CAR T-Cell Antigen Recognition and Activation

From Antigen Contact to T-Cell Activation

A CAR T cell can encounter a tumor cell that expresses the target antigen, bind that antigen through its chimeric antigen receptor (CAR), and receive signals that initiate activation. These are related but distinct events. A contact does not guarantee receptor binding; binding does not guarantee activation; and activation does not by itself specify how much the CAR T-cell population will later expand or how much tumor it will kill.

A useful QSP representation therefore separates the steps. The first part describes the chance that a contact has relevant CAR-bound recognition. The next part describes the conditional chance that this recognition triggers activation. Expansion and killing belong to downstream processes with their own parameters and state variables. Keeping these stages distinct prevents a fitted activation parameter from being mistaken for a measure of proliferation or cytotoxicity.

The probability \(q_{\mathrm{bound}}\) is a property of the contact-level event being modeled. It is not generally the fraction of CAR molecules that are occupied. Denote the fraction of CAR molecules occupied by antigen by \(\theta\). A value of \(\theta=0.5\) means that half of the receptors are occupied on average; it does not, without further assumptions, mean that half of the contacts involve any bound CAR.

Defining the Recognition Event at the Right Level

One possible model connects receptor occupancy to contact-level recognition. Suppose a CAR T cell has \(R_T\) receptors, each receptor is independently occupied with probability \(\theta\), and a contact counts as recognized if at least one receptor is occupied. Then the probability that none of the \(R_T\) receptors is occupied is \((1-\theta)^{R_T}\). The complementary probability gives the contact-level recognition probability:

This expression is conditional on all three assumptions: the receptor number is fixed at \(R_T\), occupancy is independent across receptors, and any occupied receptor is sufficient to classify the contact as having relevant bound recognition. If recognition requires multiple simultaneous bonds, particular spatial arrangements, or a minimum duration of binding, this expression is not adequate without modification. In those settings, use a contact-level probability estimated or derived for the required recognition event.

The rate equation follows by considering contacts in a short time interval. If contacts occur at rate \(c\), the expected number of contacts in a small interval of length \(\Delta t\) is approximately \(c\Delta t\). Each contact results in activation with probability \(q_{\mathrm{bound}}p_{\mathrm{act}}\), so the expected number of activation events in the interval is approximately \(c\Delta t q_{\mathrm{bound}}p_{\mathrm{act}}\). Dividing by \(\Delta t\) and taking the short-interval limit yields \(\lambda_{\mathrm{act}}=c q_{\mathrm{bound}}p_{\mathrm{act}}\). This multiplication is valid because the probabilities are defined at successive stages of the same contact event.

A common modeling assumption is that contacts occur as a Poisson process at a constant rate, and that contact outcomes are independent with the same probabilities \(q_{\mathrm{bound}}\) and \(p_{\mathrm{act}}\). Under this assumption, activation events also form a Poisson process, with rate \(\lambda_{\mathrm{act}}\). If the modeled event is a cell’s first activation, the probability it has activated by time \(t\) is:

This is a first-event probability, not a prediction of the number of activated descendants. It assumes the cell remains available to be activated until its first activation and that the rate is constant over the time interval. If contact rates or activation probabilities change with time or cell state, the constant-rate expression must be replaced with a time-varying model.

Activation Is Not Expansion or Killing

A population model can make the separation explicit. Let \(U(t)\) denote the number of not-yet-activated CAR T cells in a cohort, and let \(N_0=U(0)\). If all cells have the same constant first-activation rate \(\lambda_{\mathrm{act}}\), then:

The cumulative number of cells that have experienced first activation by time \(t\), in a model that does not yet include expansion or loss, is \(N_0-U(t)\). This bookkeeping describes entry into an activated state. To represent post-activation expansion, one needs a separate population balance with an expansion or net-growth rate. To represent tumor killing, one needs a tumor-cell loss term driven by effectors and a stated killing mechanism. Neither process follows automatically from the activation rate.

For example, even if two settings have the same fraction activated, they may produce different later CAR T-cell numbers because their post-activation net growth differs. Conversely, a large activated population need not imply a proportionally large tumor response if the killing process is limited. An activation model answers a narrower, useful question: how often does antigen-dependent CAR recognition initiate activation under the modeled conditions?

Assumptions, Diagnostics, and Model Choices

The simple rate model is useful when its event definitions and assumptions match the question. Before interpreting its parameters, check that the contact rate refers specifically to contacts with antigen-bearing tumor cells, that \(q_{\mathrm{bound}}\) refers to the defined recognition event, and that \(p_{\mathrm{act}}\) is conditional on that event. If \(c\) instead counts all tumor contacts, including antigen-negative contacts, the antigen-positive fraction must be included or the contact rate must be redefined. Otherwise, the model overstates opportunities for antigen-dependent activation.

Diagnostics should focus on the event sequence and the time scale represented. If a constant-rate model predicts a constant activation hazard but observed activation rates change systematically over time, the assumptions may be failing: contacts may change, cells may become less or more responsive, or the available cell population may change. If estimated \(q_{\mathrm{bound}}\) exceeds one, or if \(c q_{\mathrm{bound}}p_{\mathrm{act}}\) has units other than inverse time, there is a definition, data, or unit error. When receptor occupancy is used to construct \(q_{\mathrm{bound}}\), check whether independence and the “any occupied receptor” rule are biologically defensible for the intended application.

There are two common modeling choices. A contact-level model specifies \(q_{\mathrm{bound}}\) directly, which can be practical when recognition frequency is measured at the contact level or receptor-level data are unavailable. A receptor-level model specifies receptor number and occupancy and derives the contact probability; it is more mechanistic but requires additional assumptions and information. A deterministic population equation, such as \(dU/dt=-\lambda_{\mathrm{act}}U\), gives the mean cohort behavior under a shared rate. An event-based stochastic model can instead represent individual contact and activation events when cell-to-cell variation or small populations matter. These alternatives change the level of detail, not the need to distinguish recognition, activation, expansion, and killing.

Important limitations remain. A single \(q_{\mathrm{bound}}\) and \(p_{\mathrm{act}}\) can average over variation in antigen-bearing contacts, receptor abundance, and cell responsiveness. If that variation is important, a population-average model may conceal distinct subgroups. The independent-occupancy expression is only one possible bridge from receptor occupancy to contact recognition; it is not a universal law of CAR engagement. Finally, a fitted product \(q_{\mathrm{bound}}p_{\mathrm{act}}\) may not identify the two probabilities separately without data that inform each stage. A good model makes that distinction explicit rather than assigning an unsupported biological meaning to the product.

Common Pitfall

A related mistake is to interpret a high activation rate as a direct prediction of strong expansion or tumor killing. The correction is to assign separate states and parameters to those processes. Activation supplies an input to downstream models; it is not a substitute for them.