Phenotypic Models of CAR T-Cell Activation Elucidate the Pivotal Regulatory Role of CAR Downmodulation

This article is featured in Highlights of This Issue, p. 761

The notable success of cancer-specific adoptive immunotherapies raised much interest in recent years, as these methodologies have shown durable clinical responses for treating hematopoietic malignancies and recently also for solid malignancies (1–3). These new immune-based methodologies, such as bispecific antibodies (bsAb) and chimeric antigen receptors (CARs), emerged as tools to redirect T cells to efficiently eliminate tumor cells (4, 5). Still, the limited outcome in solid malignancies highlights the need for better understanding of the molecular mechanisms that lead to optimal response and of how these mechanisms can be manipulated to rationally design such techniques.

CAR T cells are engineered to express hybrid receptors composed of extracellular high-affinity variable fragments of an antibody (scFv) linked to intracellular activatory signaling domains. Typically, scFvs with high affinity are selected; however, increased affinity might lead to nonbeneficial or reduced efficacy (6–9), or to reduced specificity (10, 11). Likewise, we now demonstrated that CAR T-cell response is nonmonotonic, and medium affinity CAR T cells outperformed high-affinity CAR T cells (12). We compared between the contribution of affinity, avidity, and antigen density in a unique and quantitative experimental system, and described the relation between these parameters. Antigen density also has a nonmonotonic behavior, whereas avidity had a monotonic behavior. Although each of these parameters seems to affect T-cell functionality separately, avidity seems to be related to antigen density, thereby increasing sensitivity and the magnitude of response.

The rational design of optimal CAR properties for adoptive cell therapy can be aided by a model of CAR T-cell activation that can predict responses of CARs of varying affinities and avidities to a range of antigen density. Classical T-cell receptor (TCR) activation that occurs upon engagement with its cognate antigen depends on the affinity, avidity, and antigen density, and consequentially initiates a signaling cascade that induces activation. Models that predict how the binding parameters determine T-cell response include regulatory elements such as kinetic proofreading, serial triggering, signaling network motifs, and receptor downmodulation (13–15).

Based on our phenotypical observations (12), we set to model CAR activation using classical mathematical TCR models; however, these models only partially explained CAR T-cell response. We therefore generated an improved dynamical model that addresses the number of receptors. This revised model suggests that receptor downmodulation provides an impetus for CAR T-cell regulation. To test this hypothesis, we measured and compared receptor downmodulation and intracellular inhibitory signaling processes. CAR downmodulation was evident rapidly following encounter with target cells. We observed similar kinetics of intracellular activatory and inhibitory processes, contrary to the kinetics described following TCR activation (16). Finally, CAR internalization inhibition, but not inhibition of regulatory phosphatases (Shp1/Shp2), increased CAR T-cell responses.

Unless otherwise mentioned, cells were cultured in RPMI 1640 medium containing 10% FBS, 100 units/mL penicillin, 100 μg/mL streptomycin, 0.5 mmol/L HEPES, 1 mmol/L sodium pyruvate, nonessential amino acids, and 2 mmol/L glutamine at 37°C in 5% CO₂.

Peripheral blood mononuclear cells (PBMCs) and T cells were cultured in T-cell culture medium (T-cell CM) that additionally was supplemented with 0.05 mmol/L 2-mercaptoethanol and the described units of IL2 at 37°C in 5% CO₂.

For peptide loading of APCs, cells were incubated with BioGro medium, which contained RPMI 1640 medium, 2% BioGro (Biological Industries), 100 units/mL penicillin, 100 μg/mL streptomycin, 0.5 mmol/L HEPES, 1 mmol/L sodium pyruvate, nonessential amino acids, and 2 mmol/L glutamine at 37°C in 5% CO₂.

For drug inhibition of actin polymerization, cells were incubated with either 5 μmol/L or 1 μmol/L Cytochalasin D (CytD, Sigma Aldrich) or control 0.1% DMSO in T-cell CM for 45 minutes prior to beginning of functional assay.

For drug inhibition of Shp1/Shp2 phosphatases, cells were incubated with either 20 or 10 μg/mL of freshly dissolved Sodium Stibogluconate (SSG, Cayman) in T-cell CM for 16 hours prior to beginning of functional assay.

Unless otherwise mentioned, all media and cell-growth reagents were purchased from Thermo Fisher Scientific. Oligonucleotides were manufactured at high-performance liquid chromatography quality by Sigma-Aldrich. All peptides were ordered from LifeTein.

The following commercial antibodies were used in this work: Anti-mouse Ig kappa light chain (BioLegend), Anti–HLA-A2 (clone BB7.2, Bio-Rad), Anti-CD8 (BD Biosciences), Anti-CD3 (clone SK7, BD Biosciences), Anti-CD3 (clone OKT3, eBioscience), Anti-IFNγ (eBioscience), Anti-IL2 (eBioscience), Anti-TNFα (eBioscience), Anti-Biotin (BioLegend), Anti-CD19 (eBioscience), Anti–p-LAT (pY-226, BD Biosciences), Anti–p-Zap70 (pY-292, BD Biosciences), Anti–p-Slp76 (pY-128, BD Biosciences), Anti–p-cCbl (pY-700, BD Biosciences), and Anti–p-Shp2 (pY-542, BD Biosciences).

The TCR-like antibodies described in the work were received from Adicet Bio and were previously characterized (12).

Single-chain MHC–peptide complexes were produced by in vitro refolding of inclusion bodies produced in Escherichia coli upon isopropyl β-d-1-thiogalactopyranoside induction, and MHC tetramers and monomers were generated using labeled streptavidin, as described previously (12, 42).

Peptides were ordered from LifeTein and dissolved in DMSO. Epstein-Barr virus–transformed JY B-cell lymphoblastoid lines were incubated overnight with the indicated peptides and concentrations in BioGro medium at 37°C as described previously (12). After pulsing, the cells were washed to remove free peptide.

Primary PBMCs were obtained from donors, after obtaining their informed consent according to the Technion Institutional Review Board, using density-gradient centrifugation according to the Ficoll-Hypaque technique. CAR constructs described previously (12) were transduced to primary T cells, using the Phoenix amphotropic packaging cells, as described previously (7, 12).

pMHC and receptor were analyzed using TCR-like Fabs or pMHC monomers, respectively, as described previously (43).

Transduced CAR T cells and JY cells loaded with the mentioned peptides were added at 1 × 10⁵ cells/well at 1:1 effector-to-target (E:T) ratio (effector cell number was calculated as the number of live Tyr-MHC pentamer-positive cells). Cells were cultured in sterile 96-well plates in 100 μL of T-cell CM containing brefeldin A and monensin(eBioscience). After 5.5 hours at 37°C with 5% CO₂, cells were fixed and permeabilized according to the manufacturer's protocol (Cytofix/Cytoperm Kit; BD Biosciences), and intracellular cytokines staining (ICS) was performed as described in the text. Anti-CD8 or anti-CD3 antibodies were used to gate effector cells from target JY cells. Finally, cells were washed and analyzed by an LSR II flow cytometer (BD Biosciences). The transduced T cells were gated by size and granulation (FSC/side scatter), GFP-positive, and CD3/8-positive cells, as demonstrated in Supplementary Fig. S5B. Reactivity was assessed by measuring percentage of reactive cells and the median fluorescence intensity (MFI) of the reactive cells (reactive-MFI).

We used a phenotypical model, composed of ordinary differential equations reflecting fast processes from receptor–ligand interaction to intracellular signaling cascades, and a slow receptor downmodulation process (see Supplementary Information for detailed model development). As the time scale for the downmodulation (minutes) is much larger than the time scale (seconds) of the receptor–ligand and intracellular signaling reactions, we used a quasi-steady-state approximation where the receptor–ligand and intracellular signaling reactions are assumed to reach local steady state, whereas the downmodulation of receptors evolves adiabatically. The final model is thus solved for the steady-state solution of receptor–ligand and intracellular signaling reactions:

where R and L reflect the number of receptors and ligands, respectively; |\ {C_N}$|⁠, Y, and P reflect the signaling receptor–ligand complexes, the activatory mediators, and the final cytokine product, respectively; and |{k_{{a}}},\ \lambda ,\ \mu ,\ \alpha ,\ {\rm{and}}\ \gamma \ $| reflect activation and deactivation constants (see Supplementary Information). The final model consists of these steady-state solutions together with the evolution equation for receptor downmodulation:

where |\chi $| is the downmodulation rate, and ϵ is a small parameter that represents the slow downmodulation time scale.

The model equations are solved using Matlab program (Mathworks) and are available upon request. Graphs show activity (P) as a function of log(L) for each different receptor affinity, for initial number of receptors of 70,000 or 700,000 per cell, corresponding to low and high CAR expression levels.

The model calculations in Fig. 1B were generated using parameters of k_a = 1, γ = 1, α = 100, λ = 1, P_T = 10,000, Y_T = 100, and k_p as indicated. For Fig. 1C–F, the same parameters were used with k_p = 0.05. In Fig. 1E and F, we used χ_F = 0.35. All other parameters (μ, δ, and χ_L) were used as indicated in Fig. 1C–F. We note that parameters of χ_F = 0.35 and χ_L = 0.005 [χ = 0.35 + 0.005 × (L)] gave similar downmodulation rates to the experimental data (Supplementary Fig. S3E). We used the k_off and k_D values calculated from the Fab surface plasmon resonance experiments (12). Because the variables L and R are given in terms of numbers (per cell), whereas k_D values are in (M) units, there was a need to convert k_D to be dimensionless. In the code, we used k_D values that give reasonable and comparable outcomes (using the original k_off values), thus maintaining the differences between affinities. To achieve this, we multiplied the original k_D by a factor of 10⁴. This multiplication was carried on all k_D values.

CAR T-cell cytotoxicity was measured using conventional [35S] methionine release assay. Target melanoma cell lines 1938, 624.38, and 501A cells were incubated with 12 μCimL [35S] methionine overnight, washed, and seeded at 5 × 10³ cells/well coincubated with CAR-positive T cells at 10:1, E:T ratio for 7 hours, and the concentrations of [35S] methionine released from target cells were assayed. The percent-specific cytotoxicity was calculated as follows:

Relative cytotoxicity was determined as: (specific lysis/maximal specific lysis) × 100.

CAR downmodulation was detected similarly to previously described methods, with some changes to prevent target cell interference with pMHC monomer staining (44). Briefly, 10⁵ peptide-loaded JY cells were washed and added to each Eppendorf tube at 10⁶ cells/mL in complete T-cell CM. 10⁵ CAR T cells were then added to each Eppendorf tubes at 10⁶ cells/mL, shortly centrifuged, and incubated at 37°C for 5 minutes. Receptor downmodulation was terminated by the addition of a double volume of ice-cold FACS buffer. All following procedures were done at 4°C. The contents of each tube were then removed, pelleted at 3,000 rpm for 1 minute, and resuspended in 100 μL of ice-cold FACS buffer. As JY:T-cell interaction could interfere with MHC staining, JY cells were removed prior to staining using negative selection with Dynabeads utilizing the CELLection pan mouse IgG kit conjugated to anti-CD19 Ab (HIB19), according to the manufacturer's instructions. The cells were stained with excess monomer–MHC–PE for 30 minutes prior to CD8-APC and CD3-PerCP staining for 15 minutes, and then washed and analyzed at the Calibur flow cytometer (BD Biosciences). CAR expression separation was determined posteriori according to GFP signal (which is correlative to CAR expression levels), gating on low, medium, and high GFP signal, as demonstrated in Supplementary Fig. S5D.

To calibrate a time point sufficient for detection of phosphorylation induction, and the antibodies concentrations to be used, we initially determined the maximal phosphorylation levels of antibodies against p-Zap70, p-LAT, p-Slp76, p-cCbl, and p-Shp2 (BD Biosciences) using flow cytometry. CAR T cells were activated using saturated levels of the anti-CD3 antibody (clone OKT3) for different incubation periods, and cells were stained at different antibody concentrations. Once the setting was calibrated, phospho-flow staining experiments were conducted similarly to previously described methods (16, 33), with the addition of labeled pMHC staining in order to separate by avidity. Briefly, 1.5 × 10⁵ CAR T cells and 1.5 × 10⁵ peptide-loaded JY cells were cultured in serum-starved medium prior to experiment beginning. CAR T cells were then treated for 2 minutes at 37°C with either coculturing with the JY cells, stimulating with OKT3 anti-CD3ϵ antibody (5 μg/mL), or left unstimulated. The reaction was stopped immediately by addition of an equal volume of ice-cold FACS buffer containing pMHC monomers (10 μg/mL) and phosphatase-inhibitor cocktail (PhosSTOP, Sigma), to allow surface staining of CARs, for 5 minutes on ice. Following washing, cells were fixed and permeabilized according to the manufacturer's protocol (BD Phosflow) and stained with streptavidin-BV421 (BioLegend), anti-CD8 antibodies, and BD Phosflow antibodies, as described in text. Finally, cells were washed and analyzed by an LSR II flow cytometer (BD Biosciences). CAR T cells were gated by size and granulation (FCS/side scatter), GFP-positive, pMHC-positive, and CD8-positive cells. CAR expression separation was determined posteriori according to pMHC staining signal, gating on low, medium, and high pMHC staining signal, as demonstrated in Supplementary Fig. S5C. Relative phosphorylation for each phospho-protein was determined as [(MFI – minimal MFI) / (maximal MFI – minimal MFI)] × 100, whereas minimal and maximal MFIs were the lowest and highest MFI values for each experiment.

Trogocytosis flow cytometry was analyzed by measuring antigen transfer using antigen-specific staining. CAR T cells (5 × 10⁴) that were treated with 0.3, 1, or 5 μmol/L CytD inhibitor or with DMSO cocultured with melanoma cells or loaded JY cells at 1:1 E:T ratio in V-shaped 96-well plates. Unless specified differently, cells were incubated for 15 minutes at 37°C. Antigen transfer was measured by staining of cells with the high-affinity TCR-like Abs that recognize specifically the Tyr-HLA-A2 antigen, followed by labeling with secondary antibodies and measured by flow cytometry. Antigen transfer was determined by calculating the MFI of the effector CAR⁺ cells (by gating on GFP⁺ cells). Initial tests were carried on in order to compare whether the target cell presence can alter the staining of the antigen on the effector cells.

Trogocytosis imaging was analyzed using antigen-specific staining. CAR T cells (5 × 10⁵) that were pretreated with 1 or 5 μmol/L CytD inhibitor or with DMSO for 45 minutes were washed and cocultured with loaded JY cells at 1:1 E:T ratio in V-shaped 96-well plates and incubated for 15 minutes at 37°C. Antigen trogocytosis was measured by staining cells with the high-affinity TCR-like Abs that recognize specifically the Tyr-HLA-A2 antigen, followed by labeling with secondary antibodies and anti-CD8 antibody for surface staining. Samples were seeded with 1% agar-PBS in 15μ-Slide VI 0.4 chambers (IBIDI) and imaging at confocal microscope, and z-stacks were taken at optimal imaging parameters with a LSM710 confocal microscope with 40× water immersion objective (Carl Zeiss Microimaging). IMARIS software was used to generate the figures. Antigen trogocytosis was determined by calculating the MFI on the surface of effector CAR⁺ T cells, determined by CD8⁺ and GFP⁺ cells.

FACS data were analyzed using FCS Express 5 (De Novo). MFIs were calculated after subtraction of the background staining calculated from isotype-control staining (for primary conjugated antibodies) or from secondary antibody staining (for nonconjugated primary antibodies). Reactive-MFI was measured in a similar fashion on positive cells, with a threshold of at least 3% positive cells. In case of less than 3%, the reactive-MFI was determined as zero.

A standard principal component analysis (PCA) analysis was performed using the Perseus software (45). The MFI data matrix was arranged such that each CAR T-cell line (e.g., CAR T of 16 nmol/L affinity with high expression) was represented in rows and the averages of all markers p-Zap70, p-LAT, p-Slp76, p-cCbl, and p-Shp2, or on four markers excluding p-LAT. Following replacement of negative values by zeros, the matrices were log transformed, normalized, and then analyzed by PCA.

All statistical analyses of the described measurements (MFIs, reactive-MFIs, percent-reactive cells, receptor downmodulation, and relative killing) were performed using The R Project for Statistical Computing (46). P values were corrected for FDR using the Benjamini–Hochberg procedure (47). According to the correction, for the ICS analysis, P values under 0.021 were considered significant; for the phospho-flow analysis, P values under 0.023 were considered significant; for the downmodulation analysis, P values under 0.023 were considered significant. Significant values are represented with stars as follows: *, 0.05–0.01; **, 0.01–0.001; ***, <0.001 adjusted values. A three-way ANOVA (CAR expression × CAR affinity × antigen density) was performed on MFI (phosphoflow assay), reactive-MFI (ICS assay), or percent-reactive cells (ICS assay) of the CAR T cells, using the aov function. For two-way ANOVA calculation of the variance explained by affinity or CAR expression, we divided each effect sum of squares by the total sum of squares.

Statistical analysis of cytotoxicity was performed using The R Project for Statistical Computing. A two-way ANOVA (CAR affinity × tumor cell line) was performed using the aov function, followed by Tukey Honest Significant Differences (TukeyHSD).

Statistical analysis of receptor downmodulation at coincubation with 100 μmol/L peptide-loaded JY cells was performed using The R Project for Statistical Computing. A one-way ANOVA (drug treatment) was performed on MFI, using the aov function, followed by TukeyHSD with FDR correction using the Benjamini–Hochberg procedure.

Statistical analysis of ICS following drug treatment (reactive-MFI and percent-reactive) was performed using The R Project for Statistical Computing. Two-way ANOVAs (drug treatment x antigen density) were performed using the aov function, followed by TukeyHSD with FDR correction using the Benjamini–Hochberg procedure.

We have previously generated an experimental system in which CAR affinity, avidity and target cell's antigen density can be manipulated. This system revealed several key features regarding the effect of affinity, avidity, and antigen density on CAR T-cell function (Fig. 1A):

• A bell-shaped response as a function of antigen density.

• Different affinities had similar antigen densities inducing peak response.

• Peak response was different between affinities, with mid-affinities yielding the highest responses.

• High avidity induced a higher response.

• High avidity had a wider window of maximal response.

To explain these findings from a mechanistic perspective, we mathematically modeled T-cell activation. Although the T-cell response is an intricate network of many elements, it can be phenotypically modeled through fewer parameters representing the complex signaling cascade (13). Based on previous approaches to formulate a mathematical model of native TCR-induced activation and response (13, 18), we tested several models while gradually increasing the model's complexity. We started with a kinetic proofreading model. This model considers the antigen (ligand, L) and CAR (receptor, R) binding kinetic constants (k_on and k_off). The complex that is formed upon binding (C₀), however, does not immediately trigger signaling, and only after sufficient binding, the complex becomes a signaling complex (C_N; k_p is the kinetic proofreading constant) that induces a downstream signaling cascade and consequentially cytokine production (P). Higher k_p reduces the difference between sensitivities of the various affinities (Fig. 1B). This model however does not predict a ligand-dependent bell-shaped curve. We therefore tested a model incorporating also an intracellular incoherent-feedforward (IFF) motif (Fig. 1C), similarly to previous works (14, 17), to reflect the inhibitory intracellular signaling molecules (such as Shp1 and Shp2 phosphatases) that can induce ligand-dependent inhibition (14, 16). As expected, increasing the inhibition constant (μ) produced lower responses at high ligand concentrations. Yet, this model predicts highest sensitivity for the highest affinity receptor (4 nmol/L affinity, blue lines in Fig. 1). High-affinity–induced inhibition was described in TCRs. High-affinity interactions are thought to prevent continuous signaling due to slower dissociation constants (13). This assumption, the limited signaling model, assumes a state of a nonsignaling complex (C_N+1, formed by a δ constant from C_N; Fig. 1D). Incorporating this element into the kinetic proofreading with IFF model (KPL-IFF) predicts reduced responses for high affinities and for high antigen densities as well. Nonetheless, the model was negated, as the peak amplitude of cells with high receptor expression (700,000 receptors per cell) is similar to cells with low expression level (70,000 receptors per cell), consistent with previous models that compared receptor expression levels (17). Equally, none of the above-described models displayed different magnitudes of response as a function of receptor expression levels. We also note that all the described models show complete symmetry between the receptor and ligand. Because the numbers of receptors per effector cell are several orders of magnitude higher than the number of antigens (ligands) per target cell, it stands to reason that further increasing receptor number cannot increase activity, as all ligands are already fully occupied. We therefore examined a model without R and L symmetry, in which T-cell responses were more dependent on avidity. Previous works have shown that both TCRs and CARs are internalized and degraded following target cell encounter, thus regulating T-cell activation (18–21). We created a model that incorporates a slow downmodulation of the receptor by a constant χ, which increases as a function of the number of ligands. This model predicts higher response when receptor expression increases. Surprisingly, due to the ligand-induced receptor downmodulation, the response at high ligand concentrations is reduced, thus creating a bell-shaped curve, even without the IFF motif (Fig. 1E). Moreover, this model now predicts a similar ligand concentration that induces peak response across all different affinities, and induces a wider window of high response in high avidity cells, corresponding to our experimental data. Integrating all the above-described elements (proofreading, limited signaling, IFF, and receptor downmodulation), the model predicts a bell-shaped response as a function of antigen density, and as a function of affinity, with similar antigen densities between affinities that induce maximal responses, where high avidity contributes to a higher and wider maximal response (Fig. 1F). This model describes best our observed data. We note that in both downmodulation-incorporating models (Fig. 1E and F), the inhibitory part of IFF had negligible effect on the outcome.

Our model predicted the antigen-dependent bell-shaped curve to be mostly dependent on CAR downmodulation, whereas works on TCR signaling attribute this feature mainly to intracellular inhibitory molecules, such as Shp1 and Shp2 (16). To address this discrepancy, we examined both processes, looking at the differential phosphorylation pattern of known proximal TCR signaling molecules, and at ligand-dependent receptor downmodulation. First, we assessed the level of phosphorylation of three activatory signaling molecules (p-Zap70, p-LAT, and p-Slp76) and two inhibitory signaling molecules (p-cCbl and p-Shp2) using flow cytometry. We incubated CAR T cells of different affinities (k_Ds of 4, 16, 35, and 434 nmol/L) with peptide-loaded APC target cells for 2 minutes, and analyzed the relative phosphorylation of each molecule (i.e., normalized by the maximal phosphorylation). We used unstimulated T cells for basal activation measurements, and anti-CD3 Ab (OKT3) for CAR-independent activation measurements. We post hoc separated the cells by expression level, gating according to CAR surface staining. All five phosphorylated molecules showed an antigen-dependent increase in phosphorylation, reaching highest phosphorylation at the highest antigen density. The averages of relative phosphorylation for representative signaling molecules are shown graphically in Fig. 2 arranged either by expression level in each individual CAR affinity (Fig. 2A and C) or by their avidity (Fig. 2B and D; see Supplementary Fig. S1 for the remaining signaling molecules and separated graphs for each CAR affinity, and Supplementary Table S1 for statistical analysis). The phosphorylation seemed to depend mainly on receptor avidity, and less on receptor affinity, as seen when results are arranged by avidity (Fig. 2B and D). Higher avidity cells displayed higher phosphorylation level (including higher phosphorylation to nonspecific antigens, as seen for unloaded target cells).

Our previous work on antigen-induced inhibition in TCRs revealed that activatory signaling molecules respond sharply at low and medium antigen densities, whereas inhibitory signaling molecules respond gradually, reaching maximal phosphorylation only at high antigen densities (16), thus signal summation of these modules creates a bell-shaped response of the T cells. However, when we compared activation of the inhibitory and the activatory molecules, both modules responded gradually to antigens, reaching maximal phosphorylation only at the high antigen density, and with no detectable gap between the two modules (Fig. 2E–H). Although we clearly see activation of the inhibitory molecules (reinforcing the existence of an IFF motive), we cannot attribute the antigen-dependent bell-shaped curve to a gap between the activatory and inhibitory modules.

Next, we applied a systematical analysis to assess the contribution of affinity and avidity to proximal signaling responses, by applying a two-way ANOVA for each antigen density, and analyzed the percentage of variance attributed to each effect (the effect's sum of squares) out of the total variance (the total sum of squares), thereby denoting the percentage of variance they could explain for each antigen density. As observed above, apart from the p-LAT molecule, the level of phosphorylation was strongly dependent on receptor avidity (Fig. 3A–C). p-LAT showed high dependence on avidity only at high antigen densities. We then applied PCA on all five phosphorylated molecules over all affinities and avidities (Fig. 3D). The first principal component resembles avidity and corresponds to more than 70% of the variance. The second component seems to reflect the different affinities. As the phosphorylation rates of p-LAT displayed different trends from all other four molecules (Supplementary Fig. S1G), we repeated the PCA, this time excluding the p-LAT molecule (Supplementary Fig. S2F). Now the avidity component explained more than 80% of variance. The first principal component strongly correlated with CAR expression levels (Fig. 3E).

Whereas CARs are activated solely by their antigen targets, the native TCR requires the recruitment of additional coreceptors and the formation of the TCR complex, including the various CD3 subunits (21). We were therefore surprised that the OKT3 antibody, which activates TCR complexes and not CARs, induced discrepant levels of phosphorylation between different CAR affinities and avidities (Supplementary Fig. S2A–S2E). The deviations across different avidities and affinities to OKT3 activation resembled differences at the basal activation state. As basal TCR or CAR activity can tune cell function and sensitivity toward its cognate antigen (8, 22, 23), we compared the basal level of phosphorylation with the PCA's first principal component that describes most of the variance in phosphorylation, and detected a strong correlation to the basal level of p-Zap70, p-Slp76, and p-cCbl phosphorylation levels (Supplementary Fig. S2G). This suggests that the differences in phosphorylation following an antigen encounter can be predicted by the basal phosphorylation level prior to that encounter. When comparing phosphorylation levels between activated and resting cells, high Pearson correlations are observed (Supplementary Fig. S2H). Because the basal level of phosphorylation can be affected by avidity (24), it is yet to be concluded which parameter is causative for the differential phosphorylation levels. In aggregate, we demonstrated a rapid antigen-dependent activation of both inhibitory and activatory modules, which is correlated with basal activation levels and avidity; however, it could not explain signal summation that results in nonmonotonous behavior.

Next, we set out to characterize receptor downmodulation occurring immediately after receptor and ligand contact. CAR T cells that were shortly incubated (5 minutes) with peptide-loaded APCs target were assessed for the surface expression of CD3, CD8, and the CAR. Surprisingly, we detected a significant CAR downmodulation (i.e., reduction in CAR surface staining) even after this short incubation period (Fig. 4A; Supplementary Fig. S3D). The amount of CAR downmodulation was higher when initial CAR expression was higher. All CARs, regardless of affinity, showed minor CAR downmodulation in response to nonspecific antigens (unloaded APCs). Smaller changes were detected in both CD8 and CD3 coreceptors (Supplementary Fig. S3A and S3B). As the CAR expression is proportional to GFP expression (Supplementary Fig. S3C), we post hoc gated each CAR population into three CAR expression levels according to GFP expression level (Fig. 4B), consequently highlighting the dependence of CAR downmodulation on initial CAR expression levels. In fact, the amount of CAR downmodulation is proportional to the initial CAR expression (Fig. 4C). In other words, the fraction of downmodulated CARs depends solely on antigen density, whereas the total amount of downmodulated CARs depends also on the initial number of receptors. CAR affinity had no evident effect over receptor downmodulation. CAR downmodulation occurred rapidly, and no further downmodulation was observed at longer incubation times (Supplementary Fig. S3D). Taken together, we examined and detected both receptor downmodulation and intracellular proximal signaling processes. Both processes occurred rapidly and were dependent on CAR expression levels, and are plausible mechanisms that can result in antigen-dependent nonmonotonic behavior.

Finally, to determine which mechanism is associated with the hyporesponsiveness at high antigen densities, we compared functional avidity of CAR T cells following drug-induced Shp1/Shp2 phosphatase inhibition (using SSG; ref. 25) or drug-induced internalization inhibition (using the actin polymerization inhibitor CytD; ref. 26). As actin depolymerization can interfere with immunologic synapse formation and T-cell activation (27, 28), we worked with the lowest CytD concentrations which could detectably inhibit antigen-induced CAR downmodulation (Fig. 5A; Supplementary Fig. S4A; Supplementary Table S2). We examined CAR downmodulation of the 35 nmol/L CAR following CytD treatment, or treatment with 10 or 20 μg/mL SSG, concentrations that inhibit both Shp1 and Shp2 activities (25). The CAR surface expression of treated cells was compared with surface expression of the unstimulated (i.e., without target cells encounter) control cells (Supplementary Fig. S4A), or with each treatment's unstimulated surface expression (Fig. 5A). We noticed that all treatments reduced CAR expression in the basal unstimulated state, thus lowering initial avidity of the CAR T cells; however, cells treated with CytD had lowered antigen-induced CAR downmodulation, in a drug dose–dependent manner. Alongside downmodulation, trogocytosis is a process that depends on actin polymerization and even on receptor internalization, and therefore can also be the cause of the hypofunctionality at high antigen densities due to antigen removal or even fratricide. CAR T-cell fratricide could not be detected using reciprocal killing assay. Nonetheless, using antigen-specific staining, we observed trogocytosis that occurred rapidly following encounter with peptide-loaded cells (Fig 5B–D; Supplementary Fig. S4A–S4C; Supplementary Tables S5 and S6). However, this phenomenon is evident at the higher antigen densities, and was reduced following CytD treatment (Fig 5B–D), but not following SSG treatment (Supplementary Fig. S4C).

Next, we investigated the antigen-induced nonmonotonous curve following treatment with these drugs, by measuring production of the IFNγ, TNFα, and IL2 cytokines (Fig. 5E–G; see Supplementary Fig. S4B–S4D for percent-reactive cells and Supplementary Table S3 for statistics). Surprisingly, we detected only minor changes following SSG treatment, however with a trend of increased response at high antigen densities. CytD-treated cells responded differently for the two close concentrations. CytD treatment (1 μmol/L) leads to higher production of the three cytokines, resembling the effect of increasing CAR expression levels, whereas 5 μmol/L CytD treatment showed similar or even lower cytokine production, along with lower percentages of responding cells (Supplementary Fig. S4D–S4F). This might be attributed to interruption of synapse formation. These data suggest that reducing internalization and trogocytosis processes and increasing CAR expression level have a stronger effect than interfering with the intracellular inhibitory signaling module. We speculate that there is a narrow window of CytD concentrations that can interfere with trogocytosis and internalization feedback mechanisms without significantly inhibiting synapse formation. Nonetheless, this treatment could not prevent a bell-shaped response, which corresponds to the observation that CAR downmodulation was still evident at this CytD concentration.

Finally, we corroborated the effect of actin polymerization inhibition on target lysis of tumor cell lines. We compared responses of high-affinity CAR T cells treated with CytD toward two Tyr-HLA-A2–positive and one Tyr-HLA-A2–negative melanoma cell lines (501A, 624.38, and 1938, respectively, see Supplementary Fig S4G). Receptor downmodulation was evident on antigen-presenting melanoma cell lines (Fig. 5H), whereas trogocytosis was hardly evident (Fig. 5I; Supplementary Fig. S4B and S4C). Congruently, CytD interfered with CAR downmodulation but not with trogocytosis. Accordingly, we measured CytD effect on antitumor activity, as it interfered only with one process, thus allowing to distinguish between receptor downmodulation and trogocytosis effects. Treatment with 1 or 0.3 μmol/L CytD that prevented CAR downmodulation could enhance antitumor response (Fig. 5J; Supplementary Table S5). Altogether, reducing receptor downmodulation is equivalent to increasing avidity (as reported by us previously, ref. 12) and to increasing CAR T-cell response.

Mechanistic understanding of CAR T-cell regulation and activity is currently limited, despite being clinically approved and even expedited for additional clinical treatments. Adjusting CAR T cells design by the biophysical properties of binding can furthermore lead to optimal T-cell responsiveness. We previously described the binding properties that lead to differential T-cell responses. Here, impelled by a theoretical model, we investigated the underlying mechanism of CAR T cells response. We first built a novel model to describe the dependence of CAR T-cell activity on antigen density, CAR expression, and CAR affinity. Surprisingly, this model predicted that CAR downmodulation is more dominant in shaping the antigen-dependent nonmonotonous behavior. We experimentally corroborated this hypothesis, using drug-induced Shp1/Shp2 inhibition or actin polymerization inhibition, and demonstrated that reducing CAR internalization could increase cytokine production. These observations stress the fundamental regulatory role of CAR internalization, which can therefore be targeted for manipulating CAR T-cell response.

There is an increased interest in manipulating CAR affinity in order to optimize antitumor response. Recently, works aimed at understanding the dependency of functional avidity on affinity described different outcomes for increased receptor affinity, ranging from (i) an inhibitory outcome and noneffective effect (7–9, 11), (ii) a neither beneficial nor disadvantageous response (6, 29), and (iii) an increased response (10, 30). We previously measured CAR T-cell activity and documented that CARs of medium affinities outperformed the high-affinity CAR (12), supporting the first group of studies. Nevertheless, our new proposed model can predict the responses observed by groups who documented increased affinity effect or no additional beneficial effect (6, 10, 29). Hence, the effect of increased affinity depends on the avidity and antigen density, which can alter the affinity effect from beneficial to disadvantageous. This strengthens the importance of considering all of these binding properties when designing a CAR for treatment.

Manipulation of CAR T-cell regulatory pathways can serve to tune their activity. We noted that high avidity cells have a wider window of antigen densities that induce maximal response, as they are less affected by the internalization inhibitory process. CAR downmodulation has been observed and suggested to inhibit CAR T-cell responses to secondary challenges (18, 31). We tested here a drug-based manipulation of receptor expression levels, which resulted in improved responses. Likewise, genetically controlling CAR expression showed the dependence of CAR T cell exhaustion and efficacy on both the basal and dynamic CAR expression levels (18). Altogether, accumulating evidence demonstrates the eminence of receptor expression levels on CAR T-cell activity and can be considered an effective candidate for manipulating and adjusting CAR T cell responses.

Based upon validated models describing mathematically the T-cell activation through its TCR, we modeled CAR T cell activation. Similar regulatory mechanisms could explain T-cell activation, supported by the observation that similar antigen densities between CAR and TCR targets lead to maximal response. Correspondingly, a recent study (17) compared CAR and TCR activities and described their antigen-dependent nonmonotonic response. Through modeling TCR and CAR activation, reduced activity at high antigen densities was explained by a negative feedback mechanism in the form of an intracellular IFF motive. Contrarily, we used here a system that manipulated and quantified several binding properties of CAR binding, and we consequently inferred that these models were insufficient to fully describe CAR activation because they predict similar peak responses to different CAR expression levels. We resolved this by introducing an avidity-dependent regulatory mechanism. The model demonstrated that receptor downmodulation regulation is more dominant in shaping CAR T-cell activation, and could furthermore describe an antigen density range that induces peak response across all CAR affinities and avidities, in accordance with our experimental data.

The observation that antigen sensitivity and hypofunctionality at high antigen densities were affected by CAR avidity encouraged us to search for the mechanism that regulates antigen responses, a mechanism that would depend on CAR avidity. Our study described two regulating mechanisms that are affected by avidity—receptor downmodulation and proximal signaling. Proximal inhibitory signaling is known to inhibit T-cell function due to high antigen densities or high TCR affinities (16, 32, 33). Although CARs are built using intracellular signaling moieties of the TCR complex and are thought to activate similar downstream signaling, the proximal signaling of CAR T cells remains understudied. Compared with TCRs, CARs display rapid kinetics of proximal activatory signaling molecules, along with fast actin accumulation (34). Different signaling kinetics are observed even when comparing CARs constructed from different signaling domains (35). Our study further characterized CAR activation and described the involvement of several inhibitory and activatory signaling molecules. Moreover, we observed correlation between avidity, the basal proximal signaling activity, and the amount of proximal signaling following stimuli. Basal signaling is known to correlate with T-cell sensitivity and magnitude of response (35, 36). However, increasing basal phosphorylation is not sufficient to induce full T-cell response. For example, inhibition of the Lck-negative regulator Csk resulted in higher basal proximal signaling even without TCR engagement, yet recapitulation of a full TCR activation was accomplished only after CytD drug treatment, indicating the essential role of actin remodeling for T-cell stimulation (28).

We recognize that proximal signaling and receptor internalization processes are not separated, but in fact interconnected. Clustering of proximal signaling molecules such as PLCγ, LAT, and Slp76 in the immunologic synapse depends on actin polymerization (37, 38), and on the other hand, induction of proximal signaling can downregulate the number of receptors (39). Nonetheless, using the CytD and SSG drugs, we could see reduced CAR downmodulation only following CytD treatment and not following SSG treatment. This implies that we can find drugs that can affect, at least in part, one mechanism without altering the other. Induction of receptor downmodulation occurs through its ubiquitination via cCbl, and not through Shp1/Shp2 (40, 41); therefore, the SSG drug should not interfere with receptor downmodulation. Another possibility to differentiate between proximal inhibitory signaling and receptor downmodulation is to target receptor degradation. It was demonstrated in TCRs that internalization occurs constantly, but only internalized and activated TCRs are degraded, resulting in receptor downmodulation (19). When we compared CAR-T cell responses following CytD and SSG treatments, we observed an increased response in CytD-treated cells alone. Peak response of treated cells was higher, with wider window of maximal response, similar to the differences seen for increasing avidity. We suggest that actin rearrangement, and specifically receptor internalization, is the mechanism underlying the distinct responses between different avidities, as predicted from our mathematical model.

We acknowledge that both receptor downmodulation and antigen trogocytosis rely on a similar molecular basis, which depends on actin rearrangements, and therefore the distinction of the two processes can never be completely clear. Hamieh and colleagues described the association between antigen trogocytosis, receptor downmodulation, T:T interaction and activation, CAR T-cell exhaustion, reduction of the tumor's antigen expression, and the consequent tumor escape (48). Nevertheless, we observed very low levels of trogocytosis after incubation with solid tumor cells. Likewise, low CytD concentrations impinged on receptor downmodulation and not antigen trogocytosis, and could increase CAR T-cell antitumor activity. Nevertheless, both trogocytosis and receptor downmodulation processes provide valid mechanisms that might regulate CAR T-cell response. The clinical relevance of trogocytosis in CAR T-cell activity, and its distinction from receptor internalization should be further characterized.

Collectively, we propose a new model to assess the outcome of CAR T cells, as a function of receptor affinity, receptor expression levels, and antigen density. This novel knowledge could impinge upon adoptive cell therapy therapies, as we might better control CAR T-cell activity and tumor eradication, for example, by regulating receptor downmodulation.

Y. Reiter reports other from Adicet Bio outside the submitted work. G. Denkberg reports personal fees from Adicet Bio Israel during the conduct of the study and outside the submitted work. S. Shen-Orr reports personal fees and other from CytoReason outside the submitted work. No disclosures were reported by the other authors.

R. Greenman: Conceptualization, validation, investigation, visualization, methodology, writing–original draft. Y. Pizem: Validation, investigation, visualization, methodology, writing–original draft. M. Haus-Cohen: Validation, investigation, methodology. G. Horev: Software, methodology. G. Denkberg: Investigation, visualization, methodology. S. Shen-Orr: Software, methodology. J. Rubinstein: Resources, software, methodology. Y. Reiter: Conceptualization, supervision, writing–original draft.

The authors are grateful to Professor Zelig Eshhar (Weizmann Institute of Science) for the gift of Phoenix-Ampho cells and the pBullet vector. Dr. Yair Lewis and Professor Izhak Kehat kindly provided the CytD reagent. The authors also thank Drs. Stewart Abbot, Daulet (Dau) Satpayev (Adicet Bio), and Arie Admon (Technion) for helpful comments on the article.

This work was supported by a grant from the Israel Science Foundation (ISF-461/15) to Y. Reiter.

The costs of publication of this article were defrayed in part by the payment of page charges. This article must therefore be hereby marked advertisement in accordance with 18 U.S.C. Section 1734 solely to indicate this fact.