Dynamic changes in chromatin accessibility in CD8⁺ T cells responding to viral infection

Summary

In response to acute infection, naive CD8⁺ T cells expand, differentiate into effector cells and then contract to a long-lived pool of memory cells after pathogen clearance. During chronic infections or in tumors, CD8⁺ T cells acquire an “exhausted” phenotype. Here we present genome-wide comparisons of chromatin accessibility and gene expression from endogenous CD8⁺ T cells responding to acute and chronic viral infection using ATAC-seq and RNA-seq. Acquisition of effector, memory or exhausted phenotypes was associated with stable changes in chromatin accessibility away from the naive T cell state. Regions differentially-accessible between functional subsets in vivo were enriched for binding sites of transcription factors known to regulate these subsets, including E2A, BATF, IRF4, T-bet and TCF1. Exhaustion-specific accessible regions were enriched for consensus binding sites for NFAT and Nr4a family members, indicating that chronic stimulation confers a unique accessibility profile on exhausted cells.

Introduction

During acute viral infection, naive antigen-specific CD8⁺ T cells expand and differentiate to yield effector T cells that enable resolution of the infection (Williams and Bevan, 2007). After viral clearance, most expanded cells die, but a small proportion survive as a long-lived memory population that rapidly produces cytokines and reacquires cytotoxic activity upon secondary exposure to antigen, thus providing protective immunity. These phases of the immune response are typical of acute infections where the virus is eliminated, such as with the Armstrong (Arm5) strain of lymphocytic choriomeningitis virus (LCMV) (Wherry and Ahmed, 2004). When virus persists during chronic or latent infection – as in mice or humans infected with LCMV clone 13, human immunodeficiency virus, hepatitis B virus and hepatitis C virus – CD8⁺ T cells often enter a state of unresponsiveness to further stimulation that has been termed “exhaustion” (Moskophidis et al., 1993; Wherry, 2011). Exhausted CD8⁺ T cells are hyporesponsive to stimulation with decreased cytokine production, reduced ability to lyse target cells and increased expression of several inhibitory cell surface receptors including PD-1 (programmed death 1), LAG3 (lymphocyte-activation gene 3), TIM3 (T cell immunoglobulin and mucin domain-containing protein 3), and CTLA-4 (cytotoxic T lymphocyte-associated protein 4) (Barber et al., 2006; Blackburn et al., 2009; Wherry et al., 2003; Yamamoto et al., 2011). The exhausted phenotype is also common in tumour-infiltrating CD8⁺ cells (Crespo et al., 2013), and antibody therapy targeting inhibitory receptors (“checkpoint blockade”) has been remarkably effective in cancer immunotherapy (Pardoll, 2012).

While several transcription factors (TF), including BATF, IRF4 and the T-box family members T-bet and eomesodermin (Eomes) are known to participate in the formation and function of effector and memory CD8⁺ T cells (Dominguez et al., 2015; Intlekofer et al., 2005; Kurachi et al., 2014), the molecular mechanisms that determine the exhausted phenotype are not well understood (Speiser et al., 2014; Wherry and Kurachi, 2015). Persistent antigen stimulation appears to be a dominant factor in inducing exhaustion in tumor-infiltrating T cells (Schietinger et al., 2016); consistent with these findings, we previously implicated a gene expression program driven by the antigen-activated TF NFAT in CD8⁺ T cell exhaustion (Martinez et al., 2015). However, given the limiting cell numbers available, it was technically difficult to confirm these findings by performing ChIP-seq (chromatin immunoprecipitation followed by sequencing) for NFAT1 and other TFs in exhausted cells.

Changes in transcriptional programs are controlled not only through the action of TFs near TSSs, but also through epigenetic changes in a variety of DNA and histone modifications at regulatory elements throughout the genome (Kouzarides, 2007). Active regulatory elements that bind TFs can be defined operationally by their “accessibility” to enzymes such as DNase I and micrococcal nuclease (Consortium, 2012; Vierstra et al., 2014). While DNase I hypersensitivity assays are relatively cumbersome and require large numbers of cells (Pipkin et al., 2010), accessible regions of chromatin can be identified reliably even in small cell numbers by ATAC-seq (assay for transposase-accessible chromatin using sequencing), a technique that measures accessibility to an engineered Tn5 transposon containing flanking sequencing adapters (Buenrostro et al., 2013).

Here we have mapped accessible regulatory elements by ATAC-seq in naive, effector, memory and exhausted CD8⁺ T cells from mice with acute or chronic LCMV infection. We identified dynamic changes in chromatin accessibility in CD8+ T cells, with clusters of regions with shared accessibility profiles between different subsets. By motif enrichment analysis of the regulatory elements and comparison to transcriptional profiles obtained by RNA-sequencing (RNA-seq), we have confirmed the involvement of NFAT in the exhaustion program and implicated new NFAT-induced TFs in CD8⁺ T cell exhaustion. Our data constitute a comprehensive analysis of the similarities and differences among functionally distinct CD8⁺ T cell subsets, and provide a valuable resource for future comparisons of these cell types in different tissues and disease models, or after genetic manipulation or antibody blockade.

Results

Identification of accessible regions in CD8⁺ T cells

We used ATAC-seq to assess the accessible regions genome-wide in CD8⁺ T cells responding to viral infection. Antigen-specific CD8⁺ T cells, defined as CD44-high and staining with H-2D^b tetramers containing a peptide from LCMV glycoprotein amino acids 33-41 (H-2D^b-gp33-41), were collected by fluorescence-activated cell sorting (FACS) from mice after LMCV Arm5 or clone 13 infection (Figures 1A, S1A,C–E). Total effector cells and KLRG1⁺IL7Rα⁻ (short-lived effector, SLEC) or KLRG1⁻IL7Rα⁺ (memory precursor effector cells, MPEC) subsets (Figure S1A–C) were collected 8 days after LCMV Arm5 infection. The frequencies of SLEC and MPEC were around 50% and 8% of total H-2D^b-gp33-41 tetramer-positive cells, respectively (Figure S1B,C). At 35 days after infection with LMCV Arm5, we sorted cells expressing IL7Rα (memory) or KLRG1 (d35 KLRG1⁺) (Figure S1D). Cells with an exhausted phenotype (PD-1^hi,Tim3^hi) were isolated from mice infected with LCMV clone13 for 20 days (Figure S1E). We also compared ATAC-seq signals in naive splenic CD8⁺ T cells and in cultured CD8⁺ T cells after in vitro stimulation under various conditions or after transduction with retroviral constructs (Figure 1A).

Figure 1. High ATAC-seq signal in CD8⁺ T cells at conserved regions in promoters and distal regulatory elements.

A) CD8⁺ T cell populations collected for ATAC-seq comparison. B) Mean ATAC-seq coverage at the 70kb Ifng locus with a scale of 0-1200 for all tracks. C) k-means clustered heat map of mean normalized counts or log₂ fold-change from global mean at all peaks. D) Pairwise euclidian distance comparison of asinh transformed ATAC-seq signal per peak for all populations using all peaks accessible in at least one cell type. Data in B,C,D are from mean of at least 2 independent samples, except for a single d35 KLRG1+ replicate. See also Figure S1.

As in previous studies (Buenrostro et al., 2013), ATAC-seq coverage in these samples was concentrated with high signal in narrow regions throughout the genome (illustrated for Ifng in Figure 1B). We called peaks on individual replicates (Figure S1F), merged these to 91,099 regions, counted the number of Tn5 insertions from each replicate in all peaks, and normalized these data using signals at regions with low variance with DESeq2 (Love et al., 2014, and Supplemental Experiment Procedures). Regions with higher signal had lower inter-replicate error (Figure S1G, H), thus we restricted our analysis to 70,532 regions that met a minimum threshold of mean ATAC-seq signal in at least one cell type (Supplementary File 1 and Supplemental Experimental Procedures). Comparing the signal at all regions between all samples, each biological replicate clustered with the other replicates from the same group, and two technical replicates were as similar to each other as they were to independent biological replicates (Figure S1I). We also profiled transcriptional differences by RNA-seq of naive, effector, memory, and exhausted cells; each biological replicate clustered with replicates in the same cell type (Figure S1J, Supplementary File 1).

In 70 kb of the Ifng locus, we identified 19 peaks that were accessible in at least one cell type (Figure 1B). The conserved region located at −22 kb relative to the transcription start site (TSS) was accessible in all samples, whereas other conserved regions were not accessible in naive T cells, but were highly accessible in effector, memory and exhausted T cells (Figure 1B). Cells that were pre-activated, cultured and then briefly reactivated for short periods (2 hr) in vitro showed variable accessibility at these regions, as did cells transduced with a constitutively-active NFAT1 that cannot cooperate with AP-1 (NFAT-CA-RIT, Figure 1B) (Martinez et al., 2015). The sequences around ATAC-seq peaks were highly conserved (Figure S1K), and the peaks were frequent in promoters, introns, and distal intergenic regions (Figure S1L). While the Ifng locus illustrates changes in accessibility near the TSS and distal conserved elements, the signal at most TSSs was less variable compared to distal elements (Figure 1C).

Considering all peaks, T cells that had responded to LCMV infection were most similar to each other, while naive T cells or cells activated and cultured in vitro were part of a separate cluster (Figures 1D, S1I). To identify individual regions that were differentially-accessible between CD8⁺ T cell subsets, we computed differential coverage using DESeq2. Comparing CD8⁺ T cells activated and cultured in vitro to effector cells at day 8 post-LMCV Arm5 infection, at least 9,000 regions were differentially-accessible (Figure S1M). Of these, differentially-accessible regions that were more accessible in cells cultured in vitro than in ex vivo isolated effector cells were also more accessible in naive CD8⁺ T cells; conversely, regions that were less accessible in cells cultured in vitro than in ex vivo isolated effector cells also tended to be less accessible in naive cells, compared to effector, memory, and exhausted CD8⁺ T cells (Figure S1N). Thus the accessibility profile of CD8⁺ T cells activated and cultured in vitro under these conditions remained more similar to that of naive CD8⁺ T cells and did not completely recapitulate the accessibility profile of effector CD8⁺ T cells that develop over several days of viral infection in vivo. We previously found that high (100 U/ml) or low (10 U/ml) concentrations of IL-2 can bias CD8⁺ T cells cultured in vitro toward effector or memory phenotypes respectively (Pipkin et al., 2010). However, very few regions were differentially-accessible between high and low IL-2 conditions (Figure S1O), indicating that exposure to IL-2 under these conditions did not alter chromatin accessibility substantially, despite clear differences in transcriptional profiles.

Stable and dynamic changes in accessible chromatin of CD8⁺ T cells in acute viral infection

We assessed chromatin accessibility changes during in vivo responses to acute viral infection by comparing naive, effector and memory CD8⁺ T cells. Of the 45,489 regions that were accessible in any of these subsets, more than 12,000 were differentially-accessible when comparing naive and effector T cells (Figure 2A); fewer (~5100) were differentially-accessible between naive and memory (Figure 2B); and effector and memory T cells were the most similar with only ~3000 differentially-accessible regions (Figure 2C). Most of these differentially-accessible regions were located distal to the TSS, while commonly accessible regions were more evenly distributed between TSS proximal and distal elements (Figures S2A–C). Similarly, comparison of RNA-seq data identified more than 1500 differentially-expressed genes between naive and effector cells (Figure S2D), but only 600–800 differentially-expressed genes between naive and memory (Figure S2E) or memory and effector (Figure S2F) cells. The differentially-accessible regions were positively associated with changes in gene expression at genes located within 25 kb of the peaks (Figure S2G–I); for example, genes near peaks with higher ATAC-seq signal in effector calls than in naive had higher expression in effector cells compared to naive (Figure S2G).

Figure 2. Dynamic changes in chromatin accessibility occur in antigen specific effector and memory CD8⁺ T cells responding to acute viral infection.

A–C) Scatterplots of mean ATAC-seq counts per peak comparing the indicated samples. D–F) Boxplots of ATAC-seq counts per peak from the indicated samples (labeled at bottom) at common or differentially-accessible regions from the comparison labeled above. Box indicates interquartile range with whiskers +/−1.5 times this range and outlier points. G–K) Mean ATAC-seq coverage at Il7r (G), Ccr7 (H), Gzma (I), Gzmk (J), Dmrta1 (K) loci with a scale of 0-1200 (left) or RNA-seq gene expression for the indicated genes (right). L,M) Venn diagrams illustrating intersection of differentially-accessible regions from pairwise comparisons of naive, effector, and memory CD8⁺ T cells characterizing regions “specific” to a subset (L) or “not” in a subset (M) with p values and odds ratios from Fisher's test comparisons. ATAC-seq data in A–K are from at least 2 independent replicates. RNA-seq data in G–K are mean of two independent replicates for RNA-seq. See also Figure S2.

Regions that were less accessible in effector than in naive cells also tended to be less accessible in exhausted cells (Figure 2D). In memory cells, some of these regions had a high signal similar to naive (Il7r upstream region, Figure 2G) whereas others had a low signal similar to effector (Ccr7 downstream region, Figure 2H). Similar patterns were apparent at regions more accessible in effector than naive. For example, within the Gzma locus, regions such as the +4kb and +19kb peaks had higher signal in effector than naive, however the +19kb peak was high in memory cells, while the +4kb peak was low in memory cells (Figure 2I). The signals at some activation-induced genes, including Fasl and Prf1, were highly similar between naive, effector, and memory CD8⁺ T cells despite differences in gene expression (Figure S2J), suggesting that changes in expression of these genes would most likely result from differential activation by TFs binding to these commonly accessible regions.

We next focused on differentially-accessible regions between memory cells and either naive or effector cells. There were more differentially-accessible regions between naive and memory (Figure 2B) than between effector and memory (Figure 2C). Regions with lower signal in memory than naive also tended to have lower signal in effector and exhausted cells, while peaks with higher signal in memory than naive, tended to have higher signal in effector and exhausted cells (Figure 2E). Taken together, these comparisons identified regions, similar to the −12kb and +2.5kb peaks in the Gzmk locus and the +19kb peak in the Gzma locus, which were stably altered in naive T cells after viral infection (Figure 2I and J). Despite the low number of differentially-accessible regions between effector and memory cells (Figure 2C), many of the regions were associated with the greatest variation among naive, effector, memory, and exhausted T cell populations (Figure 2F).

In an effort to identify accessible regions that were biased towards naive, effector, or memory CD8⁺ T cells, we assessed the intersection of the differentially-accessible regions from each comparison (Figure 2L). While a few regions were memory-specific (Figure 2L), such as the Dmrta1 TSS (Figure 2K), substantial fractions of differentially-accessible regions between memory and naive or effector cells were shared with other cell types (Figure 2M), and fewer than expected were absent in memory, but shared by effector and naive (Figure 2M).

These data indicated that at least 25% of the accessible regions changed at some point during CD8⁺ T cell activation and differentiation during acute infection and that 5–10% of the CD8⁺ T cell accessible landscape underwent a stable change from a naive T cell after response to an acute viral infection. The effector population had some unique changes, with increased accessibility at some regions and reduced accessibility at others compared to naive and memory CD8⁺ T cells. Finally, most of the accessibility profile of memory CD8⁺ T cells was shared with either naive or effector T cells, with only a small fraction of total accessible regions unique to this population.

Chromatin accessibility profiles of SLEC and MPEC are similar to those of effector and memory cells

We next compared cells with a terminal effector phenotype (SLEC) or a memory precursor phenotype (MPEC) at day 8 post-LMCV Arm5 infection to assess the similarities of these subsets to effector and memory cells. Based on the signal at all accessible regions, SLEC were very similar to total effector and MPEC were less similar to both effector and memory cells (Figures 1D and S1I). In the direct comparison of SLEC and MPEC, we identified fewer than 1000 differentially-accessible regions (Figure 3A). Regions that were more accessible in SLEC than MPEC tended to have high accessibility in effector cells with progressively lower signal in exhausted, memory, and naive cells, supporting the overall similarity between effector and SLEC (Figure 3B). Regions that were more accessible in MPEC than SLEC tended to have high signal in naive and memory, with lower signal in effector and exhausted (Figure 3B). Together these data suggested while MPEC were very similar to effector cells, there was a partial bias towards the memory accessibility profile. This trend was also apparent when comparing the change in signal between SLEC and MPEC at differentially-accessible regions between memory and effector cells. Peaks with higher signal in memory than effector also had higher signal in in MPEC than SLEC (Figure 3C), while regions higher in effector than memory tended to have increased signal in SLEC compared to MPEC (Figure 3C), including the region near the Klrg1 TSS (Figure 3D). This bias was incomplete, as exemplified by the region 4–5 kb upstream of Aurkb, which included several peaks with comparable signal between MPEC, SLEC and effector cells but lower in memory and naive.

Figure 3. Memory precursor effector cells are similar to short lived effector cells with a slight bias towards memory.

A) Scatterplot of mean ATAC-seq counts per peak comparing the SLEC and MPEC. B) Boxplot of ATAC-seq counts per peak from the indicated samples (labeled at bottom) at common or differentially-accessible regions from the comparison labeled above. Box indicates interquartile range with whiskers +/−1.5 times this range and outlier points. C) Histograms of the log₂ fold-change between effector and memory cells at (top) or SLEC and MPEC (bottom) at regions differentially-accessible between effector and memory. D) Mean ATAC-seq coverage at Klrg1 and Aurkb loci with a scale of 0-1200. E) RNA-seq gene expression for Klrg1 and Aurkb. Data in A–D are from 3 independent replicates and E is mean of 2 independent replicates. See also Figure S3.

We also compared the accessibility between SLEC and effector as well as MPEC and memory (Figure S3A,B). Although SLEC were collected from female mice and effector cells were collected from male mice, there were fewer than 50 differentially-accessible regions between these two subsets (Figure S3A). In addition, all but 2 of the regions more accessible in female SLEC than male effector cells were on the X chromosome (data not shown) and there were many female-specific peaks within the Xist/Tsix and Firre loci (Figure S3C). By comparison, there were more than 700 regions differentially-accessible between female MPEC and male memory cells (Figure S3B). Those that were higher in MPEC tended to also have high signal in effector cells while those that were higher in memory tended to have high signal in memory and naive cells, indicating that these regions were largely associated with different CD8⁺ T cell function rather than reflecting the sex differences. Altogether, these data indicate that the accessibility assessed using the total effector population is highly similar to the SLEC subset, largely due to the relatively small fraction of MPEC (Figure S1B) and their similarity to SLEC (Figure 3A). MPEC were slightly biased towards the memory cell profile, but the chromatin profile of MPEC was more similar to that of SLEC (Figure 3A,S3B).

Exhaustion-specific changes in chromatin accessibility

We next compared effector and memory CD8⁺ T cells to CD8⁺ T cells with an exhausted phenotype, isolated from mice chronically infected with LCMV clone 13. There were approximately 5000 and 3000 differentially-accessible regions in cells with an exhausted phenotype compared to memory and effector CD8⁺ T cells, respectively (Figures 4A,B). As observed in comparisons of naive, effector and memory CD8⁺ T cells (Figure S2A–C), most of these differentially-accessible regions were distal to the TSS (Figure S4A, B). Similar trends were observed for RNA-seq comparisons between exhausted cells and either effector or memory cells (Figure S4C,D), where more genes were differentially expressed between exhausted and memory than between exhausted and effector. As observed for comparisons of naive, effector, and memory cells, the changes in accessibility were more likely to be positively associated with changes in expression at nearby genes (Figure S4E,F).

Figure 4. Chronic activation profile identified by comparison of viral antigen specific effector, memory, and exhausted CD8⁺ T cells.

A,B) Scatterplots of mean ATAC-seq counts per peak comparing the indicated samples. C,D) Boxplots of ATAC-seq counts per peak from the indicated samples (labeled at bottom) at common or differentially-accessible regions from the comparison labeled above. Box indicates interquartile range with whiskers +/−1.5 times this range and outlier points. E–G) Mean ATAC-seq coverage at Havcr2 (E), Tox2 (F), and Satb1 (G) loci with a scale of 0-1200 (left) or RNA-seq gene expression for the indicated genes (right). H,I) Venn diagrams illustrating intersection of differentially-accessible regions from pairwise comparisons of effector, memory, and exhausted CD8⁺ T cells characterizing regions “specific” to a subset (H) or “not” in a subset (I) with p values and odds ratios from Fisher's test comparisons. ATAC-seq data in A–G are from at least 2 independent replicates. RNA-seq data in E–G are mean of two independent replicates. See also Figure S4.

The accessibility profiles of effector and exhausted CD8⁺ T cells were similar, exemplified by the −7kb and +19kb peaks in the Havcr2 (encoding Tim3) locus (Figure 4E). Regions with unique accessibility in exhausted cells were most clearly identifiable in comparison with effector cells (Figures 4B,D) than by comparison with memory cells (Figures 4A,C), such as the −19kb and −61kb peaks near the Tox2 locus (Figure 4F) associated with high Tox2 gene expression (Figure 4F). Cells with an exhausted phenotype were also characterized by loss of accessibility compared to effector and memory CD8⁺ T cells (Figures 4C,D), such as intronic peaks in the Satb1 locus, associated with low Satb1 expression (Figure 4G). To identify features that were specific to, or absent from, effector, memory and exhausted CD8⁺ T cells, we compared by intersections the differentially-accessible regions among these groups. More regions were specific to memory or exhausted cells than to effector cells (Figure 4H). Relatively few regions were less accessible in effector than either memory or exhausted cells (Figure 4I), but there were 762 regions with low accessibility in exhausted cells compared to both effector and memory.

Despite exhaustion specific changes, these data indicated that CD8⁺ T cells with an exhausted phenotype responding to a chronic infection shared a substantial fraction of chromatin accessibility with effector CD8⁺ T cells responding to acute infection. In particular, the accessibility at key effector-related genes including Ifng, Gzma, Gzmk, Fasl, and Prf1 (Figures 1B,2I,2J, Fig S2J) and inhibitory receptor loci including Havcr2/Tim3, Lag3 and Ctla4 (Figures 4E, S4G) was similar between effector and exhausted cells.

Differentially-accessible regions are associated with regulators of CD8 T cell differentiation

We identified 18,043 regions that were differentially-accessible in at least one pairwise comparison of naive, effector, SLEC, MPEC, memory and exhausted cells and compared the relative location of individual replicates on multidimensional scaling plots computed using ATAC-seq signal at these regions. SLEC, effector and exhausted cells were most distant from naive cells, while MPEC and memory cells were closer to naive (Figure 5A). The substantial overlap between groups of differentially-accessible regions (Figure 2L,2M,4H,4I) suggested that there are functionally related groups of accessible regions that may explain the dissimilarity (Figure 5A). Using k-means clustering, we partitioned all 18,043 differentially-accessible regions into 12 groups with shared profiles between naive, effector, SLEC, MPEC, memory, and exhausted CD8⁺ T cells (Figure 5B). In this comparison, relatively few effector- or memory-specific regions were apparent, with slight biases in clusters 11 and 8 in favor of effector and memory, respectively. In contrast, cluster 10 and clusters 3 and 6 identified groups that were largely biased towards naive or exhausted, respectively (Figure 5B).

Figure 5. Differentially-accessible regions in CD8⁺ T cells are associated with bhLH, bZIP, HMG, T-box, NR, and RHD family TFs.

A) Two dimensional multidimensional scaling plot of ATAC-seq signal for all replicates of naive, effector, SLEC, MPEC, memory, and exhausted cells at 18,043 regions differentially-accessible regions identified from comparisons of naive, effector, memory, and exhausted cells. B) k-means clustered log₂ fold-change from mean ATAC-seq signal for all differentially-accessible regions identified from comparisons between naive, effector, SLEC, MPEC, memory, and exhausted CD8⁺ T cells. C) Enrichment of all known motifs within each cluster of differentially-accessible regions compared to all accessible regions in naive, effector, memory, and exhausted CD8⁺ T cells. All motifs with an enrichment log p-value less than −15 and found in 10% or more regions in at least one cluster are shown. D) Percent of each cluster of ATAC-seq peaks that overlap ChIP-seq peaks or the percent of all differentially-accessible regions in each cluster. The total number of ChIP-seq peaks for each TF and the fraction of these that overlap any of these differentially-accessible regions are shown below the plot. E) log₂ fold-change from mean RNA-seq counts per transcript are shown for all expressed TFs from families associated with each enriched motif. F) MeDIP-seq coverage compared to input for naive and effector CD8⁺ T cells 8 days after LCMV Arm5 infection. The top graph is for all accessible regions in CD8⁺ T cells, where each graph below is associated with the clusters indicated at left in panel B. ATAC-seq data in A and B are mean of at least 2 independent replicates and RNA-seq data in E are mean of 2 independent replicates. See also Figure S5.

Regions in clusters 1 and 4 that were shared by naive and memory cells were enriched for HMG motifs and also overlapped TCF1 ChIP-seq peaks from naive CD8⁺ T cells with high frequency (Figures 5C,D) (Steinke et al., 2014). Gene expression data indicated that among the HMG-containing TFs, the expression of Lef1 and Tcf7 was biased towards naive and memory cells (Figure 5E). Altogether, these data agree with the well-characterized contribution of these two TFs to the function of naive and memory CD8⁺ T cells (Steinke et al., 2014; Zhou et al., 2010), and suggested that their contribution is related to binding at these differentially-accessible regions. The same clusters, along with clusters 9 and 10 which also shared naive- and memory-biased accessibility, also overlapped E2A ChIP-seq peaks from mouse thymocytes (Leong, et al. 2016).

Another major group including clusters 2,5,6,7, and 11 had higher signal in effector, memory, and exhausted cells than naive (Figure 5B). These groups of accessible regions were enriched for bZIP and T-box motifs compared to all differentially-accessible regions in CD8⁺ T cells (Figure 5C). A substantial fraction of these peaks overlapped ChIP-seq peaks identified for BATF, IRF4, and T-bet from effector CD8⁺ T cells (Figure 5D), supporting the established role for these TFs in effector T cell differentiation (Dominguez et al., 2015; Intlekofer et al., 2005; Kurachi et al., 2014).

The regions with accessibility biased most towards exhausted cells (clusters 3 and 6) were enriched for consensus binding motifs characteristic of nuclear hormone receptor (NR) and rel homology domain (RHD) containing proteins (Figure 5C). Among NR-family TFs, we found that the expression of Nr4a2 was biased towards the exhausted cells compared to naive, effector and memory cells (Figure 5E,S5A). Among RHD-family TFs, we did not identify any genes that were primarily expressed in exhausted cells, including those encoding NFAT family members (Figure S5B), although Rel was reduced in exhausted cells compared to the other subsets (Figure 5E).

We observed similar patterns among all 2466 differentially expressed genes, with prominent clusters of naive- or exhausted-biased genes, as well as groups of genes with shared expression between naive and memory, effector and exhausted, or effector, memory and exhausted (Figure S5C). In an effort to associate the groups of differentially-accessible regions with global changes in gene expression, we compared the expression of genes within 25kb of peaks in each cluster from Figure 5B. The changes in accessibility were largely positively correlated with differences in gene expression, although there were examples of some peak-gene pairs with a negatively correlated association (Figure S5D). For example, some differentially-accessible peaks near the Tcf7 locus tended to have high signal in naive and memory cells, which was associated with high Tcf7 expression in these cell types (Figure S5D, clusters 4 and 9), but there were other peaks near Tcf7 with high signal in effector cells despite lower expression in these cells (Figure S5D, clusters 11 and 2). Despite these examples, there were many positive associations between changes in accessibility and activation-associated differentially-expressed genes; for instance, Ccl5, Tbx21, Gzma and Batf all showed accessibility and gene expression biased towards effector and exhausted cells (Figure S5D, clusters 2,5,6,7).

To associate accessibility with local epigenetic modifications, we compared the MeDIP-seq signal from naive cells and effector cells after LCMV Arm5 infection (Scharer et al., 2013) around these groups of differentially-accessible peaks; this technique assesses DNA methylation by immunoprecipitation with a methylcytosine-specific antibody. Most enhancers have been found to have moderate to low DNA methylation (Thurman et al., 2012) and around the center of all ATAC-seq peaks in both naive and effector cells we also found low average MeDIP-seq coverage compared to input (Figure 5F). Examining the MeDIP-seq signal around the clusters of differentially-accessible peaks, we identified several patterns. First, clusters 1,4,9 and 10, which had higher ATAC-seq signal in naive and memory T cells than in effector cells, were associated with low MeDIP-seq signal in both naive and effector cells, indicating that methylation is not strongly linked with reduced accessibility at these regions in effector cells. Similarly, cluster 2, which had higher accessibility in effector than naive, was also associated with low MeDIP-seq signal, suggesting that these regions had low methylation already in naive cells. Clusters 7 and 11 were the only clusters in which increased accessibility from naive to effector was associated with reduced MeDIP-seq signal in effector compared to naive. Finally, clusters 3 and 6, whose accessibility is biased towards exhausted, did not have substantial depletion of MeDIP-seq coverage compared to input; these regions may be demethylated only in exhausted cells, for which MeDIP-seq data are not yet available.

NFAT activity contributes directly and indirectly to the exhaustion profile

We previously showed that constitutive NFAT activity induces an exhaustion-like profile in T cells (Martinez et al., 2015). To assess how NFAT activity influences chromatin accessibility and the exhaustion phenotype, we expressed a constitutively active form of NFAT (NFAT-CA-RIT) with three amino acid substitutions (R468, I469 and T535) that disrupt the interaction between NFAT and AP-1 (Figure 1A). For comparison, cells were transduced with a version of NFAT-CA-RIT unable to bind DNA (NFAT-CA-RIT-DBDmut) or left untransduced (Mock). More than 4000 regions were differentially-accessible between mock-transduced cells and cells expressing NFAT-CA-RIT (Figure 6A); in contrast, expression of NFAT-CA-RIT-DBDmut did not change accessibility substantially (Figure S6A). While regions with increased accessibility in NFAT-CA-RIT-expressing cells were also more accessible in effector, memory, and exhausted than in naive cells, the greatest increase in accessibility was observed in exhausted cells. Conversely, regions with reduced accessibility in NFAT-CA-RIT-expressing cells were least accessible in exhausted cells (Figure 6B). Moreover, regions with higher accessibility in exhausted than in effector or memory cells had higher ATAC-seq signal in cells expressing NFAT-CA-RIT compared to mock-transduced cells (Figure S6B).

Figure 6. Constitutively active NFAT partially recapitulates the chronic activation profile in vitro.

A) Scatterplot of ATAC-seq counts per peak comparing in vitro cultured CD8⁺ T cells after transduction with retroviruses expressing the NFAT-CA-RIT mutant or left untransduced (Mock). B) Boxplots of ATAC-seq counts per peak in naive, effector, memory, and exhausted CD8⁺ T cells at common or differentially-accessible regions between Mock and NFAT-CA-RIT mutant expressing cells. Box indicates interquartile range with whiskers +/−1.5 times this range and outlier points. C) Scatter plot of NFAT-CA-RIT ChIP-seq coverage with log₂ fold-change ATAC-seq signal between Mock and NFAT-CA-RIT mutant expressing cells at regions with lower (top) or higher (bottom) ATAC-seq signal in exhausted compared to effector and memory CD8⁺ T cells. D) Mean ATAC-seq and NFAT ChIP-seq coverage at the Pdcd1 locus with a scale of 0-1200 for ATAC-seq tracks. E) Nr4 family member gene expression in CD8⁺ T cells over-expressing the NFAT-CA-RIT mutant or left untransduced (Mock) showing mean plus range. ATAC-seq data in A–D are from at least 2 independent replicates. See also Figure S6.

Focusing on regions with altered accessibility in exhausted cells in vivo, we asked whether they showed higher enrichment for NFAT-CA-RIT binding based on ChIP-seq analysis. Of the 993 regions highly accessible in exhausted cells (compared to both effector and memory, Figure 4H), most showed increased accessibility in NFAT-CA-RIT-expressing compared to mock-transduced cells (Fig. 6C, bottom, data points shifted to right of the vertical line), many of which had high NFAT-CA-RIT binding based on ChIP-seq (Figure 6C, bottom; data points shifted above the horizontal line). Conversely, of the 762 regions with low accessibility in exhausted cells (compared to both effector and memory, Figure 4I), most showed reduced accessibility in NFAT-CA-RIT-expressing compared to mock-transduced cells (Figure 6C, top, data points shifted to left of the vertical line) and fewer of these sites showed high ChIP-seq signal for NFAT-CA-RIT (Figure 6C, top, data points shifted above the horizontal line). Together these results extend our prior finding that NFAT activates the program of CD8⁺ T cell exhaustion and indicate that continuous NFAT activity contributes to the exhaustion profile during chronic infection in vivo (Martinez et al., 2015).

In contrast, brief restimulation of CD8⁺ T cells activated and cultured in vitro only partially recapitulated the effect of NFAT-CA-RIT in inducing increased accessibility at exhaustion-related regions. The combination of phorbol myristate acetate (PMA) and Ionomycin (Iono), or Iono alone, induced rapid (within 2 hr) changes in accessibility at ~3600 and ~700 regions, respectively (Figure S6C,D). These changes were partially blocked by addition of the calcineurin inhibitor cyclosporin (CsA), suggesting NFAT involvement. Unlike changes induced by chronic expression of NFAT-CA-RIT (Fig. 6B, right panels), these accessible regions induced by acute stimulation had similar signals in effector, memory, and exhausted cells (Figure S6C,D, right panels). The changes in accessibility observed at regions with high or low accessibility in exhausted cells (Figures 4H,4I) were more apparent in NFAT-CA-RIT-expressing cells than in PMA-Iono-stimulated cells (Figure S6E).

The Pdcd1 locus illustrates several aspects of the contribution of NFAT to changes in accessibility in exhausted-phenotype cells (Figure 6D). Three accessible regions in the Pdcd1 locus had higher signal in exhausted compared to naive, memory or effector cells: two of these were located 1.1 and 2.7 kb 5' of the Pdcd1 TSS whereas the third was located 22 kb 5' of the TSS (Figure 6D). All three peaks were observed in NFAT-CA-RIT-expressing cells, but only the two peaks proximal to the TSS were induced by stimulation with PMA and ionomycin and bound both NFAT and NFAT-CA-RIT. In contrast, the accessible region at +22 kb showed no NFAT binding and minimal change in accessibility after brief stimulation in vitro. Taken together, these data suggest that NFAT activity can directly and indirectly influence the chromatin accessibility profile associated with T cell exhaustion.

We compared the consensus TF binding motifs enriched in differentially-accessible regions after NFAT-CA-RIT expression or after restimulation of in vitro-cultured cells. Accessible regions associated with NFAT-CA-RIT expression and stimulation with either Iono alone or PMA and Iono together were enriched, as expected, in NFAT and NFAT-AP1 motifs (Figure S6H). In contrast, the NR-family binding motif that was found in exhaustion-biased regions, including the peak located 22 kb 5' of the TSS in the Pdcd1 locus (Figure 6D), was only enriched in accessible regions induced by NFAT-CA-RIT (Figure S6H). Activation of the NR family under these conditions may be a direct effect of NFAT activity, as NFAT-CA-RIT-expressing cells showed increased expression of Nr4a2 and Nr4a3 (Figure 6E) and NFAT-CA-RIT-occupied accessible regions within the Nr4a2 locus (Figure S6I). Altogether, these data indicated that prolonged stimulation during chronic infection was associated with changes in accessibility that were directly and indirectly induced by NFAT activity.

Discussion

In this study, we compared chromatin accessibility in CD8⁺ T cells under different conditions of stimulation and differentiation: in effector, memory and exhausted cells arising in response to in vivo viral infection; in cells pre-activated in culture and assessed before or after brief restimulation in vitro; and in cells subjected to retroviral transduction with constitutively-active versions of NFAT. We identified more than 70,000 regions that were accessible in at least one subset of activated or differentiated cells. At least half of these accessible regions had similar signals in most cell types examined, demonstrating that there is a core pattern of chromatin accessibility that is relatively stable in CD8⁺ T cells regardless of their activation or differentiation state. Many commonly-accessible regions were located near the TSS, but most were located within introns or at intergenic regions, and could potentially function as enhancers (Consortium, 2012). While these regions remain similarly accessible among CD8⁺ T cell subsets, their function may be regulated by subsequent DNA or histone modifications on flanking nucleosomes (Kouzarides, 2007). Indeed, variable enrichment of the trimethylation of H3K4 and H3K27 at promoters has been associated with differential transcription in naive, effector, and memory CD8⁺ T cells after influenza infection (Russ et al., 2014).

At least 30% of all accessible regions were differentially-accessible between any pair of cell types. Most of these regions were located far from TSSs, but were linked to differentially-expressed genes and may represent enhancers that control CD8⁺ T cell differentiation and function. Among the populations we have compared here, the largest number of differentially-accessible regions were observed between naive and effector T cells: approximately 25% of all accessible regions were differentially-accessible between these two cell types.

Regions that were differentially-accessible between naive, effector, and memory CD8⁺ T cells were enriched for binding sites of several TFs known to contribute to the state of the relevant cell types. Binding of TCF family members, in particular TCF1, was substantially enriched at regions accessible in naive and memory CD8⁺ T cells. Conversely, binding sites for bZIP, IRF, and T-box TFs were enriched within accessible regions specific to effector and memory CD8⁺ T cells.

Our analyses did not identify a substantial number of accessible regions specific to memory CD8⁺ T cells, which were characterized by a mixture of the accessibility profiles observed in effector and naive cells. Many peaks induced by viral responses at day 8 remained accessible at day 35 in memory cells, providing a potential mechanism for their more rapid responses to antigen compared to naive cells (Williams and Bevan, 2007). However when compared to effector T cells, memory CD8⁺ T cells also shared certain aspects of their accessibility profile with naive T cells, accompanied by a corresponding similarity in gene expression, as described previously (Best et al., 2013). Comparison of MPEC and SLEC subsets from day 8 after acute infection revealed that MPEC are substantially similar to SLEC and total effector populations, with a slight bias towards the memory accessibility profile. Our data do not address whether the shift from MPEC to memory cells reflects the selective survival of cells with a memory-like chromatin accessibility profile after viral infection, or whether cells that originally had an effector-like pattern of chromatin accessibility revert to accessibility profiles that more closely resemble those of memory cells arising later in the infection.

Based on comparisons to DNA methylation data from naive and effector T cells after LCMV infection (Scharer, et al., 2013), we found low enrichment of DNA methylation around accessible regions in any given cell type, as generally observed in other cells (Thurman, et al., 2012). Among regions that were accessible in effector cells after activation, some had a corresponding reduction in DNA methylation from naive to effector, whereas other regions were already demethylated in naive cells. Together, these data support the hypothesis that concurrent regulation of DNA methylation and chromatin accessibility controls the activity of regulatory elements in activated T cells, to alter gene expression and influence differentiation.

In a chronic infection, virus is not completely cleared and antigen-specific cells are continually subjected to chronic stimulation at the site of infection and they acquire a hyporesponsive or “exhausted” state (Speiser et al., 2014; Wherry and Kurachi, 2015). Based on the chromatin accessibility described here, exhausted cells were most similar to effector cells: for instance both functional subsets had similar accessibility profiles at the Ifng locus, which is strongly activated in effector cells but poorly induced upon antigen stimulation of exhausted CD8⁺ T cells (Barber et al., 2006; Kao et al., 2011). These observations may explain the effectiveness of antibody therapies targeting PD-1, PD-L1, and CTLA4 where blockade of the inhibitory signals transmitted by these receptors may enable exhausted T cells to reactivate enhancers that remained as accessible as in effector cells (Barber et al., 2006; Pardoll, 2012). Regions with similar accessibility between effector and exhausted cells were enriched for T-box motifs and often bound by T-bet in effector cells, consistent with the suggestion that variable expression of T-bet and EOMES among exhausted CD8⁺ T cells could alter cell function in cells that are similar in their overall accessibility profiles (Doering et al., 2012; Intlekofer et al., 2005; Wherry and Kurachi, 2015).

Regions that were uniquely accessible in exhausted cells were enriched for two motifs (NFAT and Nr4 family) that were not substantially enriched in any other viral-specific T cell subsets. NFAT can directly induce both Nr4a2 and Nr4a3, while only Nr4a2 was induced in exhausted T cells (this study and (Martinez et al., 2015)). Nr4a family members have been associated with the development and function of regulatory T cells (Treg) (Sekiya et al., 2013); Treg cells also fail to express cytokines characteristic of activated cells following antigen receptor stimulation, and thus Nr4a proteins may mediate similar cell-intrinsic inhibitory programs in Tregs and exhausted CD8⁺ cells. We hypothesize that the uniquely exhaustion-specific regions may be induced by continued antigen stimulation during chronic infection, as their accessibility profiles overlap substantially with the profiles of cells activated in vitro and cells expressing constitutively-active NFAT, even though only some of the accessible regions bound NFAT in ChIP-seq experiments. Some regions (for instance in the Satb1 and Il7r loci) were less accessibility in exhausted cells than in effector and memory cell types, suggesting that at least some aspect of T cell-specific chromatin accessibility is lost in exhausted cells. Loss of chromatin accessibility relative to effector and memory T cells may explain the reported incomplete restoration of function in exhausted cells after checkpoint blockade or after transfer to mice in the absence of antigen (Barber et al., 2006; Schietinger et al., 2012; Utzschneider et al., 2013). Additionally, exhaustion-biased accessible regions showed higher enrichment for DNA methylation in both naive and effector cells, compared to sites accessible in naive and effector cells. These sites may be demethylated specifically in exhausted cells, as shown for two conserved regions in the Pdcd1 locus during chronic infection in mice and humans (Youngblood et al., 2011). Thus DNA methylation is an additional factor that may govern the difference between effector and exhausted phenotypes.

The changes in chromatin accessibility observed after activation of naive CD8⁺ T cells, whether in vitro or in vivo, form a continuum reflecting the diverse ways in which TFs expressed in different cellular subsets can cooperate to influence gene expression and cellular function. Based on their patterns of chromatin accessibility, in vitro-activated cells, which have been exposed to stimulatory antibodies with limited cytokine exposure and without antigen-presenting cells, are most similar to naive cells that have never encountered antigen, while memory, effector, and exhausted cells, which have been exposed to diverse stimulation in a native environment, are progressively less similar. Future experiments will determine whether continuing antigen stimulation or inflammation induces stable changes in chromatin accessibility across an entire population, or whether stochastic changes in accessibility and TF binding and expression give rise to a heterogeneous cell population from which cells with particular phenotypes survive.

Experimental Procedures

Mice

CD8⁺ T cells were purified from splenocytes of 6–10 week old C57BL/6J mice after infection with LCMV or left uninfected. Effector and memory CD8⁺ T cell subsets were collected after infection with LCMV Arm5, based on staining with tetramer and antibodies to CD44, IL7Rα, and KLRG1 (Figure S1A,C,D). Exhausted cells were collected after infection LCMV clone 13 based on staining with tetramer and antibodies to PD-1 and Tim3 (Figure S1E).

For in vitro stimulation, CD8⁺ T cells were purified from P14 T cell receptor transgenic mice on a TCRα constant region deficient background, before stimulation with CD3ε and CD28 specific antibodies for two days and then expanded 10U/ml or 100U/ml IL-2 (Pipkin et al., 2010). On day 6, cells were restimulated for 2hrs under various conditions (Supplemental Experimental Procedures). For retroviral transduction experiments, CD8⁺ T cells were isolated by negative selection from C57BL/6J mice. These cells were stimulated for one day with antibodies specific for CD3ε and CD28 before infection with retroviruses followed by expansion for three days in 10U/ml IL-2.

All experiments were performed according to protocols approved by the La Jolla Institute animal care and use committee.

ATAC-seq

ATAC-seq library preparations were performed as described (Buenrostro et al., 2013). Cells were washed with PBS before treatment with lysis buffer followed by labeling with Nextera enzyme (Illumina, San Diego, CA), before PCR amplification with 10–12 cycles with barcoded primers and 2×50 cycle paired-end sequencing (Illumina, San Diego, CA). Reads were mapped to mouse genome (mm9) using bowtie (version 1.0. 0, (Langmead et al., 2009)). Unmapped reads were processed with trim_galore, re-mapped with bowtie and merged with previous mapping output. Duplicate reads identified by picard MarkDuplicates and reads mapping to chrM were excluded. Peak summits were identified with MACS2 (Zhang et al., 2008) from individual replicates using short DNA fragments (<100bp) and expanded to 500 bp regions. We merged overlapping regions from all replicates and excluded regions on chrY or those that intersected ENCODE blacklisted regions. The number of transposase insertions within each region was computed for each replicate and differentially-accessible regions were identified with DESeq2 (version 1.6) based on an fdr adjusted p value of less than 1×10⁻³ and an estimated fold change of at least 3. Full details are available in Supplemental Experimental Methods.

E2A ((Leong et al., 2016), GSE84974), TCF1 (Steinke et al., 2014), GSE52070), T-bet (Dominguez et al., 2015), SRR2075567-8,SRR2075585), BATF and IRF4 (Kurachi et al., 2014), GSE54191) ChIP-seq and naive and effector MeDIP-seq ((Scharer et al., 2013), GSE44638), including input controls where available, were downloaded from NCBI SRA database and mapped to the mm9 genome (bowtie version 1.0.0). ChIP-seq peaks were called with MACS2.

RNA-seq

SMARTseq2 RNA-seq libraries were prepared from naive, effector, memory, and exhausted cells (Picelli et al., 2014), sequenced with an Illumina HiSeq2500 and mapped to the mm9 genome (TopHat (1.4.1), Trapnell et al., 2009). Low quality reads were excluded and feature counts were computed with htseq-count (0.6.0 using the “union” option). Differentially expressed genes were identified with DESeq2 with an fdr adjusted p-value of less than 1×10⁻³ and an estimated fold-change of at least 2 and a mean of least 10 transcripts per million transcripts. Full details are available in Supplemental Experimental Methods.