, modules V and Z). While knockdown of JUN resulted in mild upregulation of stimulation-specific proliferation-associated genes as well as of derepressed genes characteristic of unstimulated cells, knockdown of KLF5 generated a gene expression profile that was more closely associated with that of unstimulated CD8⁺ T cells (Figures 7D and 7E, module W). In contrast, downregulating RUNX2 resulted in enhanced expression of proliferation genes and repression of senescence-associated genes (Figures 7D and 7E). These results provide mechanistic evidence of the mild restoration of the proliferation capacity in RUNX2-downregulated SA-β-gal-high CD8⁺ T cells and highlight RUNX2 as a potential target to restore CD8⁺ T cell function. Having established a strong foundation underlying the refractoriness of SA-β-gal-high CD8⁺ T cells to TCR stimulation, we reasoned that the information within their transcriptome could predict therapeutic response. To this end, we generated four CD8⁺ T cell senescence gene signatures from our transcriptomic datasets (overall, TFs, predicted surface genes, and predicted secreted genes) and tested their ability to predict response to CAR-T cell therapy using publicly available, clinically annotated transcriptomic data of naive CD8⁺ and CAR-T cell products of a cohort of patients with diffuse large B cell lymphoma (DLBCL).⁵² We found that all four CD8⁺ T cell senescence gene signatures faithfully identified DLBCL that was insensitive to CAR-T cell therapy (Figures 7F–7H), irrespective of the cell source of the gene expression data. These results highlight senescence profiling of the starting cellular material used in the generation of CAR-T cell products as a potentially critical decision-making step during the management of DLBCL. We also tested CD8⁺ T cell senescence gene signatures as prognostic markers for SLE, an autoimmune disease characterized by autoreactive B cells and aberrant CD8⁺ T cell function,⁵³ using published annotated transcriptomic datasets of a cohort of 48 patients.⁵⁴ We found that 3 of 4 of our CD8⁺ T cell senescence gene signatures were significantly enriched in the transcriptomes of peripheral CD8⁺ T cells of patients with active SLE (Figure S7A) and, to a lesser degree, in patients with inactive SLE (Figure S7B), indicating a potential role of CD8⁺ T cell senescence in SLE pathology. In summary, our integrative analysis of CD8⁺ T cell senescence highlights the potential of senescence-associated gene signatures as prognostic tools in clinical settings, with a strong contribution of CD8⁺ T cells to disease progression and/or response to therapies.

Figure 7. CD8⁺ T cell senescence is a barrier to TCR-mediated proliferation and predicts the response to CAR-T cell therapy in DLBCL.

(A) Proliferation curves of CD8⁺ T cells isolated from peripheral blood of 19 donors and cultured in the presence of IL-2 with periodic stimulation using soluble anti-CD3/28. Non-linear fit comparison revealed statistical significance between groups (p < 0.0001; n = 11 SA-β-gal-low donors; n = 7 SA-β-gal-high donors).

(B and C) Flow cytometry profiles (B) and corresponding quantification (C; mean ± SD) of Cell Trace Violet (CTV) dilution assays in SA-β-gal-low and SA-β-gal-high CD8⁺ T cells electroporated with the indicated siRNAs, from three donors.

(D) Heatmaps showing stimulation-specific DEG modules in unstimulated and stimulated (anti-CD3/anti-CD28 Dynabeads for 16 h) SA-β-gal-low and SA-β-gal-high CD8⁺ T cells electroporated with the indicated siRNAs. Heatmaps show the average of three donors. The scale bar indicates the row Z score of the regularized log-transformed counts.

(E) Functional overrepresentation analysis map showing significant associations of the MSigDB hallmark gene sets for each DEG module shown in (D).

(F–H) Gene set enrichment analyses (GSEAs) showing normalized enrichment score (NES) plots and false discovery rate (FDR) values for overall senescence (SenSig), senescence-specific TF (TF SenSig), surface-predicted (Surf SenSig), and secreted-predicted (Sec SenSig) gene signatures in transcriptomes of naive CD8⁺ T cells (B), CAR-T cell products (C), and combined transcriptomes of (D) patients with DLBLC (GSE223655) profiled at diagnosis and classified into relapse and no-relapse groups.

DISCUSSION

The age-associated accumulation of senescent cells in tissues is one of the driving causes of aging and age-related disease.^55,56 Although senescent cells upregulate the expression of MICA/B ligands and secrete chemokines that facilitate the recruitment of cytotoxic immune cells for their elimination, senescent cells accumulate systemically with advancing age,^57–59 suggesting that immune cell-mediated elimination of senescent cells becomes impaired with age. Recent evidence highlights a dysfunctional adaptive immune system as a potential cause for the age-associated accumulation in senescent cells.^55,60,61 In older humans, both CD4⁺ and CD8⁺ T cells acquire features of senescence,^15,62,63 which is linked to defective immune responses.⁶⁴ However, the gene regulatory mechanisms that promote senescence of CD8⁺ T cells in aging humans, as well as the contribution of senescent CD8⁺ T cells to disease, remain poorly understood. Knowledge of the epigenetic mechanisms driving senescence of CD8⁺ T cells is likely to unveil opportunities for the development of therapies to restore function of damaged CD8⁺ T cells, potentially increasing healthspan and improving the efficacy of immune cell-based therapies. Using senescent cell isolation, integrative multi-omics profiling, and targeted perturbations, we defined and validated the TF networks that control the senescence program in CD8⁺ T cells of a cohort of younger and older donors. Our study therefore lays the foundation for understanding the epigenetic regulation of senescence in CD8⁺ T cells.

One key finding of our study is that the acquisition of senescence is the main driver of epigenomic and gene expression dynamics of CD8⁺ T cells, with a minor contribution of chronological age. These epigenetic and transcriptional features are distinct from those of naive and stimulated CD8⁺ T cells,⁴¹ establishing CD8⁺ T cell senescence as a bona fide cell state. The transition to the senescence state is a major event in the epigenetic life of CD8⁺ T cells, as it involves the differential expression of 40% of all detectable TFs and requires extensive enhancer remodeling for the implementation of the senescence transcriptional program. These findings are in line with previous work demonstrating the dynamic regulation of the enhancer landscape as a general gene-regulatory mechanism required for establishment of the senescence state in human fibroblasts.^17,65 However, an important characteristic of CD8⁺ T cell senescence is the decommissioning of active enhancers as the most prominent chromatin regulatory event, which is linked to the repression of genes characteristic of CD8⁺ T cell function, including immune surveillance and proliferation. In contrast, widespread enhancer activation is required for the establishment of replicative senescence in cultured human fibroblasts.^17,65 Thus, although remodeling of the enhancer landscape appears to be a conserved regulatory event during the implementation of the senescence program, defining the chromatin-based mechanisms underlying cell type-specific differences in its execution will be required to manipulate senescence in a cell type-specific manner.

Our data also reveal that the senescence competence of CD8⁺ T cells is precoded at least during early adulthood, as shown by the detection of prebound, chromatin state-specific, and hierarchical TF networks of remarkably similar architecture in the CD8⁺ T cells of both younger and older individuals. The prior association of these TFs with accessible chromatin likely facilitates rapid activation of the senescence program, independent of differentiation state. Accordingly, we observed near-identical activation of the senescence TF network in both age groups, with a few additional TF interactions in the CD8⁺ T cells of older donors, likely representing more differentiated CD8⁺ T cells characteristic of this age group. Our results suggest that the ability to undergo senescence is a property of all circulating CD8⁺ T cells, irrespective of their differentiation state. This mechanism is poised to maximize the regulatory flexibility of potentially long-lived CD8⁺ T cells⁶⁶ to undergo senescence when required. In line with this interpretation is our quantification of SA-β-gal-positive CD8⁺ T cells across differentiation states in younger and older donors¹⁵ as well as recent evidence characterizing TF activity across CD8⁺ T cell differentiation states in aging humans.^21,23 Our comprehensive analysis of epigenomic data confirmed the hierarchical architecture of the CD8⁺ T cell senescence TF networks, consistent with our previous work,^16,36 and provides a logical platform for the manipulation of senescence in CD8⁺ T cells. Accordingly, we precisely modulated the transcriptome of senescent CD8⁺ T cells by targeting the key nodes c-JUN (AP1), RUNX2, and KLF5 through pharmacological and RNAi approaches. Collectively, these findings provide proof of principle that targeting the senescence TF network is a viable approach to diminish the potentially damaging effects of senescent CD8⁺ T cells.

Another important finding of our study is the robust barrier to proliferation stimuli that is present in senescent CD8⁺ T cells, with potentially important implications for the effectiveness of CAR-T cell therapy and the development of autoimmune disease. We identified a blunted transcriptional response to TCR stimulation in senescent CD8⁺ T cells, which likely underlies the inability of these cells to proliferate. Consistent with this, we observed robust enrichment of CD8⁺ T cell senescence gene signatures derived from our datasets in the starting cellular material (naive CD8⁺ T cells) and the CAR-T cell product of a cohort of patients with DLBLC who were insensitive to CAR-T cell therapy. This, in combination with our data showing the decreased proliferation potential of bulk CD8⁺ T cells with a higher senescence burden under conditions of CAR-T cell product expansion, highlights senescence as a critical block to the development of effective CAR-T cell products. Similarly, our CD8⁺ T cell senescence gene signatures were highly enriched in the peripheral blood of individuals with active and inactive SLE. Although the role of senescent CD8⁺ T cells in SLE pathogenesis is presently unknown, their highly inflammatory nature, along with the reduced expression of functional genes, is consistent with the presence of previously described overreactive inflammatory CD8⁺ T cell subpopulations in patients with active SLE.⁶⁷ Future work will be required to establish the molecular mechanisms leading to CD8⁺ T cell senescence in the context of autoimmunity and CAR-T cell therapy effectiveness.

In conclusion, our study provides a comprehensive foundation of the gene-regulatory mechanisms governing senescence in CD8⁺ T cells and provides a logical platform for its modulation through rational targeting of the senescence TF network. Our study also highlights senescent cell isolation followed by multi-omics profiling as a powerful strategy to study the contribution of immune cell senescence to aging and age-related disease.

Limitations of the study

Our study focused on defining the gene-regulatory mechanisms regulating senescence in circulating CD8⁺ T cells by leveraging the increased activity of SA-β-gal in a subset of these cells. Although we provide compelling flow cytometry evidence showing that senescence occurs across all CD8⁺ T cell differentiation states, our transcriptional, epigenomic, and TF network perturbation studies were performed in bulk SA-β-gal-high CD8⁺ T cells. Therefore, we cannot formally exclude the contribution of CD8⁺ T cell differentiation states to the gene-regulatory networks. Similarly, although we provide a proof of principle that perturbation of the senescence TF network modulates the transcriptional output of SA-β-gal-high CD8⁺ T cells, we were unable to restore their proliferation capacity. Although this highlights the robustness of the senescence state of CD8⁺ T cells, systematic work on the stable modulation of critical TF network nodes such as AP1, KLF5, and RUNX2, but also of additional TF nodes, will be required to fully reprogram senescent CD8⁺ T cells into a cell state with restored immune functionalities.

RESOURCE AVAILABILITY

Lead contact

Requests for further information, resources, and reagents should be directed to and will be fulfilled by the lead contact, Ricardo Iván Martínez-Zamudio (ricardo.martinezzamudio@rutgers.edu).

Materials availability

This study did not generate new unique reagents. All unique/stable reagents generated in this study are available from the lead contact without restriction.

Data and code availability

• The RNA-seq, ATAC-seq, and CUT&Tag data generated in this study are hosted on the Gene Expression Omnibus (GEO: GSE310382, GSE310383, GSE310384, GSE310393, GSE310625, and GSE310729) and are publicly available as of the date of publication. We also analyzed previously published RNA-seq data (GEO: GSE169009).

• All code has been deposited at Zenodo (https://doi.org/10.5281/zenodo.17662132) and is publicly available as of the date of publication.

• Any additional information required to reanalyze the data reported in this paper is available from the lead contact upon request.

STAR★METHODS

EXPERIMENTAL MODEL AND STUDY PARTICIPANT DETAILS

The Institutional Review Board of New Jersey Medical School approved this study (ID Pro2020000006, approved May 14, 2024). Peripheral blood mononuclear cells (PBMCs) were obtained from the blood of consenting healthy donors (aged 20–75 years) collected in heparinized tubes (BD Biosciences, Franklin Lakes, NJ) or Leukopacks were commercially obtained from the New York blood center. Samples were assigned into two groups ‘young’ and ‘old’ based on the age range described in Table 1, as deficits in immune function start manifesting around the sixth decade of life.⁸³ Details of human subjects analyzed are listed in the following table.

Table 1.

Donor information

Donors	Younger	Older
n	31	21
Age range	23–32	49–75
Median age	25.5	64
Male, %	51.61290323	47.61904762
Female, %	48.38709677	52.38095238

METHOD DETAILS

Peripheral blood mononuclear cell (PBMC) isolation and CD8⁺ T cell enrichment

PBMCs were isolated using density gradient centrifugation with lymphocyte separation medium (Corning, Manassas, VA). Briefly, heparinized blood was diluted 1:1 with Hanks’ Balanced Salt Solution (HBSS), (Corning, Manassas, VA), layered over lymphocyte separation medium at a 2:1 ratio, and centrifuged at 400g for 30 min with reduced deceleration (deceleration factor of 2). The buffy coat containing PBMCs was carefully harvested and processed further. CD8⁺ T cells were enriched from PBMCs using negative selection with the EasySep Human CD8⁺ T cell Isolation Kit (StemCell Technologies, Cambridge, MA), following the manufacturer’s protocol. The purity of the isolated CD8⁺ T cells was assessed via flow cytometry by staining for surface markers PerCP/Cyanine7-CD3 (HIT3a), and PE-CD8 (RPA-T8), BioLegend, San Diego, CA. Briefly, all surface staining was conducted at 4C for 20 min in 1xPBS+ 2% FCS whilst protected from light. Samples were subsequently washed 1xPBS+ 2% FCS and kept on ice before being acquired on the BD LSRFortessa X-20 (BD Biosciences, Franklin Lakes, NJ). Samples consistently achieved >90% CD3⁺CD8⁺ double-positive cells.

Surface, SA-βGal staining, FACS and cyCONDOR analysis

Isolated CD8⁺ T cells were resuspended in RPMI-1640 medium containing 10% fetal bovine serum (FBS, Atlanta Biologicals) and 1× penicillin/streptomycin (Corning, Manassas, VA). The Cellular Senescence Detection Kit-SPiDER-βGal (Dojindo Molecular Technologies, Inc, Rockville, MD) was used to identify senescent CD8⁺ T cells as previously described. Briefly, CD8⁺ T cells were incubated with a 1:1000 dilution of bafilomycin A-1 for 30 min before the addition of 1:1000 dilution SPiDER-βGal for an additional 30 min all whilst at 37C, 5% CO2. SA-βGal stained cells were subsequently washed with 1x PBS (Corning, Manassas, VA) and resuspended in 1xPBS +2% FBS for surface staining. Depending on the experiment, cells were surface stained for CD3, CD8, CCR7 and CD45RA, subsequently washed in 1x PBS and sorted. For ATAC and RNA sequencing, enriched CD8⁺ T cells were sorted based on live, 4^′,6-diamidino-2-phenylindole (DAPI (BioLegend, San Diego, CA) DAPI-, SA-βGal intensity. For CUT&TAG experiments, CD8⁺ T cells were sorted based on the SA-βGal intensities of DAPI-CD3⁺, CD8⁺ T cells. For pharmacological inhibition and siRNA KD of transcription factors, enriched CD8⁺ T cells were sorted based on the SA-βGal intensities of DAPI-, CD8⁺ T cells. All cells were sorted using the BD FACSAria Fusion (BD Biosciences, Franklin Lakes, NJ). All analysis were performed with FlowJo software (FlowJo 10.6.0). For cyCONDOR analysis, live DAPI- CD8⁺ T cell populations from 6 younger and 8 older donors all acquired on BD Symphony (BD Biosciencences) were exported from FlowJo per the developers’ recommendation. 5,000 cells from each donor were extracted randomly, auto-logically transformed, and utilized for downstream analysis. Batch effects were corrected for using the integrated harmony package with batch being defined as different days samples were stained and acquired. UMAP projections and scaled marker expressions were plotted using the standard cyCONDOR workflow.

CD8⁺ T cell RNA interference and ex vivo stimulation

FACS sorted SA-βGal low and high CD8⁺ T cells, or purified CD8⁺ T cells (>90% CD3⁺CD8⁺) were prepared for AMAXA 4D-Nucleofector electroporation (Lonza) per manufactures protocol. Briefly, cells were resuspended in 4D nucleofector solution, with siRNAs (20uM) targeting JUN, KLF5, RUNX2 as well as non-targeting controls (Dharmacon) and electroporated using the preset high functionality (E0115) for primary human T cells. Immediately after nucleofection, 100ul of prewarmed RPMI-1640 + 10% FBS, 1%PS, was added to each well and the plate was placed in a 37C, 5% CO2 incubator for 10 min to allow cells to recover. Afterward, cells were resuspended at 1×10⁶ cells/ml and grown under standard conditions. 24 h later, electroporated cells were collected for RNA, or stimulated with anti-CD3/CD28 dynabeads (Thermo) for an additional 16 h before being collected for RNA or collected for CellTrace Violet (ThermoFisher, Waltham, Ma) staining according to manufacturer’s protocol. Labeled cells were stimulated using soluble anti-CD3 and anti-CD28 antibodies supplemented with 20 U/mL IL-2 (StemCell Technologies). After 7 days, cells were collected, stained with DAPI, and acquired on the BD Symphony (BD Biosciences, Franklin Lakes, NJ).

Transcription factor knockdown efficiency

RNA from primary human CD8⁺ T cells 24 h post electroporation with siRNAs targeting JUN, KLF5, RUNX2 as well as non-targeting controls was purified using a Macherery-Nagel RNA XS Plus Kit per manufactturer’s instructions (Macherery-Nagel, Duren, Germany). Equal amounts of purified RNA were used to generate cDNA with a Bio-Rad iScript cDNA synthesis kit per manufacturer’s instructions (Bio-Rad, Hercules, CA). Qiagen QuantiTect primers were used for JUN, KLF5, RUNX2, and GAPDH as a reference housekeeping gene. Real-time RT-qPCR was performed on a Bio-Rad CFX96 Real-Time PCR detection system (bio-Rad, Hercules, CA). Data were analyzed on the Bio-Rad CFX Manager, software 3.1.

In vitro CD8⁺ T cell expansion

Purified CD8⁺ T cells (>90% CD3⁺CD8⁺) were cultured at 37°C in a humidified atmosphere with 5% CO₂. Cells were maintained in ImmunoCult Expansion Medium (StemCell Technologies, Cambridge, MA) supplemented with 20 U/mL IL-2 (StemCell Technologies). T cells were stimulated every 15 days using soluble anti-CD3 and anti-CD28 antibodies (StemCell Technologies). The growth medium was replenished every 2–3 days to ensure optimal cell viability and proliferation. Cell growth was monitored by counting viable cells using a hemocytometer. Population doublings (PD) were calculated using the formula PD = 3.32*[log(N final)-log(N initial)], where N final is the number of cells recovered and N initial denotes the number of cells initially seeded at each flask.

RNA-seq

RNA from sorted CD8⁺ fSA-βGal low and high from young and old donors (7 each) was purified using a Macherey-Nagel RNA XS Plus kit according to the manufacturer’s instructions (Macherey-Nagel, Duren, Germany). RNA integrity was evaluated in a Bioanalyzer 2100 system, and only RNA with an integrity of number of ≥ 7 was used for library preparation. Libraries were constructed using the SMARTer Stranded V2 (TakaraBio) according to the manufacturer’s instructions (TakaraBio, Mountain View, CA). Paired-end sequencing was performed on an Illumina HiSeq 2500 instrument. At least 40 million reads per sample (20 million per strand) were obtained and used for downstream analyses.

ATAC-seq

The transposition reaction and library construction were performed as previously described in.³⁰ Briefly, 50,000 cells from sorted CD8⁺ fSA-βGal low and high from young and old donors (4 each) were collected, washed in 1× in PBS, and centrifuged at 500 × g at 4°C for 5 min. Nuclei were extracted by incubating cells in nuclear extraction buffer (containing 10 mM Tris-HCl, pH 7.4, 10 mM NaCl, 3 mM MgCl2, 0.1% IGEPAL CA-630) and immediately centrifuged at 500 × g at 4°C for 5 min. The supernatant was carefully removed by pipetting, and the transposition was performed by resuspending nuclei in 50 μL of Transposition Mix containing 1× TD Buffer (Illumina, San Diego, CA) and 2.5 μL Tn5 (Illumina) for 30 min at 37°C. DNA was extracted using a Qiagen MinElute kit (Qiagen, Germantown, MD). Libraries were produced by PCR amplification (12 cycles) of tagmented DNA using an NEB Next High-Fidelity 2× PCR Master Mix (New England Biolabs, Ipswich, MA). Library quality was assessed using an Agilent Bioanalyzer 2100 (Applied Biosystems, Santa Clara, CA). Paired-end sequencing was performed in an Illumina Hiseq 2500 system. Typically, 30–50 million reads per library were required for downstream analyses.

CUT&Tag

We performed CUT&Tag on 75,000–100,000 sorted CD8⁺ fSA-βGal low and high from young donors (n = 3) with the CUTANA CUT&Tag kit (Epicypher) using the following antibodies: H3K4me1 (Epicypher: 13–0057), H3K27me3 (Epicypher: 13–005), H3K27ac (Active Motif: 39,133) and rabbit IgG as a negative control (Epicypher: 13–0042), according to the manufacturer’s instructions. Libraries were amplified for 16 cycles and quality assessed using a TapeStation 4200 instrument. Libraries were paired-end sequenced on a Novaseq-X instrument. 10–20 million reads per library were used for downstream analyses.

Preprocessing of high-throughput sequencing data

Paired-end (RNA-seq, CUT&Tag and ATAC-seq) reads were processed as we previously described³⁶ aligned to the GRCh38.d1.v1 version of the human genome using bowtie2⁷⁵ using the local mode (RNA-seq) or with the following parameters for ATAC-seq and CUT&Tag: -N 0 –no-mixed –no-discordant –maxins 2000 -x –trim3 8. Low-quality reads and adapters were removed using fastq-mcf v.1.0.5 and cutadapt. Alignments were further processed using samtools v.1.1.1, and PCR and optical duplicates were removed with PicardTools v.2.2.2. Enriched regions ATAC-seq were identified using MACS v.2.2.7.1⁷⁶ using a relaxed q-value threshold of 0.01. Master peaksets were constructed using a custom bedops script that merges common peaks between samples as we previously described.¹⁶ CUT&Tag data were processed using chromstaR (see below). For RNA-seq samples, reads were counted using summarized overlaps and normalized with DESeq2 for visualization.

Self-organizing maps (SOMs)

SOM expression portraits were generated using the unsupervised machine learning method deployed in oposSOM.²⁹ Metagenes were visualized in a 60 × 60 grid of rectangular topology, wherein expression portraits are projected by metagene distance matrix similarity using a logarithmic fold-change scale.

Differential expression and accessibility

Differentially accessible regions (DARs) and differentially expressed genes (DEGs) were identified from ATAC-seq and RNA-seq data, respectively using DESeq2. Raw reads are internally normalized by DESeq2 using the median of ratios method.⁸⁴ Reads per peak (ATAC-seq, using the master peakset as feature input) or exon (RNA-seq, using the GRCh38.107 genome model) were quantified using the summarizeOverlaps package and peaks/genes with at least 10 reads in at least 10 libraries were kept. Correction of batch effects was performed with limma⁶⁹ using the nucleic purification date (batch) as the surrogate variable. Data transformation (regularized-log [rld] transformation), exploratory visualization (PCA and hierarchical clustering) and differential expression analysis were performed with DESeq2 using the default parameters as previously described⁷⁷ and base R functions. We focused on highly significant DARs and DEGs by using an adjusted p-value filter of 0.05.

WGCNA and integration of chromatin accessibility with differential expression

Differentially expressed genes (DEGs) and accessible regions (DARs) identified with DESeq2 were used as input for unsupervised clustering using WGCNA. We used the “signed” option with default parameters, except for the soft thresholding power which was set to 17 for RNA-seq data and 6 for ATAC-seq data. The minimum size for the DEG and DAR modules was set to 200 features for the initial set of modules, which were then merged by a dissimilarity threshold of 0.5 for RNA-seq data and 0.2 for ATAC-seq data. For RNA-seq, WGCNA modules were functionally profiled using clusterProfiler⁷¹ using the Molecular Signatures Database Hallmark gene sets.⁸⁵ Statistical significance was calculated by a hypergeometric test with a cut-off of an adjusted p-value of ≤ 0.1 with Benjamini–Hochberg correction. Integration of DARs with DEGs was achieved by annotation of DARs to the nearest gene using ChIPSeeker⁷⁹ followed by filtering against the list of DEGs.

Chromatin state transitions

We analyzed the genome-wide combinations of H3K4me1, H3K27ac, H3K27me3 and ATAC-seq signals modeling senescence as a transition from fSA-βGal low to fSA-βGalusing the chromstaR package essentially as we previously described.³⁶ The algorithm was run on differential mode and configured to partition the genome in 200 bp non-overlapping bins and count the number of reads of histone modification/ATAC-seq mapping into each bin and modeled using a univariate HMM based on a two-component mixture, the zero-inflated binomial distribution. Subsequently, a multivariate HMM assigns every bin in the genome to one of the multivariate components considering 2 (2 cell states × 4 genomic enrichment variables [histone modifications and ATAC-seq]) possible states. We focused on robust transitions of an enrichment score of ≥ 1.5 for further analysis. Integration with gene expression data was achieved by annotating the nearest DEGs with ChIPSeeker⁷⁹ to the 10 most frequent chromatin state transitions and monitoring their expression using violin plots. The association between chromatin state transitions and DEG modules was performed through Correspondence analysis (CA) and visualized on an asymmetric biplot after filtering for the top contributing chromatin state transitions (square cosine >0.5).

Heatmap and metaprofile visualizations of ATAC-seq and CUT&Tag datasets

ATAC-seq and CUT&Tag alignments were normalized using deeptools v3.3.1100103⁸⁰ using the RPGC approach to obtain 1X coverage (bamCoverage -b–normalizeUsing RPGC–effectiveGenomeSize 2864785220 –ignoreDuplicates–binSize 10 –verbose -o).

Transcription factor footprinting

Footprinting, TF motif enrichment, and differential binding activity were performed as we previously described³⁶ with HINT-ATAC⁷⁴ using the JASPAR position weight matrix database for vertebrate TFs⁸⁶ on merged ATAC-seq datasets, focusing on differentially accessible chromatin in fSA-βGal CD8⁺ T cells. TF co-binding was assessed across ≤500bp bins of differentially accessible chromatin using a hypergeometric test for each pairwise co-binding possibility for each condition (fSA-βGal low and high) and adjusted p-values for multiple testing using Bonferroni correction.

Transcription factor networks

We built TF hierarchical networks as previously described in,¹⁷ at opening and closing DARs. Each identified TF is represented as a node, connected by directed edges representing co-binding events along the chromatin and during the transition from fSA-βGal low to fSA-βGal high. An edge with TF A as the source and TF B as the target indicates that TF A binds to at least 15% of the same chromatin regions as TF B. For each TF, we computed their normalized total number of bound regions and dynamicity and represented those properties in the networks as node size and node color, respectively. Each edge is associated with a weight in the interval [0, 1], representing the fraction of TF B binding instances previously occupied by TF A. We filtered the networks for edges with a weight higher than 0.15 and performed a transitive reduction to simplify the obtained networks while keeping essential topological features. Nodes included in the same strongly connected component (SCC), i.e., connected both by incoming and outgoing paths, were merged into a single node. We represented the identified transcriptomic module distribution for TFs included in the same SCC as Doughnut Charts.

Pharmacological perturbation of transcription factor networks

Sorted SA-βGal high CD8⁺ T cells were young donors (n = 3) cultured for three days in RPMI-1640 medium supplemented with 10% fetal bovine serum (FBS; Atlanta Biologicals), 1× penicillin/streptomycin (Corning), and 20 U/mL IL-2. Cells were treated with either a vehicle control (DMSO) or one of the following inhibitors at a concentration of 30 μM: SRI8662, T-5224, or CADD522 (Selleckchem, Houston, TX), targeting KLF5, AP1, and RUNX2, respectively. Following the three-day treatment period, cells were collected and RNA and nuclei isolated for RNA-seq and ATAC-seq as described above. To evaluate the effect of pharmacological inhibition of TF network nodes, we annotated KLF5, AP1 and RUNX2 footprints in DARs of fSA-βGal high CD8⁺ T cells to the nearest DEG as defined above to identify their potential gene targets, and the effect of TF inhibition on target DEG expression in both opening and closing chromatin relative to DMSO-treated cells was evaluated by clustering the expression levels of TF target DEGs using WGCNA as above (soft-threshold power of 6 and merge distance of 0.4 for opening chromatin; soft-threshold power of 12 and merge distance of 0.4 for closing chromatin) and visualized on a heatmap.

GSEA

GSEA of 4 CD8⁺ T cell senescence gene signatures derived from our transcriptomic datasets was performed on transcriptomes (GEO: GSE223655) of naive CD8⁺ T cells and CAR-T cell products from diffuse large B cell lymphoma (DLBCL) patients undergoing immunotherapy were phenotypically classified as ‘progressive disease or ‘complete response’ and transcriptomes from peripheral CD8⁺ T cells from control and individuals with active and inactive systemic lupus erythematosus (SLE) (GEO: GSE97264) using the GSEA GUI version 4.2.2. Probe sets were collapsed to the gene level using the correlation-based approach. The correlation of probe sets representing the same gene was computed to decide whether to average probe sets (c > 0.2) or to use the probe set with the highest average expression across samples (c ≤ 0.2). Probe sets without known annotation were removed. The signal-to-noise ratio (μA − μB)/(σA + σB) (μ represents the mean, σ the standard deviation) was used as a ranking metric and statistics based on 500 gene set permutations.

QUANTIFICATION AND STATISTICAL ANALYSIS

With the exception of sequencing analyses, all statistical test utilized in this study were performed in GraphPad Prism as indicated in the legend of their respective figure. Unless otherwise stated, all experiments were performed with at least 3 biological replicates (at least 3 independent donors) and conducted as at least 2 independent experiments.

Supplementary Material

SUPPLEMENTAL INFORMATION

Supplemental information can be found online at https://doi.org/10.1016/j.celrep.2025.116795.

KEY RESOURCES TABLE.

REAGENT or RESOURCE	SOURCE	IDENTIFIER
Antibodies
H3K4me1	Epicypher	Cat# 13-0057; RRID: AB_3076424
H3K27me3	Epicypher	Cat# 13-0055; RRID: AB_3665059
H3K27ac	Active Motif	Cat# 39133; RRID: AB_2561016
CD3-APC (SK7)	BD biosciences	Cat# 340661; RRID: AB_400530
CD8-PE (SK1)	BD biosciences	Cat# 555635; RRID: AB_395997
CD8-AF700 (SK1)	Biolegend	Cat# 344724; RRID: AB_2562790
CD45RA-APC (HI100)	Biolegend	Cat# 304112; RRID: AB_314416
CCR7-PE Cy7 (G043H7)	Biolegend	Cat# 353226; RRID: AB_11126145
Biological samples
Primary human CD8+ T cells	Lab sourced or New York Blood Center	Table 1
Chemicals, peptides, and recombinant proteins
T-5224	Selleckchem	Cat# S8966
SR18662	Selleckchem	Cat# S8900
CADD522	Selleckchem	Cat# S0790
Critical commercial assays
BD Vactunainer	BD Biosciences	Cat# 366480
EasySep Human CD8 T cell Iso Kit	Stem cell technologies	Cat# 17953
SPiDER-βGal	(Dojindo Molecular Technologies	Cat# SG04-10
NucleoSpin® RNA Plus XS	Macherey-Nagel Inc	Cat# 740990.250
Illumina Tagment DNA Enzyme and Buffer Kit	Illumina	Cat# 20034210
iScript™ cDNA Synthesis Kit	Bio-Rad	Cat# 1708891
iTaq Universal SYBR® Green Supermix	Bio-Rad	Cat# 1725124
Human CD3/CD28 T cell Activator	Stem cell technologies	Cat# 78220
Dynabeads Human T-Activator CD3/CD28	Thermo Fisher	Cat#11161D
ImmunoCult-XF T cell Exp Medium	Stem cell technologies	Cat# 78220
Human Recom IL-2	Stem cell technologies	Cat# 78220
Cell Trace Violet	Invitrogen	Cat# C34557ZO
RPMI-1640	Thermo Fisher Scientific	Cat# 61870036
Deposited data
GEO: GSE310382	Gene expression omnibus	https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE310382
GEO: GSE310383	Gene expression omnibus	https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE310383
GEO: GSE310384	Gene expression omnibus	https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE310384
GEO: GSE310393	Gene expression omnibus	https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE310393
GEO: GSE310625	Gene expression omnibus	https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE310625
GEO: GSE310729	Gene expression omnibus	https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE310729
GEO: GSE169009	Gene expression omnibus	https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE169009
Oligonucleotides
GAPDH QuantiTect Primers	Qiagen	Cat# QT00079247
JUN QuantiTect Primers	Qiagen	Cat# QT00242956
KLF5 QuantiTect Primers	Qiagen	Cat# QT00074676
RUNX2 QuantiTect Primers	Qiagen	Cat# QT00998102
JUN ON-TARGETplus siRNA	Dharmacon	Cat# LQ-003268-00-0010
KLF5 ON-TARGETplus siRNA	Dharmacon	Cat# LQ-013571-00-0010
RUNX2 ON-TARGETplus siRNA	Dharmacon	Cat# LQ-012665-00-0010
Software and algorithms
oligo v.1.62.0	Carvalho et al.68	https://doi.org/10.1093/bioinformatics/btq431
Limma v.3.54.0	Ritchie et al.69	https://doi.org/10.1093/nar/gkv007
oposSOM v.2.16.0	Loffler-Wirth et al.29	https://doi.org/10.1093/bioinformatics/btv342
WGCNA v.1.71	Langfelder et al.70	https://doi.org/10.1186/1471-2105-9-559
clusterProfiler v.4.6.0	Wu et al.71	https://doi.org/10.1016/j.xinn.2021.100141
DREM v.2.0.4	Schulz et al.72	https://doi.org/10.1186/1752-0509-6-104
fastq-mcf v.1.0.5	Aronesty73	https://code.google.com/p/ea-utils/
samtools v.1.1.1	Li et al.10,74	https://www.htslib.org/
PicardTools v.2.2.2	Broad Institute	https://broadinstitute.github.io/picard/
Bowtie2 v.2.4.4	Langmead et al.75	https://doi.org/10.1038/nmeth.1923
MACS v.2.2.7.1	Liu76	https://doi.org/10.1007/978-1-4939-0512-6_4
DESeq2 v.1.36.0	Love et al.77	https://doi.org/10.1186/s13059-014-0550-8
chromstaR v.1.22.0	Hanna et al.35	https://doi.org/10.1038/s41594-017-0013-5
FactoMineR v.2.6.0	Le et al.78	https://cran.r-project.org/web/packages/FactoMineR/index.html
Vennerable v.3.1.0.9000	Swinton, Github Inc.	https://github.com/js229/Vennerable
ChIPseeker v.1.32.0	Yu et al.22,79	https://www.bioconductor.org/packages/release/bioc/html/ChIPseeker.html
deeptools v3.3.1	Ramirez et al.80	https://deeptools.readthedocs.io/en/latest/
HINT-RGT v.0.13.2	Li et al.10,74	https://reg-gen.readthedocs.io/en/latest/hint/introduction.html
TFBSTools v.1.34.0	Tan & Lenhard	https://bioconductor.org/packages/release/bioc/html/TFBSTools.html
Igraph v.1.3.4	Gustavsen et al.	https://doi.org/10.12688/f1000research.2088.3
Rcytoscape v.2.13	Bioconductor Open-Source Software for Bioinformatics	https://www.bioconductor.org/packages//2.12/bioc/html/RCytoscape.html
CyREST API	Ono et al.	https://www.bioconductor.org/packages/release/bioc/html/RCy3.html
Bedtools v.2.30.0	Quinlan et al.	https://bedtools.readthedocs.io/en/latest/
Bedops v.2.4.41	Neph et al.81	https://bedops.readthedocs.io/en/latest/
GraphPad Prism version 9.5	GraphPad By Dotmatics	https://www.graphpad.com/
cyCONDOR	Kröger et al.	https://github.com/lorenzobonaguro/cyCONDOR
FlowJo (v10.10)	BD Biosciences	https://www.flowjo.com/
Harmony	Korsunsky et al.82	https://portals.broadinstitute.org/harmony/articles/quickstart.html
FlowSOM	Van Gassen et al.28	https://www.bioconductor.org/packages/release/bioc/html/FlowSOM.html

Highlights.

• Aging humans accumulate senescent CD8⁺ T cells across all differentiation states

• Age-insensitive transcription factor networks control the transition to senescence

• Targeting the senescence TF networks partially restores response to TCR stimulation

• CD8⁺ T cell senescence gene signatures predict the response to CAR-T cell therapy