To study the homeostatic dynamics of tissue-resident CD4⁺ memory T cells in healthy mice, we used in concert two genetic fate mapping tools in which cohorts of peripheral T cells and their offspring can be induced to express permanent fluorescent markers (Figure 1A). These reporter strains were previously used separately to study the turnover of naive and circulating memory T and B cells (Verheijen et al., 2020; Lukas et al., 2023; Bullock et al., 2024). In the Ki67^{mCherry-CreERT} Rosa26^RcagYFP system, henceforth Ki67-DIVN, fluorescent reporters are linked to the expression of Ki67, a nuclear protein expressed during cell division and for 3–4 days afterwards (Gossel et al., 2017; Miller et al., 2018). Specifically, these mice express both a Ki67-mCherry fusion protein and inducible CreERT from the Mki67 locus, together with a Rosa26^RcagYFP Cre reporter construct. Treatment of mice with tamoxifen therefore induces YFP in cells expressing high levels of Ki67, and YFP is then stably expressed by these cells and their offspring. Expression of Ki67-fused mCherry gives a constitutive live readout of Ki67 expression, independent of tamoxifen treatment and YFP expression. In the second fate reporter, CD4^CreERT Rosa26^RmTom mice (Cd4-FR), the Cre reporter is constructed such that cells expressing CD4 during tamoxifen treatment permanently and heritably express the fluorescent reporter mTomato (mTom).

In a closed population of cells at steady state, self-renewal must be balanced by loss and so, following tamoxifen treatment, the frequencies of cells expressing YFP or mTom within any such population would remain constant. Therefore, any decline in the frequency of either reporter among T_RM after treatment must derive from the influx of label-negative cells from an upstream (precursor) population. In the Ki67-DIVN mice, these will be descendants of cells that were not dividing at the time of tamoxifen treatment; in the Cd4-FR mice, labelled CD4⁺ cells will slowly be replaced by the descendants of those generated in the thymus after treatment. The shape of this decline will be determined by the combination of the net loss rate of T_RM from the tissue (through death, egress, or differentiation, offset by any self-renewal) and the label content of immigrant T_RM (Figure 1B). To refer to the persistence of individual T_RM cells, we will use the term ‘residence time’ rather than ‘lifespan’ to reflect the multiple potential mechanisms of loss from tissues.

To quantify these processes, we treated cohorts of both reporter mice, aged between 4 and 15 weeks, with a single 2 mg pulse of tamoxifen (Figure 1C). Over 9-week (Ki67 reporter) and 57-week (CD4 reporter) chase periods, we measured the frequencies of labelled cells among antigen-experienced CD4⁺ T cell subsets isolated from skin and the LP of the small intestine, and within circulating naive and memory T cell subsets derived from lymph nodes (Figure 1—figure supplement 1A). By combining these frequencies with measures of Ki67 expression and describing the resulting set of time series with simple mathematical models, we aimed to estimate the basic parameters underlying T_RM kinetics.

We considered two populations within both skin and LP, identified as tissue-localised by virtue of their protection from short-term in vivo labelling (‘Methods’ and Figure 1—figure supplement 1B). One was effector-memory (EM) phenotype (CD4⁺CD44^hi CD62L^lo) T cells in bulk, which we studied in order to gain the broadest possible picture of memory T cell dynamics at these sites. We also considered the subset of these cells that expressed CD69, a canonical marker of CD4⁺ T cell residency across multiple tissues (Szabo et al., 2019). We saw no significant changes with mouse age in the numbers of either population within skin (Figure 2A, p>0.67\begin{document}$p \gt 0.67$\end{document}) or LP (Figure 2B, p>0.39\begin{document}$p \gt 0.39$\end{document}). There were also no significant changes in any of these quantities with time since tamoxifen treatment (p>0.24\begin{document}$p \gt 0.24$\end{document}). We therefore assumed that the skin- and LP-localised T cell subsets we considered were at, or close to, homeostatic equilibrium during the chase period. For brevity, we refer to tissue-localised CD4⁺CD44^hiCD62L^lo T cells in bulk as EM, and their CD69⁺ subset as T_RM.

During the first few days after tamoxifen treatment, YFP and mTom expression increased continuously within the skin and LP subsets (Figure 2C), as well as among CD4⁺ naive (CD44^lo CD62L^hi), central memory (T_CM, CD44^hi CD62L^hi), and effector memory (T_EM, CD44^hi CD62L^lo) T cells recovered from lymph nodes (Figure 2D). These initial increases were driven in part by the intracellular dynamics of the induction of the fluorescent reporters. We therefore began our analyses at day 5 post-treatment, by which time induction was considered complete and the subsequent trajectories of label frequencies reflected only the dynamic processes of cell production and loss. A key observation was that mTom expression within the skin and LP subsets then declined slowly (roughly seven- to eight-fold over the course of a year, Figure 2C), indicating immediately that these populations were being continuously replaced from precursors. Early in the chase period YFP⁺ T_RM expressed Ki67 at higher levels than YFP⁻ cells, as expected, but Ki67 expression in the two populations converged at later times. We return to the interpretation of these kinetics below.

We then investigated the extent to which the simple model illustrated in Figure 1B could explain these trajectories. Given the observation of continued recruitment, any time variation in the label content of T_RM precursors might leave an imprint on the label kinetics of skin or LP T_RM themselves and thereby help us identify their developmental pathways. We reasoned that plausible T_RM precursors were LN-derived CD4⁺ naive, T_CM or T_EM; we also considered the possibilities that CD69⁻ cells within skin and LP are the direct precursors of the local CD69⁺ populations. Therefore, we used empirical functions to describe the time courses of the frequencies of YFP⁺ and mTom⁺ cells within these populations (Figure 2D) and used these to represent the label composition of cells entering the tissue subsets.

For each tissue subset (‘target’) and precursor pair, we fitted the model simultaneously to six time courses; the frequencies of (i) YFP expression and (ii) mTom expression among target cells, the proportions of Ki67^high cells among (iii) YFP⁺ and (iv) YFP⁻ target cells, and the (v) YFP and (vi) mTom expression kinetics within the precursor (‘Methods’ and Appendix 1). For each precursor/target pair, we considered three modes of influx – one in which new immigrant T_RM are Ki67^low (‘quiescent’ recruitment); another in which their Ki67 expression directly reflects that of the precursor (‘neutral’ recruitment); and a third in which immigrants have recently divided (Ki67^high), perhaps through an antigen-driven process (‘division-linked’ recruitment).

For each combination of target population, potential precursor, and potential mode of recruitment, we were able to estimate rates of influx, mean residence times, and mean interdivision times for the target population (Figure 3 and Appendix 1—table 1; prior and posterior distributions of the parameters of the best-fitting models are shown in Figure 2—figure supplement 1). The mean residence times of both EM and T_RM within skin and LP were ∼3 weeks and 2 weeks, respectively. The means of production of new cells differed at the two sites, however. In skin, around 2% of both populations were replaced daily by influx, comparable to the rates of constitutive replacement of circulating memory CD4⁺ T cell subsets (Gossel et al., 2017; Hogan et al., 2019; Bullock et al., 2024), and EM and T_RM self-renewed every 6 and 7 weeks, respectively. In contrast, within LP these subsets divided less often (every 7–9 weeks) and relied on higher levels of recruitment (4–6% per day) for their maintenance.

From these basic quantities, we could derive several other important measures of T_RM behaviour. First, the balance of the rates of loss (δ\begin{document}$\delta$\end{document}) and self-renewal (ρ\begin{document}$\rho$\end{document}) defines the persistence of a cohort of T cells, which is distinct from the lifespan of its constituent cells (Costa Del Amo et al., 2018; De Boer and Yates, 2023). Specifically, the quantity ln⁡(2)/(δ−ρ)\begin{document}$\ln(2)/(\delta-\rho)$\end{document} is the average time taken for a cohort and their descendents to halve in number. While we studied polyclonal populations here, this quantity applies equally well to measuring the persistence of a TCR clonotype, so we refer to it as a clonal half life (Bullock et al., 2024). The substantial rates of self-renewal in skin led to clonal half lives of just over a month. The lower levels of self-renewal and higher levels of replacement in LP resulted in shorter clonal half lives of approximately 2 weeks.

Importantly, our estimates of these quantities depended to varying degrees on the choice of precursor and mode of recruitment (Figure 3). For example, intuitively, given the observed level of Ki67 within each target population, the greater the levels of Ki67 within newly recruited cells, the less must derive from self-renewal within the tissue; hence, if one assumes that recruitment is division-linked, estimated division rates are reduced. Similarly, as discussed above, the label content of the precursor influences the net loss rate of label in the target, which was most clearly reflected in the loss of mTom⁺ cells over the longer chase period (Figure 2C and D). For example, mTom was lost most rapidly within naive CD4⁺ T cells (Figure 2D) due to the influx of label-negative cells from the thymus. Models in which naive T cells were the direct precursor of T_RM therefore predicted greater clonal persistence within tissues.

We saw very little decline in YFP expression levels during the 2-month chase period (Figure 2C) due in part to the sustained levels of YFP within the putative precursor populations (Figure 2D), which ‘topped up’ YFP-expressing T_RM. As a result, YFP kinetics within the tissues were not strongly informative regarding rates of replacement. However, the rate of convergence of Ki67 within YFP⁺ and YFP⁻ cells (Figure 2C) put clear constraints on the duration of Ki67 expression, which at approximately 3 days (Figure 3) was consistent with previous estimates. This quantity in turn was informative for estimating rates of self-renewal. Further, the observed convergence of Ki67 expresssion within these two subsets is consistent with the assumption of homogeneity in the rates of division and loss within skin and LP.

To more precisely quantify the kinetics of each target population, we assessed the relative support for each combination of precursor and mode of recruitment (Figure 4A). Each of these weights summarises the magnitude and uncertainty of a model’s out-of-sample prediction error of the label kinetics within the target population (‘Methods’).

We found that the data were quite strongly informative regarding the immediate ancestors of tissue subsets. From the candidate set of models, the weighting strongly favoured lymph-node-derived EM as the closest precursor to CD4⁺ EM within both skin and LP. However, within skin, local CD69⁻ cells were the favoured precursor to CD69⁺ T_RM. The evidence was generally more equivocal regarding the mode in which cells are recruited into skin and LP, although for skin we saw substantial evidence (66% of model support) for a division-linked transition from CD69⁻ to CD69⁺ cells. Figure 4B summarises the developmental trajectories and kinetics of T_RM in skin and LP that were supported most strongly by our analyses. Parameter estimates and credible intervals for these models are highlighted with vertical shaded regions in Figure 3 and are detailed in Appendix 1—table 1. Visual differences between models are shown in Figure 2—figure supplement 2, where for each target population we overlay the fits from top-ranked, second-ranked, and lowest-ranked models.

As a consistency check, when a population is at or close to steady state, bounds on the mean residence time of cells can be estimated using only the measured frequency of Ki67 within the target population, and the daily rate of replacement (Appendix 1 section ‘Estimating average cell residence times using Ki67’). In skin, both EM and T_RM are replaced at the rate of 2% per day and have Ki67 expression frequencies of around 0.15 (Figure 2C). The approximation then yields residence times in the range 22–28 days, depending on whether immigrant T_RM are Ki67^low or Ki67^high, respectively. In LP, with 5.5% daily replacement and Ki67 frequencies of around 0.07, we estimate residence times of 14–25 days. Both estimates are in good agreement with those from the model fitting. Further validation of these results, and dissection of the kinetics, might be achieved by manipulating cell trafficking, although this would potentially impact multiple processes at once. For example, treating mice with the sphingosine 1-phosphate receptor agonist FTY720 would block tissue ingress and egress. This would leave self-renewal as the only means of T_RM production and would also remove the component of the loss rate that is due to cells leaving the tissue. In principle, one could then gain estimates of the intrinsic lifespan of T_RM, rather than their tissue residence time. However, parameter estimation would then require accurate measurements of cell numbers within the tissue.

Ki67^{mCherry-CreERT} Rosa26^RcagYFP (Ki67-DIVN) and CD4^CreERT Rosa26^RmTom (Cd4-FR) mice have been described previously (Bullock et al., 2024). Experimental Ki67-DIVN mice were homozygous for indicated mutations at both the Ki67 and Rosa26 loci. Experimental Cd4-FR mice were heterozygous for the indicated mutations at both the Cd4 and Rosa26 loci. Tamoxifen (Sigma) was diluted to 20 mg/mL in corn oil (Fisher Scientific) and 100 μl (2 mg) was administered to mice via oral feeding on day 0. Ki67-DIVN mice were injected with 2 µg Thy1.2-BV510 (53-2.1) (BioLegend) 3 min prior to sacrifice to label T cells in the circulation. This protocol typically achieves >99% staining of circulating cells (Anderson et al., 2014), and less than 3% of cells recovered from our tissue samples were label positive (Figure 1—figure supplement 1B), supporting the assumption of very low rates of false-positive and false-negative events. Mice were subsequently taken down at specified timepoints post tamoxifen treatment for organ collection. All mice were bred in the Comparative Biology Unit of the Royal Free UCL campus and at Charles River laboratories, Manston, UK. Animal experiments were performed according to institutional guidelines and Home Office regulations under project licence PP2330953.

All peripheral lymph nodes (LNs), the small intestine (SI), and ear skin were taken from mice and processed into single-cell suspensions. LNs were mashed through two pieces of fine gauze in a Petri dish and washed with complete RPMI (Thermo Fisher) supplemented with 5% FCS (Thermo Fisher) (cRPMI). Cells were resuspended in cold PBS and counted using the CASY counter (Cambridge Bioscience). Peyer’s patches were excised from the antimesenteric side of the SI before being opened longitudinally and SI contents scraped out. SI pieces were placed in 20 mL pre-warmed extraction media (cRPMI + 10 mM HEPES [Thermo Fisher] +5 mM EDTA [Sigma]+1 mM DTT [Abcam]) and incubated in 37 °C shaking incubator for 30 min at 200 rpm. Cells were filtered over 70 μm cell strainer (supernatant containing intra-epithelial lymphocytes not used), and SI pieces were placed in cold cRPMI supplemented with 10 mM HEPES and allowed to settle. The supernatant was carefully poured off and SI pieces were finely minced, added to 20 mL pre-warmed digestion media (RPMI + 10% FCS + 1.5 mg/mL collagenase VIII [Sigma]) and incubated in 37 °C shaking incubator for 30 min at 200 rpm. After digestion, cells were passed through 70 μm cell strainer and washed with cRPMI + 10 mM HEPES. The resulting cell suspension contains cells from the LP of the SI. Ear skin was excised and separated into dorsal and ventral sides. Skin was finely minced, added to 4 mL digestion buffer cRPMI + 50 mM HEPES + 37.5 μg/mL Liberase TL [Merck] + 3.125 mg/mL collagenase IV [Thermo Fisher] +1 mg/mL DNAse I [Merck] and incubated in 37 °C shaking incubator for 2 h at 200 rpm. Cells were filtered over 70 μm cell strainer and washed through with cRPMI.

All cells isolated from skin and LP, and 5 × 10⁶ LN cells were stained for analysis by flow cytometry. Cells were stained in 100 μl PBS with combinations of CD8α-BUV395 (53-6.7), CD25-BUV395 (PC61), CD62L-BUV737 (MEL-14), TCRγδ-BV421 (GL3), CD103-BV786 (M290) (all BD Biosciences); CD25-BV650 (PC61), CD103-BV421 (2E7), CD8α-BV570 (53-6.7), TCRγδ-BV605 (GL3), NK1.1-BV650 (PK136), CD44-BV785 (IM7), CD45.1-BV605 (A20), CD4-BV711 (RM4-5), CD8b.2-APC (53-5.8) (all BioLegend); TCRβ-PerCPCy5.5 (104) (Cambridge Bioscience); CD44-APCef780 (IM7) (eBioscience); and CD3-APCef780 (2C11), CD3-biotin (145-2C11), CD69-PeCy7 (H1.2F3), CD45.2-AF700 (104), nearIR live/dead, blue live/dead, yellow live/dead (all Thermo Fisher). Cells were fixed for 20 min with IC fix (Invitrogen) and washed twice in FACS buffer (PBS + 0.1% BSA). Flow cytometric analysis was performed on either a Cytek Aurora spectral flow cytometer or a conventional BD LSR-Fortessa and analysed using FlowJo software (Treestar).

Cell numbers in LN were calculated by dividing the event count in a target population by the event count of live cells, multiplied by the total live cells in LN prep determined by CASY counter. Sizes of LP and skin populations were calculated using AccuCount (Spherotech) counting beads that were spiked into the sample prior to acquisition, as per the manufacturer’s instructions.

We fitted simultaneously the mathematical model illustrated in Figure 1B and described in Appendix 1 sections ‘The kinetics of mTom expression derived from Cd4-FR mice’ and ‘Modelling the label trajectories derived from Ki67-DIVN mice’, to the timecourses of frequencies of YFP⁺, Ki67^high in YFP⁺, Ki67^high in YFP⁻, and mTom⁺ cells in the target populations; and empirical descriptor functions describing the trajectories of the frequencies of YFP⁺ and mTom⁺ cells within precursor populations (all data shown in Figure 2C). We used a Bayesian estimation approach using Python and Stan (Stan Development Team, 2024) to perform these model fits. Code and data used to perform model fitting, details of the prior distributions for parameters, and figure generation notebooks are available at https://github.com/elisebullock/CD4TRM; copy archived at Bullock, 2025. Models were ranked based on the information criteria estimated using the Leave-One-Out (LOO) cross validation method (Vehtari et al., 2017). See Appendix 1 for full details.