QSP Models of CAR-T Cell Therapy

1. What Is a QSP Model of CAR-T Therapy?

Quantitative systems pharmacology (QSP) uses mathematical models to connect biological mechanisms, drug or cell exposure, pharmacology, and clinical outcomes. For CAR-T cell therapy, a QSP model can represent the sequence of events linking administration of engineered T cells to their expansion, interaction with target cells, immune signaling, tumor killing, and persistence.

Unlike a conventional small-molecule PK model, where the principal measured quantity may be drug concentration, CAR-T therapy involves a living cellular population that can proliferate, migrate, become activated, differentiate, contract, and persist. The therapeutic agent is therefore dynamic rather than simply being passively distributed and eliminated.

2. What Questions Can a CAR-T QSP Model Help Answer?

CAR-T QSP models can address questions that span pharmacology, immunology, translational medicine, and clinical development.

The important distinction is that QSP models connect these processes rather than treating each measurement independently. For example, tumor burden can influence antigen availability, which can influence CAR-T activation, which can influence expansion and cytokine production, which can subsequently influence tumor killing.

3. The Biological System Behind CAR-T Therapy

A CAR-T product consists of T cells genetically engineered to express a chimeric antigen receptor (CAR). The CAR provides an engineered mechanism for recognizing a target antigen and initiating intracellular signaling that can lead to T-cell activation and cytotoxic activity.

A simplified CAR-T response can be viewed as a sequence:

Each arrow represents biological processes that may contain multiple mechanisms. A QSP model translates selected parts of this network into mathematical states, parameters, and relationships.

4. Compartments in a CAR-T QSP Model

QSP models often divide the biological system into compartments. These compartments are mathematical representations of locations or functional populations rather than perfect anatomical descriptions.

The model may allow cells to move between compartments through trafficking rates. For example, circulating CAR-T cells can enter tumor tissue at a rate that depends on trafficking, vascular access, chemokine signaling, or other modeled processes.

5. What Does the Model Actually Track?

The fundamental quantities in a QSP model are often state variables. These describe the biological state of the system at a particular time.

A simplified model might track:

• \(C(t)\): circulating CAR-T cells.
• \(T(t)\): tumor-cell population.
• \(A(t)\): available tumor antigen.
• \(E(t)\): engaged CAR-T/tumor-cell complexes.
• \(I(t)\): activated or proliferating CAR-T cells.
• \(Y(t)\): concentration of a cytokine such as IL-6.

These variables can be linked through ordinary differential equations (ODEs). For example, CAR-T cells may increase through proliferation and decrease through cell death or trafficking out of the compartment.

The exact mathematical form depends on the biological assumptions and purpose of the model.

6. Modeling CAR-T Cell Expansion

One of the defining characteristics of CAR-T therapy is that the therapeutic cell population can expand after administration. This distinguishes CAR-T pharmacology from conventional drugs whose systemic amount generally decreases after administration unless repeated dosing occurs.

A simple exponential expansion model can be written as:

where \(r\) is the net growth rate. The solution is:

This model is useful conceptually but is usually too simple to describe sustained CAR-T expansion. Biological populations encounter limitations such as target depletion, nutrient constraints, inhibitory signaling, differentiation, and cell death.

A logistic representation introduces a carrying capacity \(K\):

More mechanistic QSP models may instead make proliferation dependent on antigen engagement or other immune signals.

7. Modeling Antigen Recognition and Target Engagement

CAR-T activity begins with recognition of a target antigen. A QSP model can represent antigen recognition as an interaction between CAR-T cells and antigen-bearing cells.

A simplified binding relationship can be represented as:

where:

• \(C\) represents CAR-T cells.
• \(T\) represents target-positive tumor cells.
• \(E\) represents an engaged CAR-T/tumor-cell complex.
• \(k_{\mathrm{on}}\) is the association rate.
• \(k_{\mathrm{off}}\) is the dissociation rate.

The corresponding interaction equation can be written:

This is intentionally simplified. Real CAR signaling involves receptor density, antigen density, synapse formation, intracellular signaling, co-stimulation, activation state, and multiple downstream pathways.

8. Modeling CAR-T-Mediated Tumor Killing

A central purpose of a CAR-T QSP model is to connect CAR-T engagement with tumor-cell death.

A simple mass-action killing term can be written:

More commonly, the engaged-cell population itself can be used as the driver of killing:

where \(r_T\) represents tumor growth and \(k_{\mathrm{kill}}\) represents the effectiveness of engaged CAR-T cells in producing tumor-cell death.

A saturable relationship can be used when killing does not increase indefinitely with target abundance or CAR-T exposure:

In a mechanistic CAR-T model, the driver \(C\) might instead be replaced by an activated or engaged CAR-T population.

9. Connecting CAR-T Cells to Tumor Growth

The tumor component of a QSP model determines how the malignant-cell population changes in the absence and presence of therapy.

A simple tumor-growth model might use exponential growth:

However, many QSP models use logistic or other constrained growth models:

Here \(K_T\) represents a carrying capacity and \(R_{\mathrm{kill}}\) represents CAR-T-mediated tumor-cell removal.

The interaction between these processes can generate several qualitatively different response patterns:

• Rapid tumor elimination following CAR-T expansion.
• Partial tumor reduction followed by residual disease.
• Initial response followed by tumor regrowth.
• Limited response because CAR-T exposure or engagement is insufficient.
• Response followed by relapse associated with antigen loss or other mechanisms.

These patterns are emergent properties of the coupled system rather than outputs of a single parameter.

10. Modeling CAR-T Persistence and Contraction

After an expansion phase, CAR-T cell numbers may contract. A simple model separates proliferation from loss:

where \(r\) is the effective proliferation rate and \(d\) is the loss rate.

If \(d>r\), the population declines. If \(r>d\), the population expands. In reality, both rates may change over time as antigen availability, differentiation state, immune regulation, and cellular phenotype change.

A more flexible model can divide CAR-T cells into phenotypic states:

Transitions between these states can be modeled using differentiation and reversion rates. This allows a QSP model to investigate how cellular phenotype may influence expansion, cytotoxicity, and long-term persistence.

11. Modeling Cytokine Release

CAR-T activation can stimulate production of cytokines and other inflammatory mediators. Cytokine dynamics are therefore an important component of models investigating both pharmacology and treatment-related toxicity.

A simple cytokine model can be written as:

where \(Y\) is cytokine concentration, \(P_Y\) is a production function driven by CAR-T activation or engagement, and \(k_Y\) is the effective elimination rate.

For example, production might increase with the number of engaged CAR-T cells:

More complex models may incorporate nonlinear production, immune-cell feedback, receptor-mediated consumption, multiple cytokines, and tissue-specific production.

12. Modeling CAR-T Trafficking and Tissue Distribution

CAR-T cells must reach the relevant tissue before they can interact with target cells there. A QSP model can represent trafficking between blood and tissue compartments.

For circulating CAR-T cells \(C_B\) and tumor-resident CAR-T cells \(C_T\), a simple model is:

Here \(k_{\mathrm{in}}\) and \(k_{\mathrm{out}}\) describe movement between compartments.

Trafficking parameters can be especially important when systemic CAR-T measurements do not directly represent tumor exposure. A high circulating cell count does not necessarily imply a high intratumoral CAR-T concentration.

13. Why Feedback Loops Matter in CAR-T QSP Models

Immune systems contain many feedback loops. These loops can create nonlinear and sometimes counterintuitive behavior.

For example:

This is a negative feedback loop: successful killing reduces the target that drives further activation.

Other feedback pathways may be positive. For example, antigen-driven activation can increase CAR-T proliferation, producing more CAR-T cells capable of recognizing the target.

14. Representing Interpatient Variability

CAR-T responses can vary substantially between individuals. QSP models can represent variability in biological parameters such as proliferation, trafficking, tumor growth, antigen density, killing capacity, or cytokine production.

A simple population model can represent an individual parameter as:

where \(\theta_{\mathrm{pop}}\) is the population-typical value and \(\eta_i\) represents individual-level deviation.

Alternatively, parameters can be related to measured covariates:

Potential covariates might include tumor burden, antigen expression, baseline immune-cell characteristics, or other clinically measured features, depending on the model and available evidence.

15. Modeling Mechanisms of CAR-T Resistance

A useful QSP model can explicitly represent mechanisms that reduce treatment effectiveness.

Examples include:

• Antigen loss: tumor cells become antigen-negative and therefore less susceptible to CAR recognition.
• Antigen heterogeneity: only a fraction of tumor cells express sufficient target antigen.
• Limited trafficking: CAR-T cells do not reach the tumor efficiently.
• T-cell dysfunction: CAR-T cells lose proliferative or cytotoxic capacity.
• Exhaustion: repeated stimulation is associated with a less functional cellular state.
• Immunosuppressive microenvironment: local factors reduce CAR-T activity.
• Insufficient persistence: CAR-T cells decline before complete tumor eradication.

For example, a simple antigen-loss model can divide tumor cells into antigen-positive and antigen-negative populations:

CAR-T killing can then act preferentially on \(T_+\), while \(T_-\) cells grow or persist with substantially less CAR-dependent killing.

16. Putting the Pieces Together

A minimal CAR-T QSP model might simultaneously describe CAR-T cells, tumor cells, engaged complexes, and a cytokine.

For example:

This system captures four linked ideas:

1. CAR-T cells can expand in response to engagement.
2. Tumor cells grow but are removed through CAR-T-mediated killing.
3. CAR-T cells and tumor cells form an engaged population.
4. Engagement can generate cytokine production.

The model is intentionally simplified. A research-grade QSP model could add antigen density, trafficking, multiple T-cell phenotypes, several cytokines, immune suppression, target-cell heterogeneity, and other mechanisms.

17. Worked Example: A Simplified CAR-T/Tumor Model

Consider a hypothetical model with an initial tumor burden of 1.0 × 10⁹ cells and an initial CAR-T population of 1.0 × 10⁶ cells.

Suppose the model uses the simplified equations:

Assume:

• \(r_C=0.20\) day\(^{-1}\)
• \(d_C=0.05\) day\(^{-1}\)
• \(r_T=0.015\) day\(^{-1}\)
• \(k_{\mathrm{kill}}=1.0\times10^{-10}\) cell\(^{-1}\) day\(^{-1}\)

Step 1: Net CAR-T growth rate

In this simplified representation, the CAR-T population initially has a positive net growth rate.

Step 2: Initial tumor growth rate without CAR-T killing

Thus, without treatment effects, the hypothetical tumor population would initially increase by approximately 15 million cells per day under the assumed exponential-growth component.

Step 3: Initial CAR-T-mediated killing term

At these hypothetical starting values, the modeled killing term is much smaller than the tumor-growth term.

Step 4: Biological interpretation

The important insight is not the numerical prediction itself. The example illustrates how the model connects CAR-T abundance, tumor burden, and killing efficiency into a single dynamic system.

As CAR-T cells expand, the killing term can increase. As tumor cells are eliminated, the target population decreases, which can subsequently reduce CAR-T stimulation and expansion in models where proliferation depends on antigen engagement.

18. Linking the Model to Clinical Biomarkers

A QSP model becomes especially useful when its latent biological states can be connected to observed biomarkers.

The model therefore distinguishes between latent biological processes and observed measurements. A state variable such as tumor-infiltrating CAR-T cells may not be directly observable, but it can influence several measurements that provide indirect information about the state.

19. Calibrating a CAR-T QSP Model

Calibration is the process of determining model parameters or parameter distributions that allow the model to reproduce relevant experimental or clinical observations.

A typical workflow includes:

1. Define the biological structure. Specify which mechanisms and compartments are represented.
2. Identify parameters from prior evidence. Literature, experiments, and translational data can provide informative ranges.
3. Identify parameters requiring estimation. Some rates may need to be estimated from clinical or experimental observations.
4. Fit the model to appropriate datasets. Use data that contain information about the processes being estimated.
5. Evaluate model predictions. Compare predictions with observations and assess residual patterns and biological plausibility.
6. Validate against independent information. Where possible, evaluate predictions using data not used for calibration.

20. Sensitivity Analysis in CAR-T QSP Models

QSP models can contain many parameters. Sensitivity analysis helps identify which parameters have the greatest influence on an outcome of interest.

A local sensitivity coefficient can be expressed as:

where \(Y\) is an output and \(\theta\) is a model parameter.

A normalized sensitivity can be useful when parameters and outputs have different scales:

Potentially influential parameters might include:

• CAR-T proliferation rate.
• CAR-T death rate.
• Trafficking rates.
• Antigen-binding parameters.
• Tumor-cell killing rate.
• Antigen-loss rate.
• Cytokine production rate.
• Cytokine clearance rate.

Global sensitivity analysis can additionally investigate interactions among parameters and nonlinear effects across broad parameter ranges.

21. Parameter Uncertainty and Structural Uncertainty

QSP model predictions contain uncertainty from several sources.

For this reason, a QSP result should generally be interpreted as a model-based prediction conditional on a set of assumptions rather than as a direct observation of an unmeasured biological quantity.

22. What Can a CAR-T QSP Model Simulate?

Once calibrated and evaluated, a QSP model can be used to simulate hypothetical scenarios.

• Different initial CAR-T doses.
• Different CAR-T expansion characteristics.
• Changes in target-antigen density.
• Different tumor burdens at treatment.
• Alternative trafficking assumptions.
• Changes in tumor-cell killing efficiency.
• Different persistence profiles.
• Potential effects of antigen heterogeneity or antigen loss.
• Interactions between CAR-T activity and cytokine dynamics.
• Hypothetical combination strategies when the relevant mechanisms are represented.

Simulation is particularly valuable because many clinically important biological states cannot be measured continuously in humans. The model can provide a mechanistic hypothesis about what may be occurring between observations.

23. From Cellular Exposure to Clinical Response

For conventional drugs, the conceptual sequence is often:

For CAR-T therapy, the sequence is more dynamic:

At the same time, feedback pathways can connect response back to cellular exposure:

This is one reason QSP can be particularly useful for cell therapies: the therapeutic agent itself changes dynamically in response to the biological system.

24. Extending the Model to Combination Therapies

QSP models can be extended to investigate combinations when there is a mechanistic basis for representing the interaction.

For example, a combination could theoretically influence:

• CAR-T expansion.
• Tumor-cell antigen expression.
• T-cell trafficking.
• Immune suppression.
• Tumor-cell susceptibility to killing.
• Cytokine production or clearance.

A combination model should explicitly represent the proposed mechanism rather than simply assuming that two treatments produce an arbitrary additive effect.

25. A Practical Workflow for CAR-T QSP Modeling

1. Define the scientific question. Determine whether the model is intended to study expansion, trafficking, efficacy, safety, persistence, resistance, or another question.
2. Map the biological mechanisms. Identify the minimum network of processes needed to address the question.
3. Define compartments and state variables. Decide which cellular and molecular populations need to be represented.
4. Specify mathematical relationships. Translate biological interactions into equations and parameter definitions.
5. Gather prior information. Use experimental, translational, clinical, and literature evidence where appropriate.
6. Calibrate the model. Estimate uncertain parameters using suitable data.
7. Evaluate model adequacy. Examine fits, predictions, biological plausibility, and model diagnostics.
8. Perform sensitivity and uncertainty analyses. Identify influential assumptions and quantify prediction uncertainty.
9. Test external predictions where possible. Compare simulations against independent datasets or observations.
10. Use the model for simulation. Explore scenarios that address the original scientific question.
11. Communicate assumptions clearly. Distinguish observed evidence from model-based inference.

26. What CAR-T QSP Models Do Not Tell Us Automatically

QSP models can integrate a large amount of biological knowledge, but model complexity does not automatically make a prediction correct.

• A mechanistic model is still a model. It represents selected biological processes rather than the complete immune system.
• Parameter estimates depend on the model structure. The same data can sometimes support different parameter values under different structural assumptions.
• Good calibration does not prove mechanism. Multiple mechanisms may produce similar observed trajectories.
• Unobserved states remain uncertain. A model can estimate tumor-infiltrating CAR-T cells without directly observing them continuously.
• Extrapolation can be assumption-dependent. Long-term persistence or rare resistance mechanisms may be difficult to infer from short datasets.
• Patient heterogeneity matters. A population-level model may not describe every individual equally well.
• Model complexity has costs. Adding mechanisms can introduce parameters that are difficult to identify from available data.

Where to Go Next

A natural progression is to study QSP Models of T-Cell Activation, followed by target engagement, CAR-T expansion and persistence, cytokine dynamics, tumor growth and response models, antigen-loss mechanisms, and model qualification for mechanistic cell-therapy models.

The next tutorials can build on the framework introduced here by examining how individual biological mechanisms are represented mathematically and how those mechanisms are connected into a larger QSP model.