Perfect adaptation of CD8+ T cell responses to constant antigen input over a wide range of affinity is overcome by costimulation

Maintaining and limiting T cell responses to constant antigen stimulation is critical to control pathogens and maintain self-tolerance, respectively. Antigen recognition by T cell receptors (TCRs) induces signalling that activates T cells to produce cytokines and also leads to the downregulation of surface TCRs. In other systems, receptor downregulation can induce perfect adaptation to constant stimulation by a mechanism known as state-dependent inactivation that requires complete downregulation of the receptor or the ligand. However, this is not the case for the TCR, and therefore, precisely how TCR downregulation maintains or limits T cell responses is controversial. Here, we observed that in vitro expanded primary human T cells exhibit perfect adaptation in cytokine production to constant antigen stimulation across a 100,000-fold variation in affinity with partial TCR downregulation. By directly fitting a mechanistic model to the data, we show that TCR downregulation produces imperfect adaptation, but when coupled to a switch produces perfect adaptation in cytokine production. A pre-diction of the model is that pMHC-induced TCR signalling continues after adaptation and this is confirmed by showing that, while costimulation cannot prevent adaptation, CD28 and 4-1BB signalling reactivated adapted T cells to produce cytokines in a pMHC-dependent manner. We show that adaptation also applied to 1st generation chimeric antigen receptor (CAR)-T cells but is partially avoided in 2nd generation CARs. These findings high-light that even partial TCR downregulation can limit T cell responses by producing perfect adaptation rendering T cells dependent on costimulation for sustained responses.

Primary human CD8 + T cells expressing the c58c61 TCR were stimulated using recombinant pMHC immobilised on plates with supernatant cytokine and surface TCR levels measured (see Materials & Methods). B) Cumulative TNF-α over the concentration of 9V (left) or 4A8K (right) pMHC for 1-8 hours. Mean and SD of 3 independent repeats. C) Data in panel B expressed as a rate of TNF-α secretion over time. D) Surface TCR expression measured using pMHC tetramers in flow cytometry for 4A8K (left) with a representative histogram (right). Mean and SD of 3 independent repeats. E) Recovery of surface TCR was measured by stimulating T cells for 4 hours to induce downregulation (black line) followed by transfer to empty plates without pMHC for 4 (blue) or 20 (red) hours before measuring surface TCR levels. The supernatant levels of MIP-1β, IFN-γ, and IL-2 along with raw data prior to averaging is summarised in Fig. S1-2 and single-cell intracellular cytokine staining in Fig. S3. but significant recovery in TCR surface expression on the timescale of ∼4 hours (Fig. 2E) suggesting that partial 95 downregulation is maintained by a balance of re-expression and antigen-induced downregulation. Taken together, 96 perfect adaptation cannot be explained by complete downregulation of the TCR. 97 It has also been shown that complete removal of the ligand can produce perfect adaptation (Fig. 1B). As already dis-98 cussed, the efficient removal of all pMHC ligands is not known to take place during T cell activation with previous 99 reports showing that pMHC ligands continually engage TCRs (16,17). Indeed, the removal of pMHC is unlikely to 100 be the mechanism in this experimental system because transferring T cells after they have adapted to plates newly 101 coated with pMHC did not reactivate them to produce cytokine (see below; Fig This mechanism predicts that increasing the antigen strength by increasing its concentration or affinity would induce further TCR downregulation and cytokine production. Therefore, the model was used to predict the outcome 125 of increasing the antigen concentration ( Fig. S5) or affinity (Fig. S6) and experiments confirmed that TCR levels 126 tuned to the new antigen strength with further cytokine production. As expected, reducing the antigen strength by 127 reducing antigen affinity did not lead to marked changes in TCR expression or further cytokine production (Fig. 128 S6D).

129
In summary, and in contrast to adaptation by other receptors, perfect adaptation can be explained here by imperfect 130 adaptation at the TCR by partial downregulation coupled to a switch in the pathway for cytokine production ( Fig.   131 1C).

132
T cell adaptation to constant pMHC antigen can be overridden by costimulation

133
The model predicted imperfect adaptation by TCR downregulation so that residual TCR output continued after 134 cytokine production had stopped (Fig. 3B). Given that T cells can encounter antigen in vivo with costimulation 135 through other surface receptors, and costimulation is thought to lower the signalling threshold for cytokine pro-136 duction (62-64), we determined whether costimulation can amplify residual TCR signalling to reactivate adapted 137 T cells.

138
We used the mathematical model to predict the outcome of transferring T cells from a first stimulation to a second 139 stimulation on the same antigen with or without costimulation (Fig. 4A). Note that in these transfer experiments, T 140 cells experience the same concentration of antigen in the first and second stimulation. The effect of costimulation 141 was simulated by lowering the threshold of the switch required for cytokine production and as expected, this 142 allowed T cells to produce cytokine provided they also continued to receive constant antigen stimulation (Fig. 4B).

143
In order to test whether CD28 costimulation could override adaptation, we stimulated T cells with the physio-144 logical affinity pMHC (first stimulation) before transferring them to the same titration of pMHC with or without 145 recombinant CD86, which is the ligand for CD28 (second stimulation). Consistent with the adaptation phenotype,

146
there was a dramatic reduction in TNF-α production in the second stimulation without CD86 but when CD86 was 147 present, strong cytokine production was observed (Fig. 4C). Importantly, T cells transferred to empty wells without 148 pMHC or to wells coated with only CD86 produced no cytokines.

149
In addition to CD28, the costimulatory receptor 4-1BB is also known to play an important role in the activation of 150 CD8 + T cells. We repeated the experiments with the recombinant ligand to 4-1BB showing that this TNFR is also 151 able to override adaptation but as with CD28, it critically relied on TCR/pMHC interactions (Fig. 4D).

152
Previous work on T cell anergy has described unresponsive T cell states that are induced when T cells are activated 153 in the absence of CD28 costimulation. We therefore tested whether CD28 costimulation can prevent adaptation.

154
We repeated the CD28 costimulation transfer experiments but now transferred T cells that were stimulated with 155 either pMHC alone or with both pMHC and CD86 in the first stimulation to a second stimulation that included 156 pMHC alone, CD86 alone, pMHC and CD86, or empty wells. We observed reduced cytokine production in the 157 second stimulation to pMHC alone, which was similar to empty wells, irrespective of whether CD86 was included 158 in the first stimulation (Fig. S8), suggesting that CD28 costimulation cannot prevent adaptation to constant antigen.

159
Taken together, these results indicate that perfect adaptation in cytokine production induced by constant pMHC 160 antigen stimulation does not lead to perfect adaptation in TCR signalling because extrinsic costimulation through 161 CD28 or 4-1BB can induce adapted T cells to produce TNF-α in a pMHC-dependent manner. This phenotype was 162 also observed for other cytokines ( Fig. S7-8).  showing that T cells were first stimulated for 8 hours before being transferred for a second stimulation for 16 hours with either antigen alone, costimulation alone, or antigen and costimulation. B) Predicted cytokine production by the mathematical model where costimulation is assumed to lower the threshold for the downstream switch. C) Representative TNF-α production when providing costimulation to CD28 by recombinant biotinylated CD86 and averaged E max values with SD from 4 independent experiments. D) Representative TNF-α production when providing costimulation to 4-1BB by recombinant biotinylated trimeric 4-1BBL and averaged E max values with SD from 3 independent experiments. Additional cytokines are shown in Fig. S7-8. Statistical significance was determined by ordinary one-way ANOVA corrected for multiple comparisons by Dunnett's test.

164
Given that CAR-T cells experience constant antigen stimulation in vivo, we analysed their adaptation phenotype.

165
To do this, we utilised the previously described T1 CAR (65) fused to the cytoplasmic tail of the ζ-chain (1st 166 generation CAR) that also recognises the NY-ESO-1 157−165 peptide on HLA-A2 in a similar orientation to the TCR 167 (K D = 4 nM (66)). These CAR-T cells were first stimulated with a titration of 9V pMHC before being transferred 168 for a second stimulation on the same titration of 9V (Fig. 5A). 169 We observed reduced cytokine production by CAR-T cells that experienced the antigen in the first stimulation 170 compared to CAR-T cells that were directly placed on the second stimulation (Fig. 5B, purple and orange lines, 171 respectively). Given that costimulation can override adaptation, we repeated the experiments with a 2nd generation 172 CAR containing the CD28 costimulatory domain finding that these CAR-T cells were able to partially avoid adap-173 tation (Fig. 5C). Compared to cytokine production in the 1st stimulation (100%), the production of TNF-α and 174 IL-2 were reduced in the second stimulation to 26% (p=0.0002) and 2.1% (p=0.002) in the 1st Generation CAR 175 but were only reduced to 58.7% (p=0.001) and 79% (p=0.0031) in the 2nd generation CAR (Fig. 5D-E). We note 176 that overall cytokine production was higher in the 2nd generation CAR (Fig. 5B-D) even though this receptor was 177 consistently expressed at lower levels ( Fig. S9).

178
Taken together, the adaptation phenotype observed with the TCR can also be observed with a 1st generation CAR 179 that can be partially overridden by a 2nd generation CAR that includes costimulation. Partial rescue in the 2nd 180 generation CAR is not unexpected because, unlike the complete rescue of the TCR by extrinsic co-stimulation ( Fig.   181 4), co-stimulation in the CAR is intrinsic and is therefore reduced over time as a result of CAR downregulation. : Adaptation is partially avoided in 2nd but not 1st generation CAR-T cells. A) Schematic of the experiment. T cells expressing the T1 CAR that recognises the 9V pMHC antigen were transferred to the same titration of antigen. B-C) Representative TNF-α and IL-2 production over antigen concentration from CAR-T cells expressing the B) the 1st generation variant containing only the ζ-chain and C) the 2nd generation variant containing the cytoplasmic tail of CD28 fused to the ζ-chain. D) Averaged E max values and SD for 3 independent experiments. E) Fold reduction of E max between the first and second stimulation for the 1st and 2nd generation CARs highlighting that 2nd generation CARs are more resistant to adaptation induced by constant antigen. Expression profile of both CARs and antigen-induced CAR downregulation is shown in Fig. S9. Statistical significance was determined by ordinary one-way ANOVA corrected for multiple comparisons by Dunnett's test.

183
Using a reductionist system to provide T cells with constant pMHC antigens, we observed that in vitro expanded 184 primary human CD8 + T cells do not maintain cytokine production but instead exhibit perfect adaptation across a 185 100,000-fold variation in affinity. This adaptation can be rescued by increasing the antigen concentration ( Fig. S5), 186 affinity ( Fig. S6), or when providing co-stimulation (Fig. 4).

187
Mechanism of adaptation. Adaptation by surface receptors has been termed state-dependent inactivation (5, 9), 188 which is an incoherent feedforward loop whereby ligand binding induces receptor signalling (positive regulation) 189 and receptor downregulation (negative regulation) (7). Perfect adaptation takes place if the receptor is completely inactivated or downregulated by the ligand. Unlike other receptors, the TCR is only partially downregulated in 191 response to antigen ligands leading to imperfect adaptation. To explain perfect adaptation in cytokine production, 192 an additional downstream mechanism is required, and in the present study we have invoked a switch (Fig. 3C).

193
However, other downstream mechanisms, such as additional IFFLs or NFLs, may also be able to convert imperfect 194 adaptation at the TCR to perfect adaptation in cytokine production. Consistent with imperfect adaptation at the 195 TCR, we found that extrinsic CD28 and 4-1BB costimulation can reactivate adapted T cells in a pMHC-dependent 196 manner (Fig. 4).

197
Although the downstream switch could be replaced by an IFFL or NFL, models where these motifs are responsible

206
The minimal model of TCR downregulation coupled to a downstream switch can produce bell-shaped dose-207 response curves (e.g. Fig. S6B). Previously, we argued that bell-shaped dose-response curves can be explained 208 by incoherent feedforward loops but not by TCR downregulation (47). The key assumption underlying this conclu-209 sion was that the rate of cytokine production was in the steady-state, which is the case for Jurkat T cell lines (47), 210 but the detailed analysis in the present work has revealed that it is not the case for primary T cells in the absence of the specific example of activated T cells, whose killing capacity is thought to be less dependent on costimulation, 218 perfect adaptation in cytokine production may be an important mechanism to ensure that their ability to initiate or 219 maintain inflammation is extrinsically regulated by other cells. In this way, perfect adaptation may serve to balance 220 functional immunity with excessive tissue damage.

221
Relation to in vivo studies. T cells are known to enter unresponsive states upon recognition of persistent self-and 222 viral-antigens in vivo (2, 3, 20, 31, 33-38, 67-71). While the underlying mechanisms that induce and maintain 223 these states are debated, their functional phenotype is characterised by transient cytokine production that can 224 be overcome by costimulation as observed here (Fig. 2,4). For example, it has been shown that effector CD4 + 225 T cells transiently produce cytokines despite continued antigen exposure (70) and the unresponsive (exhausted) 226 phenotype of CD8 + T cells induced by persistent antigen stimulation can be overcome by costimulation (71). 227 We note that the mechanism of adaptation that we report can take place with only minor TCR downregulation, with in vivo T cell phenotypes. As methods for the generation of large numbers of antigen specific human T cells 235 improve, it would be important to examine the response of natural quiescent T cell populations (e.g. naive CD8 236 and CD4 T cells) in this experimental assay.

237
Implications for adoptive therapy. The T cells we have used were generated using a protocol for adoptive therapy 238 with TCRs or CARs. Consistent with the adaptation phenotype we observe, it has recently been observed that CAR-239 T cells exposed to chronic antigen become unresponsive but can respond to a higher antigen dose, which correlates downregulation has yet to be explored.

Materials & Methods
Protein production. pMHCs were refolded in vitro from the extracellular residues 1-287 of the HLA-A*02:01 247 α-chain, β2-microglobulin and NY-ESO-1 157−165 peptide variants as described previously (47). CHO cell lines 248 permanently expressing the extracellular part of human CD86 (amino acids 6-247) with a His-tag for purification 249 and a BirA biotinylation site were kindly provided by Simon Davis (Oxford, UK). Cells were cultured in GS 250 selection medium and passaged every 3-4 days. After 4-5 passages from thawing a new vial, cells from 2 confluent 251 T175 flasks were transferred into a cell factory and incubated for 5-7 days after which the medium was replaced.

252
The supernatant was harvested after another three weeks, sterile filtered and dialysed over night. The His-tagged ) 1/f , where X 0 is the mean background gMFI of the TCR negative population.

319
After applying this transform, each repeat was normalised to the maximum gMFI before averaging independent 320 repeats.

321
The statistical analysis of maximum cytokine produced across different pMHC concentrations (E max ), was per-322 formed by expressing E max as a fold-change to pMHC alone before averaging independent repeats. Given that the 323 dose-response curves often exhibited a bell-shape, it was not possible to use a standard Hill function to estimate 324 E max . Instead, we used lsqcurvefit in Matlab (Mathworks, MA) to fit a function that was the difference of two Hill 325 curves in order to generate a smooth spline through the data from which the maximum value of cytokine was esti-326 mated. This procedure was used to extract E max values in Fig. 4, 5, S7, S8. In a limited number of cases, individual 327 outlier values were excluded prior to data fitting but are still shown as data points in respective figures.

333
Mathematical model. The mathematical model (Fig. 1C) is represented as a system of two non-linear coupled ordinary-differential-equations (ODEs), where R and O are the surface TCR levels (normalised to 1) and cumulative cytokine output with initial conditions 334 1 and 0, respectively. Given that TCR/pMHC binding kinetics (seconds) are faster than experimental timescales 335 (hours), the concentration of TCR/pMHC complexes (C, defined as receptor output) were assumed to be in quasi . Second, cytokine production was larger in the 2nd hour compared to the 1st hour (e.g. Fig. 2B), which 349 may be associated with transcriptional delays. To capture this delay, a multiplicative term in the equation for O 350 was introduced (H(t − t delay )) so that cytokine production was only initiated after a delay of t delay .  Using ABC-SMC, the values of R(t) and O(t) were directly fitted to the normalised surface TCR levels and 355 cumulative cytokine output, respectively, for the dose-response timecourse of 4A8K (192 data points in total). The 356 distance measure was the standard sum-squared-residuals (SSR) and all 9 model parameters were fitted (K 0 , n, 357 k 1 , k 2 , K 2 , k 3 , k 4 , K 4 , t delay ). A population of 3000 particles were initialised with uniform priors in log-space 358 and propagated through 30 populations by which point the distance measure was no longer decreasing. The final 359 population of 3000 particles, each of which had a different set of model parameters (Fig. S4), was used to display 360 the quality of the fit (Fig. 3A). Although the ODE model represents the reactions within a single cell, and hence the 361 dose-response curve for a single cell would exhibit a perfect switch (i.e. a step function), the population averaged 362 dose-response curves from the model include particles (i.e. cells) with different parameter values, accounting for 363 population heterogeneity, leading to a more gradual dose-response curve.

364
The posterior distributions revealed that only a subset of the model parameters were uniquely determined (Fig.   365   S4). Nonetheless, we were still able to make predictions using the model by simulating different experimental 366 conditions for the 3000 particles in the final population. To predict the effect of increasing antigen concentration 367 (Fig. S5B), the concentration of antigen was increased at the indicated times to the indicated value (no additional 368 parameters were required). To predict the effect of increasing antigen affinity (Fig. S6B), the TCR/pMHC binding 369 parameters in the model (K 0 and n) for each particle were reduced by 50%. To predict the effect of co-stimulation 370 (Fig. 4B), the threshold for the switch (K 2 ) for each particle was reduced by 60% for a duration of 8 hours. The  Figure S1: Cytokine production by T cells in response to constant stimulation by the supra-physiological affinity 9V pMHC. Expanded data from Fig. 2B showing TNF-α , MIP-1β, IFN-γ and IL-2 for individual repeats (6 left columns) along with averaged data (right column). The averaged TNF-α data is the same as in Fig. 2B. Error bars represent the SD. Cumulative IFN (normalised) Figure S2: Cytokine production by T cells in response to constant stimulation by the physiological affinity 4A8K pMHC. Expanded data from Fig. 2B showing TNF-α , IFN-γ and IL-2 for individual repeats (6 left columns) along with averaged data (right column). The averaged TNF-α data is the same as in Fig. 2B. Error bars represent the SD.   Figure S6: T cells adapted to a low-affinity antigen can be reactivated with a higher-affinity antigen. A) Schematic of the experiment showing that T cells were first stimulated for 5 hours on the low-affinity antigen before being transferred for a second stimulation of 5 hours with the same or different pMHC antigen (at the same antigen concentration). B) Predicted TCR surface expression (top) and cytokine production (bottom) for the transfer experiment by the mathematical model. C) TCR surface expression (top) and TNF-α production (bottom) with a detailed comparison performed at the indicated concentration (right). Consistent with the adaptation phenotype, a first stimulation with the low-affinity 4A8K pMHC (green) leads to reduced cytokine production in a second stimulation on the same pMHC (blue). However, transferring T cells to the higher-affinity 9V pMHC leads to further TCR downregulation and further cytokine production (orange). It is noteworthy that beyond the grey shaded region cytokine production is progressively reduced, which the model predicts is a result of lower levels of surface TCR prior to transfer to the higher affinity pMHC. Data are mean and standard deviation of 3 independent experiments with statistical significance determined by ordinary one-way ANOVA corrected for multiple comparisons by Dunnett's test. D) Individual repeats (left 3 columns) and averaged data (right column) showing TCR, TNF-α , and IFN-γ for all transfer conditions (see legend on right), including transfers from the high-affinity (9V) to the low-affinity (4A8K) pMHC showing that decreasing antigen affinity does not induce cytokine production (light purple). The averaged data in panel C is taken from the averaged data in panel D and shows a subset of all conditions tested.  Transfer T cells to the same pMHC +/-co-stimulation (re-suspend in fresh media) (1 hour rest) First stimulation (7 hours)

Second stimulation (16 hours)
Transfer T cells to the same pMHC +/-co-stimulation (re-suspend in fresh media) (  Figure S8: CD28 costimulation does not prevent adaptation to constant antigen stimulation. Primary human CD8 + effector T cells expressing the c58c61 TCR were stimulated with a concentration titration of the low-affinity pMHC antigen 4A8K and the supernatant concentration of TNF-α , IFN-γ and IL-2 measured by ELISA. T cells were first stimulated for 7h with the 4A8K low-affinity pMHC with or without recombinant CD86 (light blue lines). The T cells were rested for 1 hour in fresh medium before being transferred in fresh medium to pMHC alone (dark blue), pMHC and CD86 (red), CD86 alone (grey), or to empty wells (black). A) Schematic of transfer experiment with pMHC alone in the first stimulation (top) along with averaged E max values from four independent experiments for TNF-α , IFN-γ , and IL-2 (bottom). B) Schematic of transfer experiment with pMHC and CD86 in the first stimulation (top) along with averaged E max values from four independent experiments for TNF-α , IFN-γ , and IL-2 (bottom). All data is normalised to the value of E max at 7 hours with pMHC alone. Error bars represent the SD and statistical significance was determined by ordinary one-way ANOVA corrected for multiple comparisons by Dunnett's test comparing all conditions to the value of E max during the first stimulation.  Figure S9: Expression levels and antigen-induced downregulation of 1st and 2nd generation CARs. A) T cells expressing the 1st and 2nd generation T1 CAR were stained with 9V pMHC tetramers and CAR expression was measured by flow cytometry. B) CAR-T cells were stimulated with the indicated titration of the 9V pMHC antigen for 8 hours with surface CAR expression measured using pMHC tetramers. Data is average of at least 3 independent experiments and error bars represent SD.  Figure S10: A modified model where TCR surface levels do not contribute to shaping cytokine production, as a result of a highly sensitive switch, cannot explain re-activation of T cells to higher antigen strength. A) As discussed in the main text and shown experimentally (Fig. S6, S5), the proposed model of imperfect adaptation at the TCR by downregulation coupled to a downstream switch predicts that increasing antigen strength can reactivate adapted T cells. B) On the other hand, increasing the sensitivity of the switch (i.e. lowering the threshold, horizontal dashed line on receptor output) so that it remains in the on-state whenever antigen ligand is present makes TCR surface levels, and hence adaptation, have no impact on downstream signalling. Perfect adaptation in this model now requires a downstream negative feedback loop (NFL) or incoherent feedforward loop (IFFL). However, in this model, re-activation of T cells by increasing the ligand strength is not possible because the analogue information on ligand strength is filtered out by the sensitive switch.