Transcriptome dynamics of CD4⁺ T cells during malaria maps gradual transit from effector to memory

Abstract

The dynamics of CD4⁺ T cell memory development remain to be examined at genome scale. In malaria-endemic regions, antimalarial chemoprevention protects long after its cessation and associates with effects on CD4⁺ T cells. We applied single-cell RNA sequencing and computational modelling to track memory development during Plasmodium infection and treatment. In the absence of central memory precursors, two trajectories developed as T helper 1 (T_H1) and follicular helper T (T_FH) transcriptomes contracted and partially coalesced over three weeks. Progeny of single clones populated T_H1 and T_FH trajectories, and fate-mapping suggested that there was minimal lineage plasticity. Relationships between T_FH and central memory were revealed, with antimalarials modulating these responses and boosting T_H1 recall. Finally, single-cell epigenomics confirmed that heterogeneity among effectors was partially reset in memory. Thus, the effector-to-memory transition in CD4⁺ T cells is gradual during malaria and is modulated by antiparasitic drugs. Graphical user interfaces are presented for examining gene-expression dynamics and gene–gene correlations (http://haquelab.mdhs.unimelb.edu.au/cd4_memory/).

Main

CD4⁺ TCRαβ⁺ T cells confer immunity to bacteria, viruses, fungi and protozoan parasites. They are repositories of immunological memory, facilitating secondary responses both faster than and of larger magnitude than primary infection. CD4⁺ T cell memory has been studied using major histocompatibility complex (MHC) tetramers and T cell antigen receptor (TCR) transgenic T cells, mostly in bacterial and viral infections^1,2,3,4,5, and less so in parasitic infection⁶. Previous studies have employed preselected markers, including fate-mapping of cytokine or chemokine genes^1,2, adoptive transfer of cell-sorted effectors^3,4 and single-cell transfer and longitudinal sampling⁵. These indicate that a single naive CD4⁺ T cell generates memory clones that partly resemble their effector counterparts. This can be explained by a linear model wherein naive cells necessarily give rise to effector cells before transit to memory, or a branching model, wherein memory cells develop before or in parallel with effectors. Single-cell RNA sequencing (scRNA-seq) has recently revealed CD4⁺ central memory precursors during the first week of viral infection⁷, consistent with a branching model. In humans, longitudinal assessment after vaccination has indicated that effectors gave rise directly to circulating memory cells^8,9. Given the class of infectious agent can influence the response of a CD4⁺ T cell clone¹⁰, it remains difficult to extrapolate from previous studies to CD4⁺ T cells in a parasitic infection.

Malaria remains a threat to human health, with 228 million cases and 405,000 deaths in 2018 (ref. ¹¹). Although non-sterilizing immunity can be achieved via natural infection, this takes multiple exposures over many months to achieve¹². In endemic regions, antimalarial chemoprevention can improve protection for at least one year after cessation of drug treatment^13,14. Mechanisms explaining this phenomenon remain unresolved, although qualitative changes in memory CD4⁺ T cells have been reported¹⁵. In mice, immune-checkpoint blockade reduced CD4⁺ T cell exhaustion and improved immunity¹⁶, suggesting that parasitic infection impairs, but does not delete, CD4-dependent cellular immunity. Defining molecular pathways that promote effective development of CD4⁺ T cell memory may reveal new strategies for improving immunity to malaria, either during natural exposure and chemoprevention or via vaccination.

Although CD4⁺ T cells protect against Plasmodium parasites¹⁷, genome-scale characterization of the dynamics of memory development remains to be performed in malaria or other infection models. We previously employed TCR transgenic CD4⁺ T cells specific for a Plasmodium epitope (PbTII cells), in conjunction with scRNA-seq and computational modeling, to reveal mechanisms underlying T_H1 and T_FH fate bifurcation during the first week of experimental malaria¹⁸. From hundreds of single-cell transcriptomes, T_H1 and T_FH effectors emerged from a proliferating, intermediate state, the balance of which was externally influenced by monocytes or B cells. Here, we extend our previous work by studying thousands of parasite-specific CD4⁺ T cells over four weeks in the spleen using two complementary scRNA-seq platforms, single-cell assay for transposase-accessible chromatin using sequencing (scATAC-seq), computational modelling and in vivo validation. We define, at genome scale, the dynamics underlying CD4⁺ T cell memory development during experimental malaria and drug treatment, and provide user-friendly online resources to facilitate research in T cell biology.

Results

PbTIIs exhibit memory phenotypes boosted by anti-malarial drugs

We transferred naive eGFP⁺ Plasmodium-specific PbTIIs¹⁹ into C57BL/6J mice—seeding the spleen of each mouse with 100–500 cells—and infected them with blood-stage Plasmodium chabaudi chabaudi AS (PcAS) parasites¹⁷. Parasitemia peaked by day 7, followed by low-level parasitemia (Fig. 1a) and splenomegaly (Fig. 1b) for 2–3 months. We administered sodium artesunate or control saline at multiple time-points from day 7, referred to as intermittent antimalarial drug treatment (IAT).

Fig. 1: Plasmodium-specific TCR transgenic PbTII cells develop memory during infection.

a, Parasitemia during PcAS infection in C57BL/6J mice in the presence or absence (saline-treated control group) of IAT. Data are pooled from 3 independent experiments (n = 5 mice per group, per independent experiment). Statistical testing was performed using two-tailed Mann–Whitney test. Days p.i., days postinfection; pRBC, parasitized red blood cells. b, Spleen weight during PcAS infection in the presence or absence of IAT. Data are representative of 3 independent experiments (n = 5 mice group, per individual timepoint). Statistical testing was performed using two-tailed Mann–Whitney test. c, Splenic PbTII cell numbers during PcAS infection in the presence or absence of IAT. Data are pooled from 2 independent experiments (n = 5 mice per group, per individual timepoint for each independent experiment). Statistical testing was performed using two-tailed Mann–Whitney test. d, Histograms of flow-cytometric assessment for surface CD62L and CCR7 expression on PbTII cells. A representative histogram is presented for naive controls, and overlaid histograms (n = 5 mice per group) are presented for day 28 (D28) saline and IAT groups. Data are representative of two independent experiments. e, Left, analysis of direct ex vivo IFN-γ production by PbTII cells without restimulation. Right, analysis of ex vivo IFN-γ production by PbTII cells after PMA and ionomycin restimulation in vitro. The dashed gray line represents the threshold of IFN-γ production by naive PbTII cells. Data are pooled from 2 independent experiments (n = 5 mice per group, per individual timepoint for each independent experiment). Statistical analysis performed using two-tailed Mann–Whitney test. f, Representative fluorescence-activated cell sorting (FACS) plots showing direct ex vivo production of IFN-γ at 17 hours following rechallenge for naive (gray) or memory (green) PbTII cells in saline and IAT groups. The graph compares IFN-γ production between naive or memory PbTII cells in saline or IAT groups. Data are pooled from 2 independent experiments (n = 5 mice per group, per independent experiment). Statistical analysis was performed using paired two-way analysis of variance with Tukey’s multiple-comparisons test. *P < 0.05, **P < 0.01, ***P < 0.001, ****P < 0.0001. Data are presented as mean ± s.e.m. (a,b,c,e). Statistical analysis was performed between saline and IAT groups for each timepoint (a,b,c,e).

PcAS infection triggers CD4⁺ T cell responses that largely comprise T_H1 and T_FH cells^20,21 (definitions for CD4⁺ T cell states are provided in Table 1). Splenic PbTIIs peaked by day 7, remaining detectable well beyond the first month, although IAT reduced these and splenomegaly (Fig. 1b,c and Supplementary Information Fig. 1a). PbTIIs expressed CD62L and CCR7 heterogeneously during the first month, which was modulated by IAT (Fig. 1d), suggesting central memory (T_CM) and effector memory (T_EM) PbTIIs had developed. Bifurcated expression of T_H1-associated C-X-C motif chemokine receptor 6 (CXCR6) and T_FH-associated CXCR5 was partially retained, and was enhanced by IAT at day 60. (Extended Data Fig. 1a). PbTIIs were located in splenic T cell, B cell and red pulp zones at day 7 (Extended Data Fig. 1b), and in GCs and B cell zones at later times (Extended Data Fig. 1c,d). Thus, PbTIIs entered GCs and persisted long enough to become memory cells.

Table 1 T cell terminology

We next determined whether PbTIIs exhibited recall responses. Direct ex vivo interferon-γ (IFN-γ) production progressively subsided in both groups by day 28, consistent with waning of primary T_H1 responses (Fig. 1e). In contrast, in vitro T_H1 recall was substantially enhanced on a per-cell basis by IAT (Fig. 1e), despite reductions in PbTII numbers (Fig. 1c), and those capable of expressing IFN-γ or interleukin-10 (IL-10) (Extended Data Fig. 2a,b). The potential for PbTIIs to mount T_H1 and T_FH recall responses was examined by bulk ATAC-seq. Inaccessible chromatin in naive PbTIIs became accessible at day 7, and remained partially accessible at day 28 with or without IAT (Extended Data Fig. 2c). This was evident for specific T_H1 or T_FH-associated genes Ifng, Cxcr5, T-box transcription factor 21 (Tbx21) and Il21 (Extended Data Fig. 2d). To test recall in vivo, we performed homologous high-dose rechallenge, comparing day 28, antigen-experienced PbTIIs, with naive comparators transferred before rechallenge (Supplementary Fig. 1b). PbTIIs primed 28 days previously exhibited greater direct ex vivo IFN-γ production than that of naive PbTIIs (Fig. 1f), with reduced proliferative and early-activation markers (Extended Data Fig. 2e,f). Thus, PbTIIs had acquired memory phenotypes, with IAT boosting T_H1 memory (Fig. 1f).

Central memory precursors are not essential for CD4⁺ T cell memory development

We next examined the mechanisms underlying memory development. Despite reporting only two fates in our previous study¹⁸, we hypothesized that rare memory precursors existed. To test this, we assessed thousands of splenic effector transcriptomes. PbTIIs at day 7 (10,251 cells following quality control) had bifurcated, exhibiting reported T_H1 and T_FH gene signatures¹⁸, and a smaller group of proliferating cells expressing T_H1 or T_FH features (Fig. 2a–c). Consistent with our previous study, the minor cluster was likely undergoing final cell division and fate bifurcation¹⁸. Most importantly, neither a recently published CD4⁺ T_CM precursor gene signature⁷ (Fig. 2d) nor CCR7 expression (Fig. 2e) supported emergence of CD4⁺ T_CM precursors at peak infection.

Fig. 2: A lack of PbTII central memory precursors at peak infection.

a, Droplet-based scRNA-seq analysis of 10,251 PbTII cells at peak effector stage (D7 postinfection) during PcAS infection using the 10x Genomics Chromium platform. Unsupervised clustering was performed on a UMAP representation of PbTII cells at D7 p.i., calculated using principal components (PC) 1–10 with 1,522 highly variable genes as input. b, UMAP visualization showing a T_H1, T_FH and cell-cycle signature score and genes (Cxcr6, Cxcr5 and Mki67) associated with each signature on PbTII cells at D7 p.i. c, Violin plots of T_H1, T_FH and cell-cycle signature scores and genes (Cxcr6, Cxcr5 and Mki67) associated with each signature on PbTII cells at D7 p.i. The expression value for naive (D0) PbTII cells is shown for each signature and gene for reference. Median expression for each cluster is denoted for gene signature scores as a horizontal line. d, Heatmap with unsupervised hierarchical clustering showing the expression of T_H1, T_FH and T central memory precursor (T_CMP) gene signatures as defined by Ciucci et al.⁷ for PbTII cells at D7 p.i. e, Representative FACS plot of PbTII cells at D7 p.i. showing the surface expression of CXCR6 and CXCR5. PbTII cells were put in subsets as follows: CXCR6⁺CXCR5^– (red), CXCR6⁺CXCR5⁺ (purple), CXCR6^–CXCR5⁺ (blue) and CXCR6^–CXCR5^– (orange). Representative histograms showing surface CCR7 expression for the different subsets of PbTII cells at D7 p.i. and naive PbTII cells (gray). Data are representative of 2 independent experiments (n = 6 mice per group, per independent experiment).

T_H1 and T_FH effectors exhibit gradual transit towards memory

We next hypothesized that memory developed directly from effectors. However, tracking this process using single-gene approaches was difficult since T_H1 and T_FH phenotypic mixing is reported in Plasmodium infection²². scRNA-seq revealed that neither Ifng, Tbx21 nor Cxcr3 expression was confined to T_H1 cells (Extended Data Fig. 3). Similarly, selectin p ligand (Selplg) (encoding the adhesion molecule PSGL1) and lymphocyte antigen 6 complex, locus C2 (Ly6c2) (encoding Ly6C) were expressed in both fates at peak (Extended Data Fig. 3). Cxcr6 and Cxcr5 were reasonably well-confined to their respective fates (Fig. 2b), although no single gene faithfully marked all T_H1 or T_FH cells. Therefore, we applied scRNA-seq profiling over time to avoid pre-selected markers, to capture intermediate states and to examine dynamics at genome scale. As before, we transferred PbTIIs (Extended Data Fig. 4a), infected with PcAS, and administered IAT or saline from day 7 (Supplementary Information Fig. 1c). At various timepoints, splenic PbTIIs were sorted and processed via Smart-seq2 scRNA-seq (Extended Data Fig. 4b). From 4,548 wells, high-quality transcriptomes were obtained for 2,964 cells (Extended Data Fig. 4c). To study transcriptomes from naivety to memory, we integrated this Smart-seq2 dataset with our previous PbTII datasets¹⁸, resulting in 3,728 transcriptomes spanning 4 weeks of infection with and without IAT (Extended Data Fig. 4d). Principal-component analysis (PCA) and uniform manifold approximation and projection (UMAP) revealed, as before, 2 effector populations emerging by day 7, expressing Cxcr5 or Cxcr6 (Extended Data Fig. 5a). UMAP suggested that there was progressive transcriptomic change over time, with two trajectories being more apparent during IAT than during saline treatment (Extended Data Fig. 5a). Unsupervised trajectory inference using Slingshot²³ was unable to map trajectories that adhered to timepoint information (Extended Data Fig. 5a). Estimation of RNA velocity²⁴ indicated that, although change was rapid and trajectory inference was possible around fate bifurcation, subsequent timepoints exhibited slower rates of change, making inference of trajectory difficult (Extended Data Fig. 5b). In any case, our integrated scRNA-seq dataset supported the presence of intermediate transcriptomic states between effector and memory states, particularly under IAT, suggesting that there was a progressive effector-to-memory transition over several weeks.

Temporal mixture modelling of effector-to-memory transitions

We next generated a probabilistic transcriptomic model to map gene-expression dynamics with or without IAT. We employed GPfates¹⁸, which involved Bayesian Gaussian process latent variable modeling (bGPLVM), calculation of pseudotime from resulting latent variables, and finally overlapping mixtures of Gaussian processes (OMGP), run on IAT and control saline groups separately—including data from day 0–7 in both datasets (Fig. 3a). Resulting models resembled incomplete circles in two dimensions, with T_H1 and T_FH lineages coalescing and partially returning towards naivety (Fig. 3a). Our ability to assign effectors to either lineage diminished along pseudotime (Fig. 3b). This was more apparent for saline than for IAT (Fig. 3b). Also of note, the T_FH lineage tracked closer to naive cells (Fig. 3a).

Fig. 3: Temporal mixture modelling of scRNA-seq data maps effector-to-memory transit.

a, Reconstruction of trajectories along pseudotime using GPfates¹⁸. bGPLVM analysis was performed on the full time series from the combined batch-corrected PbTII datasets (first batch: Smart-seq2 (96), D0, D4 and D7 p.i.; second batch: SMARTer C1, D0–D7 p.i.; third batch: Smart-seq2 (384), D0–D28 p.i.) and 2D bGPLVM representations were plotted for each treatment group (saline, left; IAT, right). D0, D2, D3, D4 and D7 p.i. cells were replicated for each representation. OMGP analysis was performed individually for saline and IAT. All Smart-seq2 datasets mentioned from here onwards comprise the batch-corrected full time series containing all three experimental batches, unless otherwise specified. Saline and IAT groups comprise shared timepoints (D0–D7) and treatment-unique timepoints (D10–D28, saline or IAT). b, T_H1-assignment probability assigned using OMGP along pseudotime for PbTII cells for saline (left) and IAT (right) groups. BP, T_H1–T_FH bifurcation point. c, Histogram showing the number of genes expressed (left) and cell-cycle score (right) along pseudotime for the saline and IAT groups. d, Left, degree of exhaustion visualized on a 2D representation of bGPLVM. Right, violin plots showing the exhaustion score in naive cells (D0) and cells positioned at late pseudotime (pseudotime > 0.9) for saline and IAT groups (shown in the black box). For the box plot, the center line indicates the median, box limits indicate the upper and lower quartiles and the whiskers indicate the maximum and minimum measures. Statistical analysis was performed using two-sided Wilcoxon rank-sum test between saline and IAT groups. LV, latent variable. e, Left, example gene-expression dynamics for T_H1- and T_FH-associated genes interrogated on the IAT ‘Dynamic’ dataset from our GUI: http://haquelab.mdhs.unimelb.edu.au/cd4_memory/. Right, example of the optional gene–gene correlation function.

Gene numbers increased from ~2,000 at naivety to ~5,000 during clonal expansion, ~4,000 in effectors and ~2,000 by the end of pseudotime, a pattern mirrored by cell-cycling genes (Fig. 3c). IAT did not alter the decline in detected genes over pseudotime, but did reduce expression of exhaustion-associated genes (Fig. 3d). To validate our GPfates models, we employed single-cell variational inference²⁵, which recapitulated observations from PCA and GPfates, including that Cxcr5- and Cxcr6-expressing effectors appeared to gradually change over three weeks (Extended Data Fig. 5c). Thus, computational modeling suggested that effectors gradually transitioned to various quiescent cellular states that partially resembled naivety.

To facilitate interrogation of our GPfates models, we present an online graphical user interface (GUI) (Fig. 3e): http://haquelab.mdhs.unimelb.edu.au/cd4_memory/. The ‘Dynamic Model’ GUI enables visualization of expression dynamics for any gene in CD4⁺ T cells during persistent infection or IAT. We also present a gene–gene correlation function, allowing testing for coexpression of two genes in the same cell.

Transcriptome dynamics reveal immune signatures retained in memory

To discover genes associated with memory development, we categorized ~15,000 genes within our GPfates models according to expression dynamics along pseudotime (without consideration of T_H1–T_FH bifurcation). Using SpatialDE²⁶, seven distinct dynamics were identified (Fig. 4a,b). The shape of each was similar between IAT and saline groups, with minor differences, for example, in dynamic 1 (Fig. 4c). Gene Ontology (GO) analysis was conducted within each dynamic (Fig. 4b). Consistent with our previous study¹⁸, dynamic 6, associated with DNA replication, peaked before bifurcation and faded as effector phenotypes emerged. Dynamic 5 was more prolonged and was associated with energy metabolism, including ATPase and electron-carrier activity. Dynamic 7, featuring a late drop in gene expression, was overwhelmingly composed of ribosomal genes. Dynamic 1 was composed of relatively few genes (511 of 14,167 for IAT; 463 of 15,310 for saline) that peaked late and remained elevated at the end of pseudotime. GO analysis indicated that there were strong associations with immune processes, including cytokine signaling, TNF-receptor signaling, antigen binding and targeting by the immune-associated E3 ubiquitin ligase, TRAF6. Therefore, only 3–4% of genes exhibited a late peak and preserved expression, and were largely associated with immune processes.

Fig. 4: Transcriptome dynamics reveal gene signatures that were preserved during memory development.

a, Genome-wide expression patterns of 14,167 significantly variable genes along pseudotime from the IAT group from the combined batch-corrected Smart-seq2 dataset, grouped according to individual dynamics and represented as signature scores. T_H1–T_FH BP and peak effector phase (median of D7 p.i. pseudotime values) are depicted to provide temporal context. b, Summary of GO enrichment analysis of molecular function associated with genes for each dynamic (dynamics 1–7) as in a. The number and proportion of genes for each dynamic are shown. c, Expression of signature scores of genes in dynamic 1 between IAT and saline groups along pseudotime. d, Coexpression network analysis of genes in dynamic 1 for IAT (left) and saline (right) groups. Edge weight corresponds to Spearman’s rho values. e, Transcription factors in dynamic 1 for IAT and saline groups. Each transcription factor is scored for its T_H1–T_FH assignment probability. f, Gene-expression correlation with pseudotime (x axis) versus the correlation with T_H trend assignment (y axis) is shown (Pearson correlation). Genes were assessed for the T_H trend (T_H1–T_FH bifurcation) using GPfates (top 20 T_H1-associated, red; top 20 T_FH-associated, blue) for each treatment group. The top 20 pseudotime-correlated genes are enclosed within the gray dashed line window.

Correlation and gene-network analysis of dynamic 1 revealed stronger positive correlations between genes under IAT than under saline (Fig. 4d). Resulting networks suggested a linear ‘axis’ under IAT, and a single hub of weaker connections under saline (Fig. 4d,e). Several transcriptional regulators were distributed along the IAT axis with a graded T_H1–T_FH structure (Fig. 4d,e), with Bcl6 and Tox2 at one end, and T_H1-associated Id2, Tbx21 and Runx2 at the other (Fig. 4e). Foci for immune checkpoints (Tigit, Lag3, Pdcd1, Ctla4) and Tr1 (Il10, Prdm1, Bhlhe40 (refs. ^27,28)) were also noted. Thus, similar genes were detected in dynamic 1 for IAT and saline, but gene–gene correlations were stronger and T_H1, Tr1, exhaustion and T_FH groupings were clearer in IAT.

The existence of T_H1 and T_FH networks under IAT motivated examination of genes according to the strength of T_H1–T_FH bifurcation. All genes were scored for pseudotime correlation and bifurcation to either lineage (Fig. 4f). Some genes exhibited moderate pseudotime correlation and strong T_H1 or T_FH assignment, including Cxcr6, Ccr2, Ccr5, Tox2, Cxcr5 and Bcl6. Among strongly T_H1-bifurcating genes, some correlated strongly with pseudotime, including Nkg7, S100a4, S100a6 and Ccl5 (Fig. 4f). In contrast, only Folr4 (encoding folate receptor 4) was T_FH bifurcating and highly correlated with pseudotime. Instead, many T_FH-associated genes correlated with early pseudotime, including Tcf7, highlighting their expression in naivety. Finally, some genes correlated with late pseudotime with little bifurcation, including transcription factors Id2 and Maf, exhaustion-related genes Tigit and Lag3 (in saline controls more so than under IAT) and Cxcr3.

To test predictions from our models (Extended Data Fig. 6a), we examined protein expression of various markers, including distribution among lineages. ID2 was strongly upregulated at day 7, but was not substantially retained at the protein level by day 28 (Extended Data Fig. 6b). Reciprocally related TCF1 (encoded by Tcf7), was heterogeneously downregulated in effectors, and partly recovered in some memory cells during IAT (Extended Data Fig. 6b). c-Maf was upregulated at day 7, and partially retained at day 28 with or without IAT (Extended Data Fig. 6c). CCL5 was absent from CXCR5⁺ effector and memory PbTIIs, and was associated instead with CXCR6⁺ memory PbTIIs (Extended Data Fig. 6d). CXCR3 was broadly expressed on CXCR5⁺ and CXCR6⁺ cells at day 28 under IAT (Extended Data Fig. 6e).

Id3, a transcription factor related to ID2, appeared to be confined to the T_FH-associated lineage (Extended Data Fig. 6a). We confirmed this using Id3^GFP PbTIIs, with GFP expression in CXCR5⁺ but not CXCR6⁺ PbTIIs at day 28 (Extended Data Fig. 6f). Thus, Id3 and Tcf7 were retained along the T_FH lineage. To confirm Tcf7 promoted the T_FH lineage, we examined T_H1–T_FH fate in Cre^ERT2Tcf^fl/fl PbTIIs, compared with those from wild-type littermates (Extended Data Fig. 6g). TCF1-deficient PbTII effectors were defective in BCL6 and CXCR5 expression, and instead expressed more CXCR6, ID2 and T-bet (Extended Data Fig. 6h), confirming TCF1 promoted the T_FH lineage in our model.

Memory fate is determined during the first week of infection

To examine memory fate in single naive PbTIIs, we employed the diverse endogenous TCRαβ sequences in Rag-sufficient PbTIIs as barcodes¹⁸. From 2,964 transcriptomes across all timepoints and conditions, where endogenous VDJ regions were reconstructed²⁹, we detected 201 families (Fig. 5a,b), defined as sharing the same endogenous sequences. Families ranged from two to five cells (Fig. 5b), and did not cross timepoints or conditions (Extended Data Fig. 7a), because individual mice were used for each, and PbTIIs in separate mice harbor unique endogenous barcodes. Cells from one family frequently occupied both T_H1 and T_FH lineages (Fig. 5c), confirming single naive PbTIIs displayed heterogeneity in the effector and memory fates of their progeny.

Fig. 5: Memory fate of naive clones is set during the first week of infection.

a, Visualization of families sharing endogenous TCRα and TCRβ sequences on bGPLVM representation, with each family connected by green lines. Cells are colored by the level of Cxcr6 expression. b, Bar graph showing the frequency of families for each family size. c, Visualization of families on a T_H1 assignment probability along pseudotime. d, Bar graph showing the observed number of families (solid green bar) that are (from left to right): T_FH predominant, mixture of T_FH and T_H1, and T_H1 predominant. The dotted line bars represent the expected number of families by random process as described in the Methods. e–g, Representative FACS plots and graphs showing the proportion or geometric mean fluorescence intensity (gMFI) of splenic PbTII cells at D28 p.i. expressing CXCR6 and CCL5 (e), Id2 and TCF1 (f) and IFN-γ (g) for the reference and CXCR5⁺ transfer groups. Data are pooled from 2 independent experiments (first experiment: n = 6 mice per group; second experiment: n = 6 mice in the reference group and n = 5 mice in the CXCR5⁺ transfer group). Staining for TCF1 and IFNγ was performed after stimulation with PMA and ionomycin. Statistical analysis was performed using the two-tailed Mann–Whitney test. h, FACS plots and graph showing the proportion of PbTII cells in the liver expressing CXCR6 or CXCR5 at D28 p.i. A representative plot is shown for the reference control, while pooled data are shown for the CXCR6⁺ transfer. Data are representative of 2 independent experiments (n = 6 mice in reference group and n = 5 mice in the CXCR6⁺ transfer group for each independent experiment). Statistical analysis was performed using two-tailed Mann–Whitney test. Data are presented as mean ± s.e.m. (e–h). *P < 0.05, **P < 0.01, ****P < 0.0001.

We tested whether lineage choice was random within a family. We examined cells under IAT only, where binary T_H1–T_FH fates were more easily discerned than in saline controls (Fig. 5c). After excluding cells of indeterminate fate, we examined 97 families. We hypothesized cell fate was independent of family, and therefore distribution of fates within each family would follow a random, binomial distribution (Fig. 5d). Instead, sibling PbTIIs exhibited greater tendencies towards the same fate than could be explained by random binomial process (P < 0.0005, Fig. 5d), the trend being evident also during examination of smaller families, albeit with reduced statistical significance owing to smaller sample sizes (P = 0.054 for 67 families of 2; P = 0.052 for 19 families of 3). Thus, T_H1–T_FH lineage fate was not entirely random within a family.

We next hypothesized that splenic effectors remained on their trajectories with no lineage plasticity. To test this, we focused on IAT, where trajectories were better preserved over time. Since Cxcr5 and Cxcr6 were the top bifurcating effector genes (Fig. 4e), we sorted CXCR5⁺CXCR6^– (T_FH) and CXCR6⁺CXCR5^– (T_H1) PbTIIs at day 7, transferred them separately into infection-matched mice and administered IAT until day 28; the reference group of control mice harbored PbTIIs that were unperturbed after transfer, infection and IAT (Supplementary Fig. 1d). At day 28, PbTIIs derived from CXCR5⁺CXCR6^– effectors expressed neither T_H1-associated CXCR6 or CCL5 (Fig. 5e), high levels of Id2 (Fig. 5f) nor reduced expression of TCF1 relative to those in reference controls (Fig. 5f). They also exhibited reduced in vitro IFN-γ recall relative to that in controls (Fig. 5g). PbTIIs derived from CXCR6⁺CXCR5^– effectors were undetectable in spleens, but were found expressing CXCR6, not CXCR5, in the liver (Fig. 5h). Fate-mapping performed in saline-treated mice elicited similar results, suggesting that IAT had not restricted lineage plasticity (Extended Data Fig. 7b,c). Thus, single naive clones produce effector progeny that progress along T_H1 and T_FH lineages with only a modest clone-specific preference for either; having done so, memory fate is set because little plasticity exists between splenic trajectories.

T_CM and Tr1 emerge within T_FH and T_H1 lineages

CCR7 and CD62L mark naive and T_CM cells. Sell (encoding CD62L), after downregulation at priming, was re-expressed along both lineages (Extended Data Fig. 8a). Ccr7 was retained during clonal expansion and by only T_FH lineage cells (Extended Data Fig. 8a,b), supporting the idea that T_CM cells develop along the T_FH lineage. We hypothesized that T_CM cells in lymph nodes could not display a T_H1 phenotype. Consistent with this, fewer PbTIIs in inguinal lymph nodes were T_H1-like than in spleen (Extended Data Fig. 8c). We next searched for GC T_FH cells along the T_FH lineage. Pdcd1 (encoding PD1) and Ccr7 were poorly coexpressed within the T_FH lineage (Extended Data Fig. 8d,e), suggesting that if Pdcd1⁺Cxcr5⁺ GC T_FH cells and Ccr7⁺ T_CM emerged within the T_FH lineage, the GPfates model had insufficient power to distinguish them.

Finally, Ifng was not confined to the T_H1 lineage, but was expressed more so in that lineage than in the T_FH lineage (Extended Data Fig. 8f,g). In contrast, Il10 expression was largely confined to the T_H1 lineage: 34% of T_H1 lineage cells coexpressed Ifng and Il10 during persistent infection, consistent with progressive Tr1 development from T_H1 effectors, that was limited by IAT (Extended Data Fig. 8f,g). Thus, on the basis of GPfates modeling, we propose that T_FH effectors gave rise to T_CM cells, GC T_FH cells and potentially T_FH memory, while T_H1 cells gave rise to Tr1 cells and T_H1-phenotype T_EM cells.

IAT limits exhaustion in both lineages and boosts GC T_FH cells

We next determined broader effects of IAT, examining ‘late pseudotime’ in GPfates models to focus on exhaustion and memory. Differential-gene-expression analysis between IAT and saline within either lineage revealed that IAT altered the expression of 1,110 genes (false-discovery rate (FDR) < 0.05) in the T_H1 lineage (Supplementary Table 1) compared with 252 (FDR < 0.05) for the T_FH lineage (Supplementary Table 2). GO analysis revealed IAT increased expression of immune-associated genes in the T_H1 lineage, with fewer immune effects on the T_FH lineage (Fig. 6a). There were 175 immune genes upregulated in the T_H1 lineage by IAT, 17 of which were also downregulated in the T_FH lineage, including Il7r, Cd96 and Cd40lg (Fig. 6a). Fifty-four immune genes were downregulated by IAT in the T_H1 lineage, 20 of which were also downregulated in the T_FH lineage, including Ctla4, Lag3, Tigit, Icos, Il21, Cd3d and Cd3g. These data suggest that IAT influenced both lineages, but more so T_H1.

Fig. 6: IAT boosts development of GC T_FH and memory CD4⁺ T cells during malaria.

a, Venn diagrams showing the number of differentially expressed genes (genes with FDR < 0.05) involved in immune-system process (GO:0002376) performed on the Smart-seq2 (384) dataset only. An overlapping list of genes between genes upregulated in the T_H1 and T_FH branches of late cells (pseudotime > 0.9) during IAT (pink, left), or upregulated in the T_H1 and T_FH branch of late cells in the saline group (green, right). b, UMAP representation of PbTII cells isolated at D28 p.i. from saline or IAT groups, analyzed using the droplet-based 10x Genomics Chromium platform. UMAP was calculated from the first 10 principle components using 1,394 highly variable genes. Data are available as ‘Memory’ on our web-based GUI: https://haquelab.mdhs.unimelb.edu.au/cd4_memory/. c, Unsupervised clustering of UMAP representation in b. d, Visualization of Cxcr6, Cxcr5 and Mki67 on UMAP representation in b. e, Violin plots showing the exhaustion score, Il7r expression and Il10 expression for all clusters shown in UMAP representation in b. Median expression for each cluster is denoted for each gene and score. f, Top, representative FACS plots and graph showing surface expression of LAG3 and TIGIT on PbTII cells at D28 p.i. Bottom, representative FACS plots and graph showing the kinetics of IL7R expression over time. Data are representative of 2 independent experiments (n = 5 mice per group, per individual timepoint for each independent experiment) and are presented as mean ± s.e.m. Statistical analysis was performed between saline and IAT groups for each timepoint individually using two-tailed Mann–Whitney test. g, Violin plots show GC T_FH and T_CM score for all clusters shown in UMAP representation in b. Median expression for each cluster is denoted for each score. h, Pie charts show the proportion of GC T_FH and T_CM cells within the ‘T_FH’ population for PbTII cells at D28 p.i. with or without IAT as described in b and g. i, Representative FACS plots and graph showing proportion of PbTIIs as GC T_FH cells (CXCR5⁺PD-1⁺) at D28 p.i. in the presence or absence of IAT. Data are representative of 3 independent experiments (n = 6 mice per group, per independent experiment) and are presented as mean ± s.e.m. Statistical analysis was performed using two-tailed Mann–Whitney test. **P < 0.01.

Power to detect differentially expressed genes between groups of single cells is influenced by heterogeneity. We reasoned that increased heterogeneity in the T_FH lineage had obscured IAT effects, and that distinguishing GC T_FH from T_CM cells required analysis of more cells. We also questioned whether persisting infection had merged T_FH and T_H1 lineages into a single state. We assessed PbTIIs from IAT and control groups via droplet-based scRNA-seq at day 28 (3,227 and 3,380 cells, respectively, following quality control), performing unsupervised clustering on the combined datasets (Fig. 6b), available for interrogation as the ‘Memory’ GUI at http://haquelab.mdhs.unimelb.edu.au/cd4_memory/. During persisting infection, PbTIIs exhibited two main phenotypes (Fig. 6b): T_FH-like cluster 1, and T_H1-like cluster 3 (Fig. 6c,d). Minor cluster 5 was proliferative based on cell-cycling genes, including marker of proliferation Ki-67 (Mki67) (Fig. 6c,d). Finally, during persisting infection, rare PbTIIs occupied distal clusters 2, 4, 6 and 7, as well as minor cluster 8 (characterized by type I IFN genes). Separated clusters 1 and 3 suggested that persisting infection had not merged T_H1 and T_FH lineages into a single state. In addition to these, rarer phenotypes had also been generated during persisting infection.

Considering IAT effects, T_H1-like clusters 3 and 4 (Fig. 6c,d) were mainly from saline and IAT groups, respectively (Fig. 6b). They were distinct from each other, with cluster 4 being smaller in size, suggesting that IAT had uniformly altered the T_H1 lineage and reduced heterogeneity. T_FH-like clusters 1 and 2 were less distinct, remaining partly adjacent to each other. Most strikingly, in contrast to T_H1 cells, IAT increased heterogeneity among non-T_H1 clusters 2, 6 and 7 (Fig. 6b). Nevertheless, IAT reduced expression of exhaustion-related genes in both lineages, including Lag3 (Fig. 6e and Supplementary Tables 3 and 4), abrogated Il10 expression in the T_H1 lineage (Fig. 6e) and boosted Il7r expression, all of which are consistent with the GPfates models. Predicted effects of IAT on Lag-3, TIGIT and IL-7R were confirmed at the protein level (Fig. 6f). Collectively, our results revealed that IAT had reduced exhaustion in both lineages, limited Tr1 development, condensed T_H1 memory phenotypes and increased heterogeneity within the T_FH lineage.

To search for GC T_FH and T_CM cells, we employed small curated gene lists: Cxcr5, Bcl6, Il21, Pdcd1 and Icos for GC Tfh, and Ccr7 and Sell for T_CM. Cluster 6 exhibited the strongest GC T_FH signature, and cluster 7 the strongest T_CM signature (Fig. 6b,c,g). Cluster 6 also appeared to be moderately exhausted, likely owing to Pdcd1 and Cd160, markers common to exhaustion and GC biology (Fig. 6e,g). Clusters 6 and 7 were smaller than the adjacent T_FH-like cluster 2, consistent with the GPfates model having missed phenotypes present at <5%. Importantly, the proportion of T_FH-like cells with T_CM or GC T_FH signatures increased by ~80% under IAT (Fig. 6h), confirmed for GC T_FH by flow cytometry (Fig. 6i). Finally, we identified genes differentially expressed between GC T_FH and T_CM cells (Supplementary Table 5), and those elevated in GC T_FH compared with other fates at day 28 (Supplementary Table 6). In addition to T_FH-associated genes, Pdcd1, Tox2, Cxcr5, Id3, Tcf7, Tox and Bcl6, we noted those previously unstudied, including Rgs10 and Ppp1r14b (Supplementary Table 6). Thus, IAT had improved CD4⁺ T cell responses by preventing exhaustion and boosting T_CM and GC T_FH cells.

scATAC-seq reveals effector heterogeneity and partial resetting of chromatin in memory

We hypothesized that lineage convergence and partial reversion to naivety was an artefact of transcriptomics. We examined chromatin using scATAC-seq in naive (day 0), proliferating (day 4), effector (day 7) and memory (day 32 with IAT or control saline) PbTIIs (Fig. 7a), as well as sorted T_FH lineage (CXCR5^hi) and T_H1 lineage (CXCR6^hi) effectors (Fig. 7a). Across all cells, we identified 59,197 accessible peaks. Using latent semantic indexing (LSI)³⁰ and PCA, PC2 separated proliferating and effector cells from naive and memory cells, while PC3 separated T_H1 from T_FH effectors (Fig. 7a). Chromatin accessibility was substantially more heterogeneous among effectors than among their naive counterparts. PCA and UMAP (Fig. 7a,b) revealed epigenomes at day 32 that were more homogenous than effectors, and clustered close to naive cells. Targeted analysis of transcription-factor-binding motifs suggested differential and inverse patterns of accessibility for Id2 and Tcf7 in T_H1 effectors compared with in T_FH effectors, which was partially reset to naive levels by day 32 (Fig. 7c). Tbx21 motifs were enriched, and Bcl6 motifs depleted during the first week, which was partially lost by day 32 (Fig. 7c). Thus, phenotypic convergence of effectors during transit to memory was evident by epigenomics.

Fig. 7: scATAC-seq reveals partial epigenomic resetting and homogenization during memory development.

a, PCA of peaks after application of latent semantic indexing analysis³⁰. b, UMAP representation of the epigenomic landscape of PbTII cells from PC 2–10. c, Changes in scATAC-seq peaks associated with different T_H1- or T_FH-associated transcription factors (Id2, Tcf7, Tbx21, Bcl6) for different timepoints. Error bars represent mean ± s.d. Statistical test was performed using two-sided Wilcoxon rank-sum test. d, Reactome analysis of the top 10 enriched pathways for different timepoints. The P value was adjusted for multiple testing using Benjamini–Hochberg correction. e, Top, PCA analysis of 500 variable motifs (combined database of human and mice motifs from Jaspar database) calculated using chromVAR analysis for comparison between cells from naive (D0), saline and IAT groups at D32 p.i. Bottom, top 50 motifs explaining variability between cells from naive (D0), saline and IAT groups at D32 p.i. f, Changes in scATAC-seq peaks associated with the top variable transcription factors detected from chromVAR analysis as described in e. Error bars represent mean ± s.d. Statistical test was performed using two-sided Wilcoxon rank-sum test. **P < 0.01, ***P < 0.001.

Nevertheless, epigenomic differences remained between naive and day 32 cells. Reactome pathway analysis revealed enrichment for peaks associated with tyrosine-kinase and CD28 signaling at day 32 that was absent in naive cells (Fig. 7d). Thus, specific changes in chromatin had been retained in memory. We sought to infer transcriptional regulators whose access to chromatin varied between naivety and memory³¹. PCA of motif bias-corrected deviation revealed that day 32 motif variability was similar, although distinct from, that of naive cells (Fig. 7e). Binding motifs were variable for Fos, Jun, Runx, Rel, Maf, Tbx and Irf families (Fig. 7e and Supplementary Table 7), with motif enrichment for Irf8, JunD and MafF, and loss for Tcf7, Ctcf, Etv6, Nfatc3, and Runx1 in memory compared to naivety (Fig. 7f and Extended Data Fig. 9). Thus, epigenomic and transcriptomic heterogeneity induced during effector differentiation was partially, but incompletely, reset during transit to memory (Extended Data Fig. 10).

Discussion

Here, we examined the dynamics of CD4⁺ T cell memory development using a model of malaria and treatment with antimalarials. We mapped transcriptomic change as CD4⁺ T cells gradually transitioned from effector to memory over four weeks, and in doing so inferred genomic relationships between T_FH, T_H1, Tr1, GC T_FH, T_CM, T_EM, naive, proliferating and exhausted cells (terms defined in Table 1). We revealed that CD4⁺ T cell fate toward either T_H1-effector memory or various T_FH and T_CM fates was determined during the first week of infection, whose phenotypes were altered by antimalarial drugs.

Currently, the only available malaria vaccine, RTS,S/AS01, generates short-lived immunity that correlates with antibodies and CD4⁺ T cell-derived IL-2 and TNF^32,33. Since T_H1 cells also provide protection³⁴, one vaccine strategy might be dual promotion of antibodies and T_H1 memory. Our study highlights that these mechanisms can be elicited simultaneously among T cells of the same specificity during infection, although development of one likely influences and interferes with the other^22,35. One possible solution might be to temporally segregate T_H1 and T_FH differentiation, by priming and boosting with different antigens, and employing adjuvants that block and then promote T_H1 immunity.

One conclusion from our work is that memory fate is determined during the first week of infection. Therefore, diverting CD4⁺ T cells along a particular lineage may be possible at early, but not late, stages of infection or vaccination. Nevertheless, consistent with clinical studies^13,14, our data suggest that reducing parasite load profoundly affects cellular phenotypes and functional potential. We speculate that anti-malarial drug treatment in endemic regions minimizes deleterious effects of chronic infection on naturally acquired or vaccine-mediated immunity.

In contrast to LCMV infection⁷, CCR7⁺ T_CM precursors were absent during Plasmodium infection. One interpretation might be that all effectors expressed equal T_CM potential, although this appears unlikely given the restriction of Ccr7 along one lineage, and that fate-mapping suggested minimal crossover between trajectories. The discrepancy between studies may be due to differences in viral versus parasitic infection¹⁰, the specific transgenic TCRs used or specifics of infectious dose or Plasmodium species employed. Nevertheless, our data argue that T_CM precursors are not essential for CD4⁺ T cell memory development.

Clonal analysis confirmed that diversity in fate is common among clones during parasitic infection¹⁸. Mathematical modeling revealed a modest predisposition to one fate among sibling cells. This suggested heterogeneity in responses by individual clones, despite transcriptomic and epigenomic homogeneity in naivety, and predictable frequencies of each fate at a population level. Our observations are consistent with those from bacterial infection³⁶. Given heterogeneity in microanatomical location, and since B cells and myeloid cells influence effector fate¹⁸, we hypothesize that cell–cell interactions are the primary drivers of differences between clones.

A caveat of our dynamic modeling was its inability to distinguish GC T_FH from T_CM cells within the T_FH lineage. A focused analysis at a late time-point revealed that, although the bulk of T_FH lineage cells exhibited a heterogeneous T_FH phenotype, rarer GC T_FH and T_CM subpopulations were present. GC T_FH and T_CM cells differentially expressed many genes, including those previously reported such as Klf2 (ref. ³⁷) and Satb1 (ref. ³⁸). Our analysis suggested a continuum between T_FH and GC T_FH cells, with T_CM cells as a separate, adjacent population. We propose that cells progress along the T_FH lineage, with small proportions differentiating further into GC T_FH or T_CM cells.

Tcf7 promotes T_FH cell differentiation, but its role in CD4⁺ T cell memory is unclear. We noted Tcf7 expression in CD4⁺ T cells along the T_FH lineage—indeed, TCF1 supported this trajectory in our model. Tcf7 was also preserved during persisting infection, suggesting that exhaustion was not an impediment to its potential function. This is reminiscent of CD8⁺ T cell reports^39,40,41,42, wherein TCF1 supported stem-like memory cells in humans, and precursor exhausted cells were reinvigorated via immune checkpoint blockade. These observations suggest a generalized model in which TCF1 is retained in one lineage after bifurcation, supporting their subsequent recall responses.

An important feature of the T_H1 and T_FH lineages was their gradual, partial coalescence and incomplete return to naivety during memory. Nevertheless, changes in transcription-factor-binding motifs were preserved among individual memory cells—for example, Maf and Irf8 motifs. Several studies have examined epigenomic and transcriptomic changes in CD8⁺ T cells during memory development⁴³, or in humans undergoing vaccination or cancer treatment⁹. These support a model of CD8⁺ T cell memory characterized by partial reversion to naive-like states, consistent with that presented here for CD4⁺ T cells.

Drug treatment boosted T_H1 memory cells and GC T_FH formation in our mouse model, perhaps consistent with what happens in humans in malaria-endemic regions, where immunity may be improved by antimalarial drugs. A caveat of our model is that immunity induced by a single infection is so robust that improvements to parasite control cannot be assessed functionally⁴⁴. Nevertheless, we previously associated T_FH boosting with improved primary immunity⁴⁵. We speculate that exhaustion, impaired antigen presentation⁴⁶ and IL-10 immune regulation⁴⁷ impair GC T_FH responses, and are alleviated by removing parasites with antimalarial drugs. In endemic areas, children are infected multiple times each season. The effect of multiple infections and drug treatment on CD4⁺ T cells remains to be tested.

Thus, transition from effector to memory and GC T_FH cell states occurs gradually in CD4⁺ T cells during experimental malaria. The extended timeframe over which transcriptomic change occurred, and the effect of antimalarial drugs, suggests that there are opportunities for manipulating memory development after primary immune responses have been triggered. We identify potential targets for manipulating CD4⁺ T cell memory, with potential clinical utility in malaria-vaccine development or in prevention of ongoing, persisting Plasmodium infections. Given pivotal roles for CD4⁺ T cells in bacterial, fungal and viral infections, our datasets may be relevant across a range of human infectious diseases.

Methods

Experimental mice, adoptive transfer and infections

C57BL/6J and SJL.Ptprca mice were purchased from Animal Resources Centre (Canning Vale), PbTII¹⁹, nzEGFP and Cre^ROSA26Ert2 mice were bred in-house. nzEGFP mice express enhanced green fluorescent protein (eGFP) from the ubiquitously expressed CMV early enhancer/chicken beta actin promotor. Id3^GFP mice and Tcf7-floxed (Tcf7^fl/fl) were crossed with PbTII mice. Only Id3^GFP/+ mice were used as reporters. All mice were female, aged 8–12 weeks, and maintained under specific-pathogen-free conditions within the animal facility at QIMR Berghofer Medical Research Institute. Mice were housed in exhaust-ventilated cages (Opti-mice) with ≤6 mice per cage, and room temperature was maintained between 19 °C and 22 °C, with humidity between 55% and 65% and with 12-h/12-h dark/light cycle with a 15-min sunrise and sunset. All animal procedures and protocols were approved (A1503-601M) and monitored by the QIMR Berghofer Medical Research Institute Animal Ethics Committee.

Spleens were collected and homogenized through a 100-µm cell strainer to create a single-cell suspension. Red blood cells (RBCs) were lysed using RBC Lysing Buffer Hybri-Max (Sigma-Aldrich) and CD4⁺ T PbTII cells were enriched using CD4 microbeads (Miltenyi Biotec). Cells (1 × 10⁴ per mouse) were transferred to each mouse via lateral tail-vein intravenous (i.v.) injection.

PcAS parasites were used after thawing frozen, infected blood stabilites and performing a single in vivo passage in C57BL/6J mice. PcAS-infected RBCs were collected from passage mice by cardiac puncture and mice infected (1 × 10⁵ infected RBCs per mouse) via lateral tail vein i.v. injection.

Intermittent antimalarial drug treatment (IAT)

Sodium artesunate (Guilin Pharmaceutical or sourced from J. Mohrle) was prepared according to the manufacturer’s protocol by diluting in 0.9% saline (Baxter) to a final concentration of 5 mg ml^–1. Mice were treated via intraperitoneal (i.p) injection with sodium artesunate (1 mg per mouse), or vehicle control saline, twice daily from day 7 to day 9 p.i., once daily from day 10 to day 16 p.i. and then twice weekly until experimental endpoint. Mice treated with sodium artesunate were also administered pyrimethamine (0.07 g l^–1; Sigma-Aldrich) in drinking water for the duration of the treatment.

Tamoxifen preparation and administration

Tamoxifen (Sigma Aldrich) was dissolved in 10% ethanol in corn oil (Sigma Aldrich) at a concentration of 50 mg ml^–1. This was sonicated for 30 min–1 h, until all tamoxifen particles had dissolved. Mice were administered with a single 5-mg dose of tamoxifen via i.p injection.

Parasitemia assessment

Parasitemia assessment was carried out as previously described⁴⁵. Briefly, a single drop of blood was collected via tail bleed and diluted in 250 µl of RPMI medium containing 5 U ml^–1 heparin sulphate. Diluted blood was stained with Syto84 (5 µM; Life Technologies) and Hoechst 33342 (10 µg ml^–1; Sigma-Aldrich) for 30 min in the dark at room temperature (RT). Staining was quenched with ×10 volume of ice-cold RPMI medium, and samples were immediately acquired by flow cytometry.

Flow cytometry

Spleens and lymph nodes were collected and homogenized through a 100-µm cell strainer to create a single cell suspension, and RBCs were lysed using RBC Lysing Buffer Hybri-Max (Sigma-Aldrich). Cells were assessed for viability using a LIVE/DEAD Fixable Aqua Dead Cell Stain Kit (1:200 dilution, catalog no. 423102, Biolegend), according to the manufacturer’s protocol, unless otherwise specified. Prior to antibody staining, Fc receptors were blocked using antibodies against CD16 and CD32. Cells were incubated with surface marker antibodies for 20 min at 4 °C. Staining for PerCpCy5.5 CCR7 surface antibody (1:10 dilution, clone: 4B12, catalog no. 120116, Biolegend) was performed at 37 °C for 1 h as per the manufacturer’s recommendations, after Fc receptor blocking. Other surface marker antibodies used were: APC–CD4 (1:200 dilution, clone: RM4-5, catalog no. 100516, Biolegend), PE–CD4 (1:200 dilution, clone: RM4-5, catalog no. 100512, Biolegend), BV421 –CRβ (1:200 dilution, clone: H57-597, catalog no. 109230, Biolegend), PerCpCy5.5–TCRβ (1:200 dilution, clone: H57-597, catalog no. 109228, Biolegend), PerCpCy5.5–Vβ12 (1:200 dilution, clone: MR11-1, catalog no. 46-5798-80, eBioscience), AF700–CD44 (1:200 dilution, clone: IM7, catalog no. 103026, Biolegend), PE–CD62L (1:200 dilution, clone: MEL-14, catalog no. 104408, Biolegend), APC–CXCR6 (1 200 dilution, clone: SA051D1, catalog no. 151106, Biolegend), biotinylated CXCR5 (1:50 dilution, clone: 2G8, catalog no. 551960, BD Biosciences), PE–Ki67 (1:200 dilution, clone: SolA15, catalog no. 12-5698-82, eBioscience), BV605–CD69 (1:200 dilution, clone: H1.2F3, catalog no. 104530, Biolegend), BV421–IL-7R (1:200 dilution, clone: A7R34, catalog no. 135924, Biolegend), PE–CXCR3 (1:200 dilution, clone: CXCR3-173, catalog no. 126506, Biolegend), APC–Lag3 (1:100 dilution, clone: C9B7W, catalog no. 125210, Biolegend), BV421–TIGIT (1:100 dilution, clone: 1G9, catalog no. 565270, BD Biosciences), AF700–CD45.1 (1:200 dilution, clone: A20, catalog no. 110724, Biolegend). Other staining dyes or flurophores include strepavidin–PeCy7 (1:200 dilution, catalog no. 405206, Biolegend) and propidium iodide (1:400 dilution, catalog no. 12-9855-41, Sigma-Aldrich).

To assess cytokine production, cells were incubated with brefeldin-A (10 mg ml^–1) with or without ionomycin (500 ng ml^–1) and PMA (25 ng ml^–1) at 37 °C for 3 h. Intracellular staining for cytokines, transcription factors and chemokines was then performed using the eBioscience Foxp3/Transcription Factor Staining Buffer Set. Staining with intracellular antibodies was conducted at 4 °C for 1 h. Where TCF1 was stained after stimulation with PMA and ionomycin for 3 h is explicitly stated in figure legends. All other transcription factors and chemokine molecules were stained without prior stimulation. Intracellular marker antibodies used were: APC–Id2 (1:100 dilution, clone: ILCID2, catalog no. 17-9475-82, eBioscience), PE–TCF1 (1:100 dilution, clone: C63D9, catalog no. 2203 S, Cell Signaling Technology), PE–CCL5 (1:100 dilution, clone: 2F9, catalog no. 149104, Biolegend), PE–c-Maf (1:100 dilution, clone: sym0F1, catalog no. 12-9855-41, Thermo Fischer), PE–IL-10 (1:100 dilution, clone: JES5-16E3, catalog no. 505008, Biolegend), BV421–IFN-γ (1:200 dilution, clone: XMG1.2, catalog no. 505830, Biolegend), APC–T-bet (1:50 dilution, clone: 4B10, catalog no. 17-5825-82, Thermo Fischer); PE–BCL6 (1:10 dilution, clone: K112-91, catalog no. 561522, BD Biosciences). Samples were acquired on a LSRII Fortessa analyzer (BD Biosciences) and subsequently analyzed using FlowJo software (Treestar).

Livers were collected in 1% (vol/vol) FCS/PBS and homogenized through a 200-µm metal sieve. Liver suspension was resuspended in 33% (vol/vol) percoll/PBS before centrifugation at 1,700 r.p.m. for 12 min at room temperature. After removal of supernatant containing unwanted cells and debris, liver leukocyte pellet was processed similarly as splenic and lymph-node tissue described above.

PbTII memory recall during in vivo rechallenge

Mice harboring previously in vivo primed eGFP⁺ PbTIIs received CD4-microbead enriched, non-eGFP⁺ congenically-marked (CD45.1⁺) PbTII cells from naive donors (1 × 10⁵ cells per mouse) 1 day prior to homologous high-dose rechallenge with 1 × 10⁷ PcAS-infected RBCs.

Immunofluorescence microscopy

Spleens were fixed with 2% paraformaldehyde for 2–4 h at RT in the dark followed by dehydration with 30% sucrose (Chem-supply) overnight at RT in the dark. Spleens were snap-frozen in Tissue-Tek Optimal Cutting Temperature embedding medium (Sakura Finetek) on dry ice and stored at −80 °C. Spleens were sectioned at 10–30 μm on polysine slides such that consecutive sectioning is avoided. Sections were allowed to dry overnight. The dried sections were rehydrated for 15–20 min before fixation in 4% paraformaldehyde for 15–20 min at RT in dark. Slides were washed 3 times for 5 min in washing buffer (0.01% Tween20 in PBS) before permeabilization with 0.1% Triton X-100 in washing buffer for 10–15 min. After washing, slides were incubated with Medical Background Sniper (Biocare) for 30 minutes. Slides were rinsed for 2 min and endogenous biotin was blocked using an Avidin/Biotin Blocking kit (Vector Laboratories), according to manufacturer’s protocol. Tissue sections were then stained with rabbit anti-GFP (1:500 dilution, catalog no. ab6556, Abcam), rat anti-mouse CD3-AF594 (1:200 dilution, clone: 1742, catalog no. 100240, Biolegend), rat anti-mouse IgD-AF647 (1:50 dilution, clone: 11–26 c.2a, catalog no. 405708, Biolegend) and biotinylated peanut agglutinin (1:500 dilution, catalog no. B-1075-5 PNA; Vector Laboratories, Inc.) for 1–2 h at RT in the dark. Secondary antibody staining for PNA and GFP was performed using streptavidin–AF555 (1:300 dilution, catalog no. S21381, Thermo Fisher Scientific) and donkey anti-rabbit AF488 (1:300 dilution, catalog no. R37118, Thermo Fisher Scientific), respectively, for 1–2 h at RT in the dark. Tissue sections were incubated with DAPI for 10–15 min to counterstain nuclei, and slides were mounted in Dako Mounting Media (Agilent Technologies). Image acquisition was performed using an Aperio FL slide scanner or a Zeiss 780-NLO point scanning confocal microscope at ×20, ×40 and ×63 objective.

Cell detection and quantification were performed using the spot-detection function in Imaris image analysis software (Bitplane), with thresholds <10 μm. All objects were manually inspected for accuracy before data were plotted and analyzed in GraphPad Prism. Colocalization analysis was performed using Imaris colocalization functions. B cell follicle was defined as IgD-positive region, and GC was defined as IgD-negative and PNA-positive region within B cell follicle.

Bulk Fast-ATAC sequencing and analysis

Fast-ATAC sequencing was performed as previously described⁴⁸. Briefly, 5,000 viable PbTII cells were sorted by flow cytometry and pelleted by centrifugation. Supernatant was removed and 50 µl of transposase mixture (25 µl TD buffer (Illumina), 2.5 µl of TDE1 (Illumina), 0.5 µl of 1% digitonin (Promega), 22 µl nuclease-free water) was added to the cells. Cells were then incubated at 37 °C for 30 min at 300 r.p.m. in an Eppendorf ThermoMixer. Transposed DNA was amplified and purified using a Qiagen MinElute PCR Purification kit, according to the manufacturer’s protocol. Purified DNA was eluted in 20 µl Buffer EB (Qiagen) and sequenced using paired-end sequencing on a NextSeq 550 instrument (Illumina).

Raw ATAC-seq reads were mapped to mouse genome MGSCv37 (mm9) using BWA–MEM⁴⁹ (version 0.7.15). Unmapped reads, reads mapping to unassigned contigs and mate unmapped reads were removed, as well as mitochondrial genes and PCR duplicates. The resulting .bam files were first converted to bedGraph using the bedtools genomecov command from the BEDTools suite⁵⁰ (version 2–2.29.0) and then converted to bigwig format using the bedGraphToBigWig program from University of California, Santa Cruz (UCSC) Genome Browser (https://genome.ucsc.edu/). Read counts were normalized to the number of uniquely mapped reads per million. Peak calling was performed using MACS2 (version 2.1) with the following parameters:–nomodel, –shift 37, –extsize 73, –pvalue 1e-5. The bigWig tracks and narrowPeak files were visualized in a custom UCSC Genome Browser track hosting the mouse mm9 reference genome. Peaks overlapping with the mm9 blacklist region generated by UCSC were removed, and overlapping peaks between replicate samples were identified using the findOverlaps command from the GenomicRanges (version 1.38) package and used for downstream analysis. Peaks were annotated using the R package ChIPseeker⁵¹ (version 1.22.1) with transcription start site (TSS) reaching from −3,000 to 3,000 to the UCSC mm9 gene model and the org.Mm.eg.db R package.

Single-cell capture and sequencing

Droplet-based scRNA-seq

Recipient mice (n = 6 mice per group per timepoint) received 1 × 10⁴ sorted naive (CD62L⁺CD44^–) eGFP-expressing PbTII cells from a single donor 24 h prior to infection with PcAS. Spleens from 5 infected, recipient mice were pooled, and viable PbTII cells were sorted at D7 and D28 (saline and IAT groups) into cold 1% BSA/PBS buffer. PbTIIs were counted, and ~8,000 cells were loaded per channel onto a Chromium controller (10x Genomics) for generation of gel bead-in-emulsions. Sequencing libraries were prepared using Single Cell 3′ Reagent Kits v3 (10x Genomics) and then converted using the MGIEasy Universal Library Conversion Kit (BGI) before sequencing on a MGISEQ-2000 instrument (BGI).

Smart-seq2 scRNA-seq

We prepared 384-well lo-bind plates with 0.5 µl of Triton-X lysis buffer, 0.25 µl of 10 µM olido-dT30-VN, 0.25 µl dNTP mix (25 mM each) and ERCC controls (final dilution of 1:64 million) per well and stored at −20 °C until use. Plates were thawed on ice before sorting single PbTII cells isolated from 1 selected donor mouse at each timepoint (n = 6 mice per group per timepoint) into each well. Single-cell lysates were sealed and spun at 100g for 1 min, then immediately frozen on dry ice and stored at −80 °C. Reverse transcription and PCR were performed following the Smart-Seq2 protocol with the following modifications: 1 µl of reverse transcription mix and 5.7 µl PCR Master mix. After amplification, complementary DNA was subjected to quality control using 1 µl of amplified cDNA on an Agilent 2100 BioAnalyzer and Agilent High Sensitivity DNA kits (Agilent Technologies). Next, 5 µl of cDNA per cell was cleaned using Agencourt AMPure XP beads (Beckman Coulter) at a 1× ratio on a Hamilton STAR liquid handler (Hamilton Robotics). cDNA was quantified using Biotium AccuClear High Sensitivity DNA quantification reagent and normalized to 1 ng µl^–1 in a total volume of 500 nl before library preparation using Nextera XT DNA Sample Preparation Kit (Illumina). Then, 125 nl of in-house index adapters (Integrated DNA Technologies) similar to Illumina N7 and N5 indices were added to the tagmentation reaction before adding 1.5 µl of KAPA HiFi DNA polymerase (KAPA Biosystems) and performing 12 cycles of PCR according to the manufacturer’s instructions. After PCR, all samples were pooled into 384-plex pools using a 24 × 16 dual indexing approach, and the pool was cleaned using Agencourt AMPure XP beads at a 0.6× ratio. Library pools were eluted in buffer EB and quality controlled using an Agilent 2100 BioAnalyzer and Agilent High Sensitivity DNA kits before the concentration was adjusted to 4 nM. Each library was sequenced on 1 lane of Illumina HiSeq 2000 version 4 chemistry (paired-end 75-base-pair reads).

Processing of scRNA-seq data

Smart-seq2 data

Raw reads were mapped to a mouse genomic index (Ensembl version 70) using STAR aligner⁵² version 2.5.2 and transcripts per millions (TPM) were estimated by RSEM⁵³ version 1.2.30. ERCC RNA spikes were included in the index, but removed for down-stream analysis.

10x Genomics data

For the BGI FASTQ files to be made compatible with ‘cellranger count’ pipeline from Cell Ranger version 3.0.2 (10x Genomics), file names and FASTQ headers were reformatted using code from https://github.com/IMB-Computational-Genomics-Lab/BGIvsIllumina_scRNASeq⁵⁴. Data were processed using 10x mouse genome 1.2.0 release as a reference.

Quality control of scRNA-seq data

10x Genomics data

Cells were filtered to remove those expressing fewer than 500 genes and more than 6,000 genes, and those with more than 35% mitochondrial content. Only genes expressed in three or more cells were considered. Regression of the number of unique molecular identifiers was performed for each analysis individually.

Smart-seq2 data

Cells were filtered to remove those with fewer than 100,000 reads mapping to the mouse genome, those expressing fewer than 1,000 genes and more than 5,000 genes and those with more than 35% mitochondrial content. Only genes expressed at one or more TPM in three or more cells were considered, unless otherwise specified.

Batch-effect correction

Two PbTII datasets (Smart-seq2 (96) and SMARTer C1) from our previous study underwent quality control as described previously¹⁸ and were combined with Smart-seq2 (384) dataset from our current study. Genes expressed at less than one TPM in fewer than three cells were globally removed. The ‘removeBatchEffect’ function from Limma⁵⁵ was then used to remove the batch effect. All three datasets shown have been regressed as per described unless otherwise specified.

Dimensionality reduction

PCA dimensionality reduction (prcomp) was performed on Smart-seq2 dataset with the genes expressed at 100 or more TPM in 15 or more cells as input. On the combined, batch-effect corrected dataset, dimensionality reduction was performed with BGPLVM using GPfates as previously described¹⁸.

To generate UMAP of the combined dataset, the ‘removeBatchEffect’ function from Limma was used to remove the batch effect from the dataset, which contained highly variable genes selected using ‘trendVar’ and ‘decomposeVar’ from Scran (v1.6.9)⁵⁶ (FDR < 0.05, bio > 0.5), followed by PCA dimensionality reduction and usage of ‘RunUMAP’ from Seurat⁵⁷ package (v2.3.4 unless otherwise specified).

Alternative integration of the 3 datasets was performed using single-cell variational inference (version 0.3.0). The expected-counts matrix obtained from RSEM was used as an input. All parameters were kept at default except: up to 4,000 genes were considered as an input, 30 latent variables were considered, 2 hidden layers were used for encoder and decoder neural networks and 100 epochs and a 0.01 learning rate were used to train the model. The computed latent variables were used as an input to generate UMAP using ‘RunUMAP’ from Seurat.

To perform PCA dimensionality reduction on the 10x Genomics dataset, we identified highly variable genes using the ‘FindVariableGenes’ function from the Seurat package and used these as an input. For each subset of data used for dimensionality reduction, highly variable genes were computed individually and used as an input. The number of highly variable genes used in each analysis is noted in the figure legends. The PCA output was then used as an input to generate UMAP using ‘RunUMAP’ from Seurat.

Gene-signature scoring

T_H1 and T_FH signature scores were calculated using the top 50 T_H1- and T_FH-bifurcating genes derived from Lonnberg et al.¹⁸. Cell-cycle signature score was calculated using 226 cell-cycle genes derived from Cyclebase⁵⁸. Eight coinhibitory receptors (Ctla4, Pdcd1, Lag3, Havcr2, Btla4, Cd160, 2b4, Tigit) established for their role in T cell exhaustion⁵⁹ were used for exhaustion-phenotype scoring.

Each cell on the 10x Genomics dataset was scored for various signatures defined in each figure legend using the ‘AddModuleScore’ function from Seurat. To score cells in the Smart-seq2 dataset, a summative expression of all genes defined within each signature was calculated for each cell.

Pseudotime and trajectory inference

Pseudotime was inferred on the combined, batch-effect-corrected dataset based on latent variable (LV) 1 coordinates from bGPLVM dimensionality reduction. Firstly, DBSCAN clustering was performed on D0–D3 cells and D4–28 cells separately, using LV1 and LV2 coordinates as inputs. From the D0–D3 group, only cells with LV2 coordinates lower than the lowest LV2 coordinate from the non-outlier cells were considered outliers, and from the D4–D28 group, only cells with LV2 coordinates higher than the highest LV2 coordinate from the non-outlier cells were considered outliers. Outliers from each group were grouped together with the cells from the opposing groups. LV1 coordinates of the first group (D0–D3 cells and outliers from D4–D28 cells) were flipped along the lowest LV1 coordinate from all pf the dataset and concatenated with the LV1 coordinates of the second group (D4–D28 cells and outliers from D0–D3 cells) to infer the pseudotime coordinates.

Overlapping mixtures of Gaussian processes (OMGP)

Lineages were traced along the pseudotime with OMGP using GPfates¹⁸, using LV2 coordinates to model as a function of pseudotime. Trajectory inferences were performed separately for saline and IAT groups, with the same set of parameters used (two trends assumed, global variance = 0.5, per-trend variance = 3 and per-trend lengthscale = 2). Here, the saline-treated group consists of all cells from D0–D7 and saline-treated cells from D10 onwards. The IAT group consists of all cells from D0–D7 and IAT-administered cells from D10 onwards.

Slingshot

Trajectories were inferred through UMAP cell embeddings using Slingshot v0.99.12 (ref. ²³). Cluster information from unsupervised clustering was used as an input. Slingshot analysis was performed separately for saline and IAT groups. A semisupervised approach was taken whereby clusters with a high proportion of D0 were specified as starting points.

RNA velocity

RNA-velocity analysis was performed using Velocyto version 0.17.16 (ref. ²⁴). Analysis was performed separately for saline and IAT groups and only timepoints from D4 onwards were included in the analysis. All parameters were kept at default except: minimum number of spliced molecules = 40, minimum number of cells expressing spliced molecules of a gene = 30, 1,000 gs to rank, and 20 PCs and 150 neighbors considered for knn imputation. Calculated velocity was projected onto pre-computed BGPLVM embeddings.

Functional dynamics analysis

Automatic expression histology (AEH) from spatialDE²⁶ was applied on the combined, batch-effect-corrected dataset. Only significantly variable genes along pseudotime were considered (FDR < 0.05). AEH was applied separately for saline and IAT samples, with the same set of parameters (number of patterns (c) = 7, lengthscale = 9). c specifies the number of groups of genes, with each group displaying distinct expression dynamic along pseudotime. Genes in Dynamic 1 (n = 511) were identified together in a small number of groups even when c = 70, indicating that gene groupings were robust to variation in parameter choice for c.

Gene-signature scores of each dynamic were computed using the ‘AddModuleScore’ from Seurat. Correlations between expression profiles of genes in dynamic 1 were quantified by calculating Spearman’s rho. Only correlations between genes higher than rho > 0.3 were considered as input for generation of a coexpression network. The network plot was generated using ggraph v2.0.3 with layout = ‘fr’ parameter.

Differential-gene-expression analysis

As an input for differential-gene-expression analysis (DGEA), expected counts from Smartseq2 data were estimated by RSEM version 1.2.30 and then normalized and log-transformed using ‘computeSumFactors’ and ‘normalize’ functions from scran. Pairwise Welch t-test was performed using ‘pairwiseTTests’ function from scran to identify genes differentially expressed between two groups of cells. Comparisons were done as follows: (1) cells at late pseudotime from IAT group on the T_H1 arm versus cells at late pseudotime from saline group on the T_H1 arm, (2) cells at late pseudotime from IAT group on the T_FH arm versus cells at late pseudotime from saline group on the T_FH arm (Fig. 4e). Genes with a FDR below 0.05 were considered to be differentially expressed. Cells with a pseudotime value higher than 0.9 were considered to be late pseudotime.

To generate the differential gene signature from comparison of different clusters within D28 data from the 10x Genomics dataset, cluster-specific genes were defined using ‘FindAllMarkers’ from Seurat by comparing between clusters as specified in each analysis.

GO term enrichment analysis

GO terms were obtained from the ‘org.Mm.eg.db’ Bioconductor annotation package⁶⁰. Fisher’s exact test was used to identify significantly over-represented GO terms. Input gene lists were differentially expressed genes from the comparisons described above, with up- and downregulated genes considered separately. The background gene set was reduced to those used as an input for DGEA. GO term enrichment analysis was conducted for functional-dynamics analysis using genes from each dynamic as input and all significantly variable genes along pseudotime as the background gene set. Enrichment analysis for biological process and molecular function terms was performed using the ‘goana’ function in edgeR⁶¹.

Testing the distribution of cell fates among clonotypes

TCR sequences for each cell from Smart-seq2 data were reconstructed using TraCeR as previously described²⁹. Cells were classified as belonging to the T_H1 group if their T_H1 assignment probability (TAP) was greater than the median TAP + 0.05, and cells were classified as belonging to the T_FH group if their TAP was less than the median TAP – 0.05. Analysis to test whether the distribution of cell fate among families was consistent with cells acquiring fates independently of the family to which they belong was focused on only those cells from the 97 families found in IAT group, excluding cells where fate was indeterminate. To test the null hypothesis that cells fate is randomly distributed across cells, independent of their family (that is ancestry), a bootstrapping approach was used. First, we constructed a likelihood function to describe the likelihood of observing the a given number of T_H1 cells, Th_i, in a particular family, i, of size n_i, assuming the process is binomial with some probability P_h. The overall likelihood of observing the distribution of cell fates across all families is then given by:

$$L = \mathop {\prod }\limits_{i = 1}^N Bi\left( {Th_i,n_i,P_{\rm{h}}} \right)$$

where N is the number of families. Note that the P_h that maximizes the likelihood will be given by the proportion of all cells that have a T_H1 fate (demonstrated by evaluating the derivative of when log of the likelihood function and solving for when derivative is zero). In this dataset the proportion of cells that were T_H1 was P_h = 0.575. Using the observed numbers of Th_i and n_i for each family, L in equation 1 was computed and was called the likelihood of the observed distribution (L₀).

To determine whether the likelihood of the observed distribution of cell fates by family is consistent with a random process, cells were randomly permuted among the different families 100,000 times. Each permutation, p, was performed in such a way as to preserve the size of families, but redistribute the cells across the families. After permuting the cells, the likelihood, L_p, was recalculated for each permutation p. This process produced a distribution of the likelihood associated with cell fates being randomly and independently distributed across families. The likelihood of the observed distribution of cell fates across families, L₀, was compared with the set of likelihoods, L_p, associated with each random permutation of cells across the fates, and the likelihood of finding the observed distribution of cell fates by family by random chance was less than approximately 0.05%, that is p = 0.0005. Thus, the null hypothesis that cell fate is independent of family was rejected.

The histogram of the observed versus expected distribution of families with a T_H1 predominance, a T_FH predominance or a mixture of fates (Fig. 6d), was determined by first drawing a histogram of the observed proportions of T_H1 cells in each family (noting that families of size 3 or 6 that had a proportion of 1/3 or 2/3 were included in the middle bin in the histogram). To determine the expected distribution of families, the above described permutation of cells between families was performed and the histogram redrawn each time. The expected distribution of families was created by taking the mean of the bin counts across these 100,000 random permutations.

Plate-based scATAC-seq

Plate-based scATAC-seq was performed as previously described⁶². Briefly, 17,000–50,000 viable PbTII cells were sorted using flow cytometry and pelleted by centrifugation. Supernatant was removed, and cells were resuspended in 50 μl tagmentation mix (33 mM Tris-acetate, pH 7.8, 66 mM potassium acetate, 10 mM magnesium acetate, 16% dimethylformamide (DMF), 0.01% digitonin and 5 μl of Tn5 from the Nextera kit from Illumina, cat. no. FC-121-1030). Cell/tagmentation mixture was then incubated at 37 °C, 800 r.p.m. on an Eppendorf thermomixer for 30 min. Tagmentation reaction was stopped by the addition of equal volume (50 μl) of stop buffer (10 mM Tris-HCl, pH 8.0, 20 mM EDTA, pH 8.0) followed by incubation in ice for 10 min. Then, 150 µl of DPBS/0.5% BSA was added to the mixture and nuclei suspension was then transferred to a polypropylene FACS tube. 0.0006% DAPI (in-house) was added for staining the nuclei prior to single-nuclei sort into 384-well plate (Eppendorf, cat. no. 0030128508) containing 2 µl of 2× lysis buffer (100 mM Tris-HCl, pH 8.0, 100 mM NaCl, 40 µg ml^–1 Proteinase K (New England BioLabs, P8107S), 0.4% SDS. Single-nuclei lysates were sealed and spun at 100g for 1 min, then were immediately frozen on dry ice and stored at −80 °C. Tn5 release and proteinase K digestion was then performed on a PCR machine (MJ Research Peltier Thermal Cycler) at 65 °C for 15 min followed by the addition of 2 µl of 10% Tween to quench the SDS. Five microliters of NEBNext High-Fidelity 2× PCR Master Mix (New England Biolabs, cat. no. M0541L) was then added to a preprepared 384-well plate (Eppendorf, cat. no. 0030128508) containing 1 µl of 10 µM i5 and i7 indexing primer mix (5 µM each) (Integrated DNA Technologies). The indexing primers were arranged in a combinatorial manner to allow multiplexing of up to 1,536 (4×384-well plates) sorted nuclei into one sequencing pool. The plate containing the PCR Master Mix and indexing primers was stamped onto the single nuclei, mixed briefly, sealed and spun at 100g for 1 min. Amplification was performed on a PCR machine (MJ Research Peltier Thermal Cycler) with 72 °C for 10 min, 98 °C for 5 min and 20 cycles of 98 °C for 10 s, 63 °C for 30 s and 72 °C for 20 s. The 384-well plate containing the PCR product was inverted into a VBLOK200 reservoir (Clickbio cat. no. CBVBLOK200-1) and spun at 200g for 1 min. The combined reactions were then transferred to a 5-ml Eppendorf tube, (Eppendorf, cat. no. 0030119401) and the DNA was purified and size-selected with an AMPure XP workflow (Beckman Coulter, cat. no. A63880). The purified pool was quantified on an Agilent Bioanalyser and combined with 3 additional pools, each containing 384 nuclei, to produce a sequencing library containing 1,536 single nuclei, and this was sequenced on 1 lane of an Illumina HiSeq 4000 instrument.

Single-cell ATAC sequencing and processing

Raw ATAC-seq reads were mapped to mouse genome MGSCv37 (mm9) using BWA–MEM (version 0.7.15). For each single cell, .bam reads with mapping quality below 30 and duplicated reads marked with Picard (version 2.19.0) were removed. For each time point, we aggregated the single-cell .bam files and marked and removed duplicated reads again. Peak calling was performed using MACS2 (version 2.1) with parameters --q 0.01 --nomodel --shift --100 --extsize 200 for each aggregated time point. The union of aggregated time point peaks (q < 0.01) was generated using bedops⁶³. To generate the full read-count matrix over the union of peaks for each single cell we used bedtools (version 2–2.29.0) coverageBed⁵⁰.

Quality-control metrics

For each single cell, the numbers of total reads sequenced, uniquely mapped reads, duplicated reads and mitochondrial reads were calculated. In addition, we calculated the fraction of reads in peaks and the fraction of open chromatin as previously described⁶².

Filtering cells and peaks

To identify poor-quality single-cell sequencing for each time point, we marked cells with log₂-transformed uniquely mapped reads counts, mitochondrial content and fraction of open chromatin were twice as high as the median absolute deviation. If a cell deviates twice from any of the three metrics, it was removed prior to analysis. Peaks were removed if they overlapped with the mm9 blacklist region generated by UCSC genome browser. Peaks mapped to regions present in less than 30 cells across all cells were removed.

Dimensionality reduction

The filtered coverage matrix was binarized and used as input for latent semantic analysis (lsa) followed by principle component analysis (PCA) using R package lsa (https://CRAN.R-project.org/package=lsa) (version 3.6.3). UMAP was generated using ‘RunUMAP’ from Seurat (version 3.0).

Motif in peaks fraction and motif variability

To assess the role of transcription factor (TF) motif accessibility during memory formation we utilized R packages Biostrings⁶⁴ (version 2.54.0), TFBSTools⁶⁵ (version 1.24.0) and transcription factor motif database JASPAR2016 (ref. ⁶⁶) (version 1.14.0). For each TF motif of interest, we annotated each peak if the underlying sequence contained the motif with a minimum matching score of 80%. Then for each cell, we counted the number of peaks with the TF motif. We then corrected the number of peaks with a motif by the number of total peaks of that cell. In addition, we assessed TF motif variability across all time points or selected time points of interest using chromVar³¹ (version 1.8) with default parameters. In total, we assessed 514 TF motifs from mouse and human from the JASPAR2016 TF motif database.

Peak annotation and enrichment analysis

Peaks were annotated with R package ChIPseeker (version 1.5.1) with transcription start site (TSS) reaching from −3000 to 3000 to the UCSC mm9 gene model and the ‘org.Mm.eg.db’ Bioconductor annotation package. Genes with TSS closest to peaks that were identified solely during peak calling of the aggregate of D0, D32 saline or D32 IAT were assessed for pathway enrichment using R package ReactomePA⁶⁷ (version 1.30.0).

Statistical analyses

Statistical analyses for all flow-cytometry-based protein-expression data were performed using Prism 7 (version 7.0c, GraphPad software). P values are shown as *P < 0.05, **P < 0.01, ***P < 0.001, ****P < 0.0001, unless exact P values are indicated.

Reporting Summary

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