Predicting Tumor Killing and T-Cell Activation by T-Cell Bispecific Antibodies as a Function of Target Expression: Combining In Vitro Experiments with Systems Modeling

T-cell bispecific antibodies and antibody fragments are promising modalities in the field of cancer immunotherapy (1). They bind simultaneously to CD3 on T cells and a specific antigen on tumor cells. This will result in crosslinking of T cells and tumor cells, causing the formation of immune synapses that will promote activation of T cells and granzyme-induced apoptosis of tumor cells (2–4). TCBs can redirect non-naive T cells irrespective of their specificity, thereby inducing a polyclonal T-cell response (5).

Cibisatamab is a novel TCB construct that targets a membrane-proximal epitope of human carcinoembryonic antigen (CEA) and CD3 epsilon-chain. Further compound characteristics are summarized in the Supplementary Section S1.1.

CEA is a surface antigen that is mainly found on the apical side of columnar cells in the colon, where it is expressed at low densities (6). CEA is often overexpressed in gastrointestinal tumors, where also its expression is no longer polarized to the apical side. Due to being largely tumor specific, CEA is an attractive target for targeted therapies against gastrointestinal tract neoplasms (7).

For such bispecific molecules, drug effect requires formation of ternary complexes between tumor and T cells, leading to activation of T cells. These ternary complexes result in the formation of immunologic synapses that mimic naturally formed synapses between antigen-specific T cells and tumor cells. This implies that the effects of TCBs are complex and multifactorial, and cannot be directly related to the concentration of the compound. Mathematical models have been proposed to understand the complex relationship between target expression, target affinity, T-cell infiltration, activation, proliferation, and drug effects (8–10). Such models are helpful in exploring the potential impact of altering experimental conditions and predicting the outcome of untested scenarios, thus assisting in the design of new experiments and informing on desired compound properties.

Campagne and colleagues proposed a model to capture the exposure–response relationship for TCBs in cynomolgus monkeys with respect to T-cell trafficking, T-cell–mediated tumor lysis, and the formation of antidrug antibodies (8). Jiang and colleagues developed a mechanistic model, based on in vitro data sets, assuming that immune synapse formation is driving the drug effect and linked this to drug and system-specific parameters across different compounds and cell lines (9). The model predicted cytotoxicity at single time points and showed the impact on cell killing by both system and compound related properties, such as E:T ratios and binding affinities. Betts and colleagues employed a translational model during nonclinical development of LP-DART, a proprietary-format TCB targeting P-Cadherin, to support clinical dosing regimen projection. They linked the synapse-based model to in vivo pharmacokinetics (PK) to characterize TCB plasma concentrations and in vivo efficacy in xenograft mice (10). This mechanistic model was also used to project the in vitro minimally anticipated biological effect level (MABEL) from LP-DART to in vivo as an alternative to determine the first-in-human dose (11).

Here, we propose a modeling framework that predicts drug-induced T-cell activation and tumor cell killing in relation to drug concentration, target expression, and target antigen affinity. Therefore, we conducted tailored in vitro experiments and generated a longitudinal in vitro dataset to inform the model. We developed our model in conjunction with a learn-and-confirm paradigm, using tailored experimental data from a low and high expression cell line (Fig. 1). We built the model with the data of the high target expressing cell line and tested the predictive power of the model with the dataset of the low expression tumor cell line. Using that approach, we show that the model can predict in vitro cytotoxicity induced by cibisatamab, based on the surface expression density of CEA. The application of such a learn-and-confirm methodology entails the generation of an informative in vitro dataset over several time points that spans multiple tested concentrations. In summary, the proposed model is a tool complementary to experimentation. It allows the investigator to test hypotheses, to explore conditions that go beyond the performed experiments, and to probe the understanding behind the mechanism of action.

MKN45 (high CEA expressing tumor, CEA density range 230,000–690,000/cell) and CX1 (low CEA expressing tumor, CEA density range 2,000–11,000/cell) were used as target cell lines. To determine tumor cell lysis and T-cell activation, tumor cells were stained with 1.75 μmol/L eFluor670 (#65-0840-85; ebioscience) and PBMCs were isolated from fresh human blood with Histopaque-1077 density gradient method (#10771, Sigma Aldrich) and stained with 0.2 μmol/L CFSE (#21888, Sigma Aldrich). A total of 30,000 target cells (MKN45 or CX1) were seeded in flat-bottom 96-well plates and co-cultured with 300,000 PBMCs per well to attain an E:T of 10:1 (assay medium RPMI1640 + 2% FCS + 1% GlutaMax). Cibisatamab dilutions were added to reach the required total TCB concentrations (6, 32, 160, 800, 4,000, 20,000, and 100,000 pmol/L). For the negative control, 50 μL of assay medium was added. All assays were performed in triplicate. The co-cultures were incubated for 24, 48, 72, 96, and 168 hours at 37°C in a humidified incubator. At the indicated time points, supernatants were collected and FACS analysis was performed to quantify cell killing, T-cell activation, and cytokine release (Supplementary Section S1).

A set of ordinary differential equations (ODE) was developed to describe the formation of immune synapses by cibisatamab, triggering activated T cells (CD25⁺), and T-cell–mediated tumor cell killing. Figure 2 represents the schematic of the model. All parameters and their description are listed in Supplementary Table S1.

Immune synapse formation was modeled in a well-stirred setting assuming sequential and independent binding processes between the molar free fractions of cibisatamab (TCB), CEA antigen (CEA), and CD3e (CD3), which implies that binding to the first target does not affect the binding affinity to the second target. For simplicity reasons, we assume that cibisatamab has only one binding arm for CEA. When TCB binds to CEA or CD3, dimers between TCB and CEA (CEAdim) or TCB and CD3 (CD3dim) are formed, respectively. CEAdim and CD3dim can subsequently bind CD3 and CEA, respectively, to form a ternary complex (Synapse) between the three entities (Fig. 2A). Synapse formation is described by Eqs. A to D. It should be noted that cibisatamab binding to soluble CEA that was shed from tumor cells is negligible and was not accounted for in the model (3).

Total concentrations of CEA and CD3e are derived from the respective cell concentrations (Eq. E). Antigen expression per cell is assumed to remain constant. Surface expression of CD3 was fixed at 50,000/cell (internal data). The rates kon and koff are the respective association and dissociation constants between cibisatamab and CEA (kon_R and koff_R) or CD3e (kon_CD3 and koff_CD3). Only the KD for binding to CD3 and the FRET IC₅₀ for binding to CEA were known. Therefore, both koff values were taken as 3.6 h⁻¹ and both kon values were derived as kon = koff/(KD or IC₅₀).

It is assumed that the density of immune synapses per cell is one of the drivers behind T-cell activation (Fig. 2B), and that activated T cells express CD25 on their surface which we monitored in the in vitro experiments. Bimolecular complexes between cibisatamab and CEA (CEA_dim) or CD3e (CD3_dim) are inactive. The molar concentration of immune synapses is converted into a density of an absolute number of immune synapses/cell and averaged by the total number of tumor cells per microliter (Eq. F). The number of synapses/cell will elicit a delayed stimulation on CD8⁺ T cells, which was monitored by CD25 expression (Eqs. G–L).

The stimulatory drug effect on T cells (STIM) is dependent on the number of immune synapses per cell (Synapse_cell) and the relative tumor target expression, captured by a scaling factor δ, accounting for differences in CEA expression between the experimental data and a reference system (Eq. H). For the presented case, the reference system is MKN45. This stimulation of T cells is delayed relative to the formation of the synapses and is described by a sigmoidal function where E_max represents maximal stimulation by immune synapses, multiplied with the relative target expression. The number of immune synapses per tumor cell required to reach half-maximal killing is given by Complex50. The transit compartments delay₁, delay₂, and delay₃ dampen the stimulatory effect of STIM on CD25 induction on CD8⁺ cells (Eqs. I–K). K_tr is the transit rate from one compartment to the other and equals 3/tau.

The gradual increase in CD25⁺CD8⁺ T cells is captured with a delay compartment linked to a signal transduction cascade with a feedback loop and captures the transient T-cell activation (Fig. 2C). The synthesis rate of CD25⁺ T cells is given by kin, and kreg is the rate of CD25 turnover (Eq. L). Depending on the amount of Synapse_cell formed, delay₃ will add up to kin. The output rate of CD25-positivity is given by k_reg and can be calculated assuming a steady state (Eq. M). A negative feedback loop is described with three transit compartments (Transit₁, Transit₂, Transit₃). The transit compartments will flow into the Reg compartment, which has an inhibitory effect on kin (Eqs. N–Q) and is depicted by the light-grey rounded arrow at the far right of Fig. 2 (12).

A control group without cibisatamab treatment was used to study unperturbed tumor growth for both CX1 and MKN45. The control group was used to derive the tumor-related parameters growth rate (k_g) and carrying capacity (K), which are specific for each cell line. Unperturbed growth is described by a logistic function with parameters for tumor growth rate (k_g) and maximal tumor capacity (K). From the tumor growth rate, the tumor doubling time (t_d) can be derived (Eq. R)

The killing of tumor cells is driven by the increase in CD25⁺CD8⁺ T-cell concentration over baseline and is described by a sigmoidal function (Eq. S, Fig. 2D). The CD25⁺CD8⁺ T cells are baseline corrected to eliminate cytotoxic activity in the control group. The maximal rate of tumor killing is proportional to the tumor growth rate. Parameter kappa (κ) is the proportionality constant between tumor growth rate and maximal killing.

All model parameters and their description are listed in Supplementary Table S1.

To verify the robustness of the model, the estimated parameter values obtained during model building with MKN45 (kin, tau, Emax, Complex50, CD25_baseline, CD25₅₀, h, κ) were used as an input to predict the outcome of the assays on CX1 (low CEA-expressing). Cell line specific parameters (k_g, K, CEA-expression level) were changed accordingly (Table 1). Cell line-specific parameters were either measured or estimated from the untreated sample. Predictions of tumor cell and T-cell profiles were overlaid with observed data.

As an external validation, the model was used to simulate an unrelated dataset presented by Bacac and colleagues (3), where they show cytotoxicity in function of CEA expression levels. Tumor cytotoxicity at 48 hours after incubation with 20 nmol/L cibisatamab was simulated across a 7-log CEA-density range. Cytotoxicity was calculated as (1 − Treatment/Control) × 100%. Simulations were overlaid and compared with published data (3). A Monte Carlo simulation of 250 iterations was run to capture the random effects estimated by the model. Random effects for k_g and K were fixed at 0.3 to account for greater variability between cell lines.

Model parameter estimation was performed with nonlinear mixed-effects modeling in Monolix (version 2018R2; Lixoft). Monolix employs the stochastic-approximation expectation maximization (SAEM) algorithm. Model verification and simulations were performed in Berkeley-Madonna (version 8.3.18; University of California, Berkeley). Diagnostic plots comparing observed versus predicted values and visual predictive checks assessing goodness-of-fit were used to check model performance. The precision of the estimations and the identifiability of the parameters were assessed by means of diagnostic criteria such as the relative standard error (RSE) and parameter shrinkage.

MKN45 (AAC 409) and CX1 (AAC 129) were obtained from DSMZ. Authentication of MKN45 was performed by Microsynth AG in 2018 to confirm identity. No authentication was performed on CX1. No mycoplasma testing was performed prior to the study. Both cell lines' CEA expression levels were determined with QIFIKIT (Agilent Dako), as described elsewhere (3). Cell lines have been in culture for 1 to 4 months at the start of the experiments.

In high expression tumor cell line MKN45 (230,000 CEA/cell), we modeled the time course of T-cell activation captured by the number of CD25⁺CD8⁺ T cells (Fig. 3A) and of tumor cells (Fig. 3B) in the presence and absence of cibisatamab. Overlay of model prediction (solid line) with observed data (triplicate means, symbols) are shown. Individual triplicate observations are shown in Supplementary Fig. S2.

With regards to T-cell activation, we observed a dose-dependent increase in CD25⁺CD8⁺ T cells (Fig. 3A, symbols) with maximal levels of activated T cells of 164 cells/μL observed in high dose groups (4,000—100,000 pmol/L). In the control group, number of activated T cells was low (∼3 cells/μL) and only slight changes were observed throughout the observation period. T-cell activation reached maximal levels after approximately 4 days, followed by a decrease in CD25 expression observed at 7 days. We observed gradual and dose-dependent increases in immune checkpoints TIM-3 and PD-1 (Supplementary Fig. S3). The decline in activated T cells at later time points may be explained by these inhibitory mechanisms. Activated T cells remain considerably elevated compared with baseline after 1 week incubation.

Tumor cell growth was monitored over time in the presence and absence of cibisatamab (Fig. 3B) in MKN45. In the control group, tumor cell growth was characterized with exponential growth phase during the first 3 days reaching a plateau at 168 hours. From the control group, the estimated tumor growth rate was 0.058 hour⁻¹ (RSE 6.48%) resulting a doubling time of 12 hours (Fig. 3B) and the estimated carrying capacity was 326 cells/μL (RSE 9%; Table 1). In the presence of cibisatamab, we observed a dose-dependent decrease of tumor cell growth. Almost complete tumor growth reduction occurred at high cibisatamab concentrations (4,000–100,000 pmol/L). In addition to the dose-dependent effects on activated T cells and tumor growth inhibition, we observed a dose-dependent effect on cytokine release (Supplementary Fig. S3). In this study, we observed an early IL2 release that triggers T-cell activation and CD25 induction.

In summary, the model well captured the time course of T-cell activation (Fig. 3A, solid lines) and tumor cell growth (Fig. 3B, solid lines) in control and treated group across a broad dose range of cibisatamab. The model parameters were estimated with good precision, with the highest RSE below 45% (Supplementary Table S2). The diagnostic plots showed overall good model performance (Supplementary Figs. S4 and S5). At lower TCB concentrations, some observations fall outside the 90% prediction intervals in Supplementary Fig. S4A. It should be noted that this misspecification is only observed for very low T-cell numbers (∼3–18 cells/μL). At these levels, the accuracy of the assay may not be sufficient to reliably measure the exact T-cell numbers.

For model verification, we tested whether the model could predict the activity of cibisatamab acting on a tumor cell line (CX1) with much lower surface expression of CEA (11,000/cell) based on the model built on the high expression cell line MKN45 (230,000/cell). Therefore, we simulated the expected time course of CD25⁺CD8⁺ T cells (Fig. 4A, solid line) and tumor cells (Fig. 4B, solid line) and overlaid it with the observed data (Fig. 4A and B, symbols, triplicate means). For the simulations, we used the estimated parameters we obtained from fitting the model to the dataset with MKN45 and replaced only cell line-specific parameters by accounting for the differences between the cell lines. These were target expression and the tumor growth parameters which was estimated from the control group (Table 1). For CX1, the estimated tumor growth rate was 0.114 hour⁻¹ (RSE 5%), which gives a doubling time of 6.1 hours and the estimated carrying capacity was 278 cells/μL (RSE 4.84%).

Low CEA-expressing tumor cell line CX1 showed only small increases in CD25⁺CD8⁺ T cells, reaching maximal levels in the range of 7 to 10 CD25⁺CD8⁺ cells/μL at concentration of 4,000 to 100,000 pmol/L of cibisatamab. This was approximately 20-fold lower as compared with MKN45. The baseline levels (control group) as well as the levels from the low dose groups (6–160 pmol/L) fluctuated between 1.5 and 7 cells/μL. Tumor cell growth over time (Fig. 4B) was half maximally reduced at 20,000 and 100,000 pmol/L cibisatamab. Furthermore, for the tested concentration range, no increase in cytokine was detected (Supplementary Figs. S3N–S3R).

Due to the large difference in target expression, there was a considerable change in predicted synapses per cell formed. The lower CEA expression in CX1 decreased scaling factor δ, thereby restricting synapse-mediated T-cell activation.

In summary, the model correctly predicted a dose-dependent effect on CD25⁺CD8⁺ cells with a very low increase of approximately 9 cells/μL. In addition, the predicting model captured the effects on tumor cell growth, which was described as a concentration dependent partial tumor inhibition with a clear concentration-dependency (Fig. 4B).

We used the presented model to project the cibisatamab-induced cytotoxicity to a data set of 110 different tumor cell lines, with CEA surface densities ranging from 1 to 200,000 per cell (3). These cells were treated with 20 nmol/L cibisatamab and cytotoxicity was recorded 48 hours after treatment.

A simulation of cytotoxicity shows that the model can reliably predict the observed cytotoxicity over the whole CEA expression range. The model could distinguish between responding cells (>10% cytotoxicity, green squares) and nonresponding cells (<10% cytotoxicity, orange circles) under these experimental conditions. A clear switch between no/low killing and high killing is present, with a cut-off value around 10,000 CEA/cell (Fig. 5). The shaded area is delineated by the 5% and 95% percentiles of the individual predictions. A part of the cell lines falls outside the prediction interval, which may be due to cell line specific factors, assay differences or a process not yet captured by our model.

After demonstrating that the model was able to predict the drug effect across a broad range of tumor cells, we wanted to investigate the relationship between formation of synapses as a function of drug concentration (Fig. 6A), the relative T-cell activation as a function of synapses per cell (Fig. 6B) and the relative cytotoxicity as a function of activated T cells (Fig. 6C).

Figure 6A shows the predicted synapse per cell over a range of cibisatamab concentrations in MKN45 (CEA^high; Fig. 6A, blue curve) or CX1 (CEA^low; Fig. 6A, orange curve) cell line. For both cell lines, the profiles are parallel to each other with CX1 being 20-fold lower. In both cases, a bell-shaped profile is observed and the amount of immune synapse formed will increase with increasing TCB concentrations up to a maximal threshold, after which increasing TCB concentrations negatively affect synapse formation. In both cases, peak synapse formation occurs at a cibisatamab concentration of approximately 68 nmol/L.

The model suggests that stimulation of T cells depends on the number of immune synapses formed per tumor cell and on the target expression level, relative to a reference system. Figure 6B shows percent of T-cell activation as a function average synapses. Observed (symbols) and predicted (lines) data are shown for MKN45 (in blue) and CX1 (in orange). T-cell activation was expressed relative to the highest observed value of CD25⁺CD8⁺ cells, which was set to 100% and the respective maximal observed values from each treatment group was normalized accordingly. We observed the highest T-cell activation in the MKN45 cell line at 96 hours with 20 nmol/L cibisatamab and we took this value as 100% T-cell activation. We observed the highest T-cell activation in the CX1 cell line at 168 hours with 20 nmol/L cibisatamab. This level of activation corresponded to 4.3% of the maximal T-cell activation observed in MKN45. These fractions of T-cell activation were well in line with the simulated profiles (Fig. 6B). Although both cell lines differ strongly in maximal response with regards to T-cell activation, the model correctly predicted the relationship of T-cell activation and synapses per cell. Interestingly, for the same range of synapse per cell, the predicted and observed response of T-cell activation was much lower with CX1 as compared with MKN45. In line with proposed model, we concluded from these findings that the drug effect of cibisatamab on T-cell activation not only depends on number of synapses per cell but also on the relative CEA expression level.

The proposed model suggests that tumor cell killing depends on the increase in activated T cells over baseline levels, which is described as the increase in CD25⁺CD8⁺ T cells. We therefore explored the relationship between the cytotoxicity and maximal CD25⁺CD8⁺ T-cell counts. The highest observed concentration of CD25⁺CD8⁺ T cells was observed for 20 nmol/L cibisatamab in MKN45 and was set to 100% and normalized the maximal CD25⁺CD8⁺ for each treatment group and cell lines accordingly. Figure 6C shows good agreement between the predicted and observed cytotoxicity in function of maximal T-cell activation for each dose level.

In the presented work, we provided a mathematical model that extends the current knowledge on in vitro TCB activity by including a biomarker of T-cell activation. By linking ternary complex formation to T-cell activation, the proposed model predicts correctly a dose-dependent and transient increase in CD25⁺CD8⁺ T cells, which results into tumor cell killing observed in in vitro experiments with a high (MKN45) and a low (CX1) target expressing cell line tested over a broad range of cibisatamab concentrations. Furthermore, the model was capable of capturing the tumor lysis profile across a very broad range of CEA expression densities.

Bispecific antibodies redirecting T cells to kill tumor cells are a promising therapeutic modality in cancer immunotherapy. The mechanisms behind TCB mediated T-cell redirection and tumor cell killing have been investigated and it has been described that the initiation or extent of TCB activity is multifactorial as the bispecific antibody needs to bind both its binding partners before exerting its effect. Various mathematical models have been proposed to investigate these processes in vivo or in vitro and have been applied to address relevant questions related to the development of these modalities. We have expanded on existing models describing TCB activity by linking tumor cell killing to T-cell activation.

A common denominator between these models is the assumption that the formation of immunologic synapses is driving directly or indirectly the cytotoxic effects. T cells have been used as a link between synapse formation and tumor cell killing. These were either total T-cell counts (13) or a virtual pool of activated T cells without actual data (8, 11). In our study with cibisatamab, we used in vitro experiments to build a mechanistic model that integrates formation of synapses resulting into T-cell activation and tumor cell killing. The model captures the dynamics of tumor cell and CD25⁺ cytotoxic T cells over time. It was successfully applied to predict cytotoxicity as a function of target expression and to predict the time course profiles of activated T cells and tumor cell killing of a low expression cell line based on the modeling results using high expression cell line.

CD25 is the alpha-chain of the trimeric IL2 receptor and a known marker for T-cell activation [18]. CD25 expression is inducible on effector T cells and functions within auto- and paracrine feedback loops with IL2. Binding of IL2 to CD25 will result in receptor-mediated internalization of IL2, while also inducing higher expression of CD25 on the cell surface [19]. Therefore, we mechanistically linked T-cell activation—indicated by CD25 expression induction on the cell surface—to tumor cell killing in our model. We were capable to well describe killing of the high expressing tumor cell line as well as the associated T-cell activation.

The extent of synapse formation is dictated by the expression level of CD3 and tumor antigen, tumor cell concentration, the affinity of the TCB to both antigens, and the TCB concentration in the system. A change in antigen expression impacts the amplitude at which immune synapse can be formed as there will be more or less target available for the TCB to bind to. The TCB concentration has a more ambiguous impact on synapse formation. Higher TCB concentrations will increase synapse concentrations until the optimum is reached. TCB concentrations above that optimum promote the dimerization between TCB and a single antigen, rather than formation of ternary complexes (the actual immune synapses), as was discussed elsewhere (9, 14). The relationship between the binding affinities to both targets will determine at what TCB concentration the tipping point in synapse formation will occur. In the binding model that we used, the TCB concentration at which the bell-shaped curve peaks corresponds to the geometric mean of the bindings affinities from both binding arms (15). Therefore, information on binding affinities can inform on the TCB concentration range resulting in maximal synapses per cell, which may be relevant for experimentation. Our model suggests that ternary complex formation as well as T-cell activation and tumor lysis is decreased at high TCB concentration. However, it has not been shown experimentally and further in vitro tests could be conducted to confirm this hypothesis.

In this work, we used an in vitro system as this allows is to have a well-controlled environment to investigate the relationship between all elements of this complex pharmacological response. Our work is in line with previous findings showing that in vitro efficacy is closely related to target expression (3). Moreover, a clear threshold was observed in vitro at which the system switches from minimal to complete tumor killing, and which was well captured by the model. However, it should be noted that this quantitative insight may not reflect the clinically relevant threshold, which may be dependent on other factors and is beyond the scope of the work. CEA expression is the main differentiator in our model. Besides influencing how many immune synapses will be formed, CEA expression also directly affects the extent of T-cell activation in our model. Natural immune synapse formation is a very complex mechanism that involves many TCR–pMHC complexes clustering together in a small patch on the cell surface to induce T-cell activation (16). Size and stability of the cluster depends on the number of TCR–pMHC complexes that can be formed in that area (17, 18). Therefore, we hypothesized that the density of the TCB-targeted tumor antigen will affect the probability that sufficient immune synapses are formed in close proximity to elicit T-cell activation. As a result, cells with higher antigen densities have a higher chance to activate T cells when engaged. Consequently, a certain immune synapse density may exhibit more or less T-cell activation based on the target expression level. This peculiar dependency between target expression, immune synapse density, and T-cell activation predicted by the model is backed by the experimental data as depicted in Fig. 6. Conversely, tumor cell lysis is unambiguously linked to the extent of T-cell activation.

We predicted cytotoxicity in function of a broad range of CEA densities (Fig. 5). We included a Monte Carlo simulation to also account for the random effects captured by the model. As the simulation was run with system-related parameters taken from MKN45 such as k_g and K values, we allowed a higher level of individual variability on these parameters by fixing the random effects at 0.3.

Our simulation was capable of distinguishing responder from nonresponder cell lines based on CEA expressing level. Moreover, the simulation nicely captured the general profile of increasing cytotoxicity with expression level. Our model underpredicted a part of the responding cell lines, which accounts for 27.4% of the cell lines. Potential causes for this discrepancy are discussed below.

An immune response is a multifactorial process, whereas T-cell activation by formation of ternary complexes is the first step in TCB activity, and there are many processes that will impact T-cell activation and the interaction with tumor cells. The presence of costimulatory or inhibitory receptors may change the responsiveness of the system to TCB treatment. Interestingly, we observed an increase in PD1-positivity in the CD8⁺ T cells from 10% to 70% over the course of the assay (Supplementary Fig. S3L), with a clear dose dependency. Differences in baseline levels of PD1 on T cells and PD-L1 on tumor cells could be one of the factors attributing to the variability in cytotoxicity between cell lines observed in Fig. 5. Higher levels of PD1 also suggest potential susceptibility to anti-PD1 therapy. The presence of regulatory T-cells (Tregs) may also dampen T-cell response. Baseline levels of Tregs could be one of the factors contributing to donor-to-donor variability. This warrants further investigation. The impact of anti-PD1 or other combination therapies to boost the immune response could be investigated with experiments, requiring further expansion of the model to capture these therapeutic conditions.

Bacac and colleagues showed that in tumor-bearing mice, tumors became PD-L1 positive upon cibisatamab treatment (3). Cibisatamab is also being studied in patients both as a monotherapy and in combination with anti-PD-L1 antibodies. Radiologic tumor shrinkage was observed in 50% of patients treated in combination with anti-PD-L1 therapy, as opposed to 11% of patients that only received monotherapy (7).

TCBs target antigens that are overexpressed in tumors. However, the prevalence of the antigen is often not tumor exclusive and it can be found at lower levels in healthy cells. This could result in the occurrence of on-target/off-tumor toxicities. In early drug discovery, the model presented here could help with selecting the best compound properties with respect to the expected antigen expression level in tumor and healthy cells.

Besides target expression described here, there are many more variables that influence TCB activity. Bluemel and colleagues showed that antigen size and epitope location affected TCB potency (19). More specifically, MCSP and EpCAM binding BiTE constructs showcased higher potency when targeting membrane-proximal epitopes and truncated variants of antigens.

TCB valency appears to play an important role. Cibisatamab is bivalent for CEA and monovalent for CD3e to increase tumor selectivity. Multivalency is commonly used to increase the functional affinity of a molecule and often also the potency. Besides that, target affinity has been shown to affect the biodistribution in mice (20).

Many different formats are currently under development (for overview see ref. 21). The effect of compound size and format on TCB activity is not well understood. However, they do play a big role in the PK and disposition of the TCB in the body. Depending on the format, TCBs may have similar PK profiles as classical mAbs, or could differ massively (22). Ellerman performed an extensive review of the most well studied variables (23). The limitation of the current model is that it has been build and verified with one TCB molecule and further TCBs should be tested to evaluate if the model is generalizable and if it can be applied to predict the outcome of T-cell activation and killing for scenarios when the binding affinity is modified. In addition, the model assumes a simplistic binding model, which does not fully capture the bivalency of cibisatamab to the tumor target. Furthermore, the current binding model in place assumes independent binding of CD3 and tumor target to form trimolecular complexes, which is a simplified process assuming that the system is well stirred and no spatial organization between the different binding partners is taken into account.

The presented model can be used to investigate the impact of target expression on TCB activity, or which target affinity would be optimal for a given expression level. The model can be further expanded to be more broadly applicable. Betts and colleagues showed the translational value of implementing a systems model into a larger PKPD framework, by linking both PK and PD models together (10). As a result, they were able to calculate the concentration to reach tumor stasis in mice and predict TCB disposition in patients. Because our model takes TCB concentration as an input, it could be incorporated into a holistic PKPD framework, linking intratumoral TCB concentrations to T-cell activation and tumor cell killing. A model to capture tumor uptake of TCBs is being developed (Eigenmann, M. J.; unpublished data).

In general, such methodology can be used to derive relevant therapeutic indices for TCBs. To achieve this, the model needs to be supported by longitudinal data of tumor killing, T-cell activation and cytokine release.

Due to the inherently human-specific nature of TCBs, animal models are often not appropriate or limited to recapitulate the pharmacologic processes. The use of in vitro-based assays is therefore of increasing importance (24–26). Moreover, appropriate experimental design is seminal to maximize the gain from these in vitro studies. This work presents a synergy between modeling and experimentation, which helps to dissect complexity and decrease the amount of in vitro data required to quantify drug action.

Essentially, model development with in vitro data from a single tumor cell line and model verification with data from another tumor cell line with lower antigen density gave confidence in the ability to predict cytotoxicity over a broad range of antigen densities. The paradigm of learn-and-confirm proved to be sufficient to generate a model that is predictive for the in vitro activity of cibisatamab. A similar set-up can be performed for other T-cell bispecifics, allowing for simulation of efficacy under different target expression levels and characterization of a cytotoxic threshold.

T. Lehr reports grants from Aspen Pharmaceuticals, grants from Boehringer Ingelheim, grants from Neovii, grants from Bayer, and grants from Apogenix outside the submitted work. No disclosures were reported by the other authors.

A.J. Van De Vyver: Conceptualization, software, formal analysis, visualization, methodology, writing–original draft. T. Weinzierl: Conceptualization, supervision, writing–review and editing. M.J. Eigenmann: Formal analysis, visualization, methodology, writing–review and editing. N. Frances: Software, methodology, writing–review and editing. S. Herter: Conceptualization, data curation, formal analysis, investigation, writing–review and editing. R.B. Buser: Data curation, formal analysis, investigation. J. Somandin: Data curation, formal analysis, investigation. S. Diggelmann: Data curation, formal analysis, investigation. F. Limani: Data curation, formal analysis, investigation. T. Lehr: Supervision, methodology, writing–review and editing. M. Bacac: Conceptualization, resources, supervision, writing–review and editing. A.-C. Walz: Conceptualization, supervision, visualization, methodology, writing–review and editing.

This work was sponsored by F. Hoffmann-La Roche Ltd. The authors would like to thank colleagues from Roche Innovation Center Zürich; Anne Freimoser for kindly helping with the interpretation of the affinity data of cibisatamab; and Tanja Fauti for the fruitful discussions on the in vitro data sets and providing valuable feed-back.

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.