QSP Models of Immune Checkpoint Inhibitors

1. Why Model Immune Checkpoint Inhibitors?

Immune checkpoint inhibitors (ICIs) are therapies that modify inhibitory signals controlling antitumor immune responses. Examples include antibodies directed against PD-1, PD-L1, or CTLA-4. Their pharmacology is fundamentally different from that of a drug whose primary effect is direct inhibition of a single tumor-cell signaling pathway.

The clinical effect of an ICI can emerge through a sequence of biological events involving drug exposure, target binding, checkpoint blockade, immune-cell activation, trafficking, tumor-cell killing, antigen release, and changes in the tumor microenvironment.

A quantitative systems pharmacology (QSP) model provides a mechanistic framework for representing these interconnected processes mathematically. Instead of describing only the relationship between dose and a clinical endpoint, a QSP model attempts to represent intermediate biological states that connect treatment to outcome.

2. What Is an Immune Checkpoint?

The immune system contains regulatory mechanisms that prevent excessive or inappropriate immune activation. These mechanisms include receptor-ligand interactions that can reduce T-cell activation or effector function.

In cancer, tumors can exploit immune-regulatory pathways to reduce effective antitumor responses. An ICI can interfere with one of these inhibitory interactions, changing the probability or strength of downstream immune activity.

Two commonly modeled checkpoint pathways are the PD-1/PD-L1 axis and the CTLA-4 pathway. Although both are immune checkpoints, they should not be treated as biologically interchangeable. Their expression patterns, cellular context, timing, and mechanistic roles can differ.

A QSP model does not need to represent every molecular interaction. The appropriate level of detail depends on the question. A model designed to study target occupancy may require a different representation from a model designed to simulate tumor growth under combination therapy.

3. The Architecture of an ICI QSP Model

A mechanistic ICI model commonly contains several interconnected modules. These modules may represent drug disposition, target engagement, immune-cell populations, cytokine or signaling networks, tumor cells, and treatment-induced changes in the tumor microenvironment.

The modules do not necessarily correspond to separate software components. They are conceptual parts of the mathematical system. Depending on the model, a single differential equation may connect multiple modules, or an entire network of equations may be used to represent one biological process.

4. The PK Module: Getting the ICI to Its Target

The first layer of many ICI QSP models describes drug exposure. For an antibody therapeutic, the model may include systemic concentrations, distribution, clearance, and dosing. More detailed models can include tissue distribution or target-mediated processes when those features are scientifically important.

A simple one-compartment representation can be written as:

where \(C\) is systemic drug concentration, \(V\) is the apparent distribution volume, \(CL\) is clearance, and \(R_{\mathrm{in}}(t)\) represents drug input.

For repeated IV dosing, the input function can represent a sequence of administrations. The resulting concentration-time profile then supplies the exposure input for the target-engagement component.

5. Modeling Checkpoint Target Engagement

Target engagement is the point where drug exposure is translated into a mechanistic interaction with the checkpoint system. For an antibody binding to a receptor or ligand, a simple binding model can be represented by:

Here \(D\) represents drug, \(T\) represents an unbound target, and \(DT\) represents the drug-target complex. The association and dissociation parameters are \(k_{\mathrm{on}}\) and \(k_{\mathrm{off}}\).

Under an equilibrium approximation, the dissociation constant can be written as:

A simplified occupancy relationship is:

This equation is intentionally simplified. In a full QSP model, free drug concentration, target abundance, receptor turnover, binding kinetics, tissue distribution, and competing ligands may all affect target engagement.

6. From Checkpoint Blockade to T-Cell Activity

Checkpoint blockade is not itself the final biological outcome. A QSP model must specify how reducing an inhibitory signal changes downstream immune behavior.

One simplified representation might define an effective activation signal as:

where the checkpoint term represents inhibitory signaling. An ICI can reduce this term through target engagement.

The resulting signal could then influence T-cell proliferation, activation, trafficking, cytotoxic activity, or survival. For example, a generic turnover equation for an activated T-cell population \(T_a\) might be:

This is a conceptual equation rather than a universal biological law. Different QSP models may use substantially different representations of T-cell activation.

Why intermediate mechanisms matter

Suppose two treatment regimens generate similar systemic concentrations but different immune-cell exposure or target engagement. A model containing only concentration and tumor size may not be able to explain the difference mechanistically. A QSP model can introduce the intermediate states needed to investigate such hypotheses.

7. Modeling Tumor Growth and Immune-Mediated Killing

The tumor component describes how tumor burden changes over time. A simple tumor model might begin with intrinsic tumor growth and add an immune-mediated killing term.

For example:

where \(N\) is tumor-cell burden, \(k_g\) is an intrinsic growth parameter, \(T_a\) is activated immune-cell abundance or activity, and \(k_{\mathrm{kill}}\) controls immune-mediated killing.

This model is deliberately simplified. More realistic tumor modules can represent logistic growth, Gompertz growth, multiple tumor-cell states, resistant populations, necrotic compartments, spatial effects, or treatment-induced changes in tumor-cell susceptibility.

The important point is not the specific equation but the structure of the causal hypothesis: immune activity influences tumor dynamics, while tumor burden can in turn influence antigen availability and the immune environment.

8. Why QSP Models Often Need Feedback Loops

The immune system is a dynamic network rather than a one-directional pathway. Tumor cells can influence immune activation, immune cells can alter tumor-cell populations, and changes in tumor burden can modify the biological environment in which treatment acts.

These feedbacks are one reason QSP models can behave differently from simple exposure-response models. A small change in one mechanism can propagate through the network and alter the long-term trajectory of several state variables.

9. Representing the Tumor Microenvironment

The tumor microenvironment can contain tumor cells, T cells, antigen-presenting cells, macrophages, stromal cells, cytokines, chemokines, and other biological components. ICI response can depend on interactions among these components.

A QSP model may therefore include variables representing immune-cell populations and mediators rather than treating the tumor as a single homogeneous compartment.

The modeler must decide which components are necessary for the question being studied. Adding biological detail without sufficient data can increase model complexity without providing additional identifiable information.

10. Connecting Immune Activity to Tumor Response

Tumor response can be represented in several ways. A model might predict total tumor-cell burden, viable tumor volume, a biomarker associated with tumor burden, or another clinically relevant response measure.

For a simple model, tumor burden might follow:

where \(G(N)\) represents tumor growth and \(K(T_a,N)\) represents immune-mediated killing. The two functions can be made more mechanistic as needed.

For example, immune killing might saturate with increasing immune activity:

This formulation implies that increasing \(T_a\) eventually produces diminishing increases in the per-cell killing rate. Other model structures can represent different biological hypotheses.

11. Why QSP Is Useful for ICI Combinations

One important application of oncology QSP is the study of combination therapy. A combination may contain two immune-modulating agents, an ICI plus a targeted therapy, an ICI plus chemotherapy, radiation plus immunotherapy, or other treatment modalities.

The key question is not simply whether two drugs are given together. It is how their mechanisms interact inside the biological system.

A mechanistic model can therefore provide a framework for testing hypothetical combinations through simulation before all combinations can be evaluated experimentally.

Importantly, a simulated combination response is a model-based prediction. It is not equivalent to evidence from a clinical trial.

12. Incorporating Biomarkers Into ICI QSP Models

ICI response can vary substantially across patients. QSP models can incorporate biomarkers when there is a mechanistic reason to connect the biomarker to a model component.

A biomarker should not be inserted into a QSP model merely because it correlates with response. A mechanistic interpretation requires an explicit hypothesis about which model component the biomarker informs and how that component affects downstream dynamics.

13. Representing Between-Patient Variability

A major challenge in oncology is that patients can have different tumor and immune-system states before treatment begins. A QSP model can represent this heterogeneity through different initial conditions, parameter values, or both.

For example, patient \(i\) might have a different baseline tumor burden:

or a different immune-cell abundance:

Parameters can also vary between individuals. A population model might represent a parameter such as an immune-mediated killing rate using a log-normal distribution:

where \(\theta_{\mathrm{kill}}\) is the typical population value and \(\eta_i\) represents between-patient variability.

This distinction is useful because two patients receiving the same dose can experience different biological trajectories even when their systemic drug exposure is similar.

14. A Simple ICI QSP Model in Equations

To see how the pieces fit together, consider a deliberately simplified system containing drug exposure, checkpoint target engagement, activated T cells, and tumor burden.

Drug concentration

Checkpoint binding

Here \(B\) is the concentration of drug-target complex and \(T\) is the total target concentration in the modeled compartment.

Activated T cells

The inhibitory term \(I_{\mathrm{inh}}\) can itself depend on checkpoint engagement. A simple relationship might be:

where \(\phi\) describes the degree to which checkpoint engagement changes inhibitory signaling.

Tumor burden

The complete model is therefore a coupled system:

with feedbacks and additional biological processes potentially added as the scientific question requires.

15. Worked Example: From ICI Exposure to Tumor Dynamics

Consider a hypothetical ICI model with the following simplified assumptions:

• Initial tumor burden is \(N(0)=1.0\times10^9\) cells.
• Systemic drug concentration at a particular time is \(C=10\) units.
• The effective binding constant is \(K_D=5\) units.
• Activated immune-cell activity is represented by \(T_a=0.50\) model units.
• The intrinsic tumor growth rate is \(k_g=0.020\) day\(^{-1}\).
• The maximum immune-mediated killing coefficient is \(k_{\mathrm{kill}}=0.060\) day\(^{-1}\).
• The half-saturation constant for immune activity is \(K_T=0.25\) model units.

Step 1: Approximate target occupancy

The simplified model therefore predicts approximately 66.7% occupancy at this concentration. This is a model quantity, not a direct prediction of clinical efficacy.

Step 2: Calculate the immune-mediated killing rate

Step 3: Determine the instantaneous tumor-growth rate

The tumor equation is:

Therefore:

Under these simplified assumptions, the instantaneous model trajectory is decreasing at approximately 20 million cells per day.

Step 4: Interpret the result

The example illustrates how a QSP model can connect multiple mechanistic layers: concentration influences target occupancy; target engagement can influence immune activity; immune activity affects tumor-cell killing; and the balance between tumor growth and killing determines the direction of the tumor trajectory.

16. Why Concentration Alone May Not Explain ICI Response

For many conventional drugs, it can be useful to relate exposure directly to pharmacologic effect. For an ICI, however, the path from concentration to tumor response may contain multiple biological steps.

This does not mean that a direct exposure-response model is inappropriate. Rather, a QSP model provides additional mechanistic variables that can be examined when concentration alone is insufficient to answer the scientific question.

For example, two regimens could produce similar average concentrations but different patterns of target occupancy or different downstream immune dynamics. A mechanistic model can be used to investigate whether those differences could plausibly produce different long-term outcomes.

17. Modeling Primary and Acquired Resistance

Not every tumor responds to checkpoint blockade, and tumors that initially respond can later progress. QSP models can represent potential mechanisms of resistance when sufficient biological evidence exists to support them.

One conceptual approach is to divide the tumor into sensitive and resistant populations:

The sensitive population might be affected by immune-mediated killing:

while the resistant population could follow:

The switching terms represent a hypothetical transition between states. Other models may represent resistance through antigen loss, altered antigen presentation, immune exclusion, checkpoint-independent suppression, or changes in tumor-cell susceptibility.

18. What Can an ICI QSP Model Be Used to Simulate?

Once the model has been calibrated and evaluated, simulations can explore hypothetical conditions that may be difficult or expensive to study experimentally.

• Alternative dosing schedules.
• Different exposure levels.
• Changes in target expression or target turnover.
• Different baseline immune-cell states.
• Potential combination mechanisms.
• Changes in tumor growth or immune-killing parameters.
• Potential sources of patient heterogeneity.
• Mechanistic explanations for observed biomarker-response relationships.
• Hypothetical treatment sequences or treatment interruptions.

Simulation is particularly useful for separating mechanistic hypotheses. For example, one simulation may hold tumor sensitivity constant while varying immune-cell abundance, whereas another may hold immune abundance constant while varying tumor susceptibility.

The resulting differences can help identify which mechanisms are capable of generating the observed behavior, although simulations alone cannot establish which mechanism is actually operating in patients.

19. Parameter Estimation and Calibration

An ICI QSP model may contain many parameters. Some can be informed directly by experimental measurements, while others may require estimation from integrated datasets.

A central QSP challenge is that not every parameter can be uniquely estimated from every dataset. Some parameters may need to be fixed from external evidence, constrained by prior knowledge, estimated jointly, or represented through distributions rather than single values.

This is why parameter identifiability, sensitivity analysis, and uncertainty assessment are essential components of mechanistic model development.

20. Sensitivity Analysis in ICI QSP

Sensitivity analysis asks how strongly model predictions change when model parameters or inputs are changed.

A local sensitivity measure can be represented conceptually as:

where \(y\) is a model output and \(p\) is a parameter. A normalized sensitivity can also be used:

For a QSP model, outputs might include tumor burden at a particular time, target occupancy, immune-cell abundance, or probability of a specified response criterion.

Global sensitivity methods can explore parameter variation over a broader range and can help identify interactions among parameters.

21. How Should an ICI QSP Model Be Evaluated?

A QSP model should be evaluated against its intended purpose. Because a mechanistic model can be complex, matching one dataset is not sufficient evidence that the complete biological system has been adequately represented.

1. Check structural assumptions. Are the modeled mechanisms appropriate for the scientific question?
2. Verify equations and implementation. Does the software implement the intended model correctly?
3. Check parameter plausibility. Are parameter values consistent with available biological evidence?
4. Evaluate calibration. Can the model reproduce relevant observations used during development?
5. Perform sensitivity analysis. Which assumptions and parameters drive the predictions?
6. Evaluate external predictions. Where possible, compare predictions with data not used to fit the model.
7. Assess uncertainty. How much do predictions change under plausible parameter uncertainty?
8. Document limitations. Clearly identify mechanisms that are simplified, uncertain, or unsupported by direct data.

Model qualification is therefore broader than goodness-of-fit. The objective is to establish whether the model is sufficiently credible for the intended use.

22. Virtual Patients and Population Simulation

A mechanistic QSP model can be used to generate a virtual population by varying parameters or initial conditions according to distributions intended to represent patient heterogeneity.

For example, baseline tumor burden, immune-cell abundance, target expression, or tumor-cell sensitivity could vary across simulated individuals.

where \(\boldsymbol{\theta}_i\) is the parameter vector for virtual patient \(i\), and \(P(\boldsymbol{\theta})\) represents the assumed population distribution.

Each virtual patient can then be simulated under one or more treatment scenarios. The resulting population of trajectories can be summarized to explore how biological heterogeneity may influence treatment response.

23. Using QSP to Explore Biomarker-Response Relationships

One of the strengths of QSP is the ability to connect a measurable biomarker with an unobserved or difficult-to-measure biological mechanism.

For example, suppose a biomarker \(X\) is associated with baseline immune infiltration. A model might represent:

The model could then propagate this baseline difference through immune activation and tumor killing to predict different tumor trajectories.

This provides a mechanistic hypothesis:

The value of the model is that it makes the proposed causal pathway explicit and allows the implications of the hypothesis to be examined quantitatively.

24. Worked Conceptual Example: ICI Plus a Second Therapy

Suppose a second therapy increases the effective antigenic stimulus experienced by the immune system. A simplified QSP representation could introduce an additional factor \(A(t)\) into the immune-activation equation:

The ICI primarily affects \(I_{\mathrm{inh}}\), while the second therapy changes \(A(t)\). The two treatments therefore act on different parts of the mechanistic system.

A simulation could compare:

• ICI alone: \(A(t)\) remains at its baseline value.
• Second therapy alone: \(I_{\mathrm{inh}}\) remains unchanged.
• Combination: both mechanisms are active.

The model can then determine whether the combined mechanism produces a qualitatively different tumor trajectory under the assumed parameter values.

25. What ICI QSP Models Do Not Tell Us Automatically

QSP models can be powerful, but their predictions remain conditional on their assumptions, data, parameter values, and implementation.

• A mechanistic model is not a complete representation of biology. Important processes may be omitted or simplified.
• A good fit does not prove mechanism. Multiple mechanisms can sometimes produce similar observed trajectories.
• Parameter values may not be uniquely identifiable. Different parameter combinations can generate similar model outputs.
• Unmeasured biological states can be uncertain. A model may infer intermediate variables that are difficult to observe directly.
• Virtual populations depend on their assumptions. Simulated heterogeneity does not automatically reproduce clinical heterogeneity.
• Combination predictions require validation. Simulated synergy or response differences are hypotheses until supported by evidence.
• Extrapolation can be uncertain. Predictions outside the conditions represented by the calibration data may depend strongly on model assumptions.

26. A Practical Workflow for Building an ICI QSP Model

1. Define the scientific question. Determine whether the objective concerns target engagement, dose selection, mechanism of action, combination therapy, biomarker interpretation, resistance, or another question.
2. Define the model scope. Identify the biological processes that are necessary to answer the question and avoid adding mechanisms without a clear purpose.
3. Construct the biological network. Map the relationships among drug exposure, checkpoint targets, immune populations, tumor cells, and relevant signaling processes.
4. Translate the network into equations. Define state variables, parameters, initial conditions, inputs, and outputs.
5. Assign parameter sources. Use experimental, clinical, literature, or prior model information where appropriate. Clearly distinguish measured quantities from assumed or estimated quantities.
6. Implement and verify the model. Check units, mass balance where applicable, numerical stability, limiting behavior, and implementation against the intended equations.
7. Calibrate to relevant data. Use appropriate datasets to estimate parameters or constrain uncertain mechanisms.
8. Perform sensitivity and uncertainty analysis. Determine which assumptions and parameters drive the predictions.
9. Evaluate predictive performance. Where possible, compare model predictions with data not used for calibration.
10. Simulate scenarios. Use the qualified model to explore treatment regimens, combinations, biomarkers, or mechanistic hypotheses relevant to the original question.
11. Document limitations. State which mechanisms are well supported, which are simplified, and which remain uncertain.

27. QSP vs. Conventional PK/PD Models

QSP and conventional PK/PD models are not mutually exclusive approaches. A QSP model can contain PK and PD components, but typically places those components within a larger mechanistic biological network.

The distinction is therefore largely one of scope and mechanistic representation rather than a strict separation between model types.

28. Extensions of ICI QSP Models

Once the core model is established, additional mechanisms can be introduced when supported by the scientific question and available evidence.

• Multiple checkpoint pathways.
• Immune-cell trafficking between blood, tumor, and lymphoid compartments.
• Antigen presentation and priming.
• Cytokine and chemokine networks.
• Myeloid-cell and regulatory-cell populations.
• Tumor heterogeneity and resistant subpopulations.
• Spatial or tissue-specific processes.
• Combination therapy mechanisms.
• Patient-specific biomarkers and virtual populations.
• Bayesian parameter estimation and uncertainty propagation.
• Model-based dose and regimen exploration.

The model should evolve according to the scientific question rather than simply accumulating biological detail. Each additional mechanism introduces parameters, assumptions, and potential identifiability challenges.

Where to Go Next

A natural progression is to study the individual components of an ICI QSP model in greater depth: target engagement, immune-cell dynamics, tumor growth and tumor killing, biomarker modeling, QSP sensitivity analysis, parameter estimation, and model qualification.

The next step can then be to combine these components into a complete oncology QSP workflow, including calibration, virtual populations, uncertainty analysis, and simulation of immune-checkpoint inhibitor combinations.