Mechanistic States, Observables, and Clinical Endpoints

Three Different Roles in a QSP Model

A QSP model often describes biology that cannot be observed directly in a patient, while also using laboratory measurements and clinical outcomes to represent evidence that can be collected. These quantities are related, but they are not interchangeable. A useful first step is to ask of every quantity: is it part of the model’s description of the biological system, is it something an assay records, or is it an outcome that matters clinically?

For example, a model might represent the amount of activated immune cells in tissue as an internal biological state. A blood assay might measure a cytokine associated with that process. A clinical endpoint might record the number of symptom flares over a specified period. The state, the assay result, and the endpoint may be connected, but each answers a different question.

These labels describe a quantity’s role in a particular model, not an absolute property of nature. A biological state can sometimes be measured directly in a particular experimental setting. Conversely, an assay result may be used as an internal state in a simplified model. What matters is whether the quantity is being treated as part of the model’s biological representation, as recorded data, or as an outcome used to assess patient status or benefit.

From a Hidden Process to Recorded Data

Let \(x(t)\) denote a mechanistic state at time \(t\). The model specifies how that state changes over time. For instance, a state could be the amount of a protein in a tissue compartment. An assay generally does not report \(x(t)\) automatically. Instead, it produces an observation \(y(t)\), related to the state through a measurement rule. A simple representation is:

Here, \(h\) represents how the assay translates the biological quantity into its measurement scale, and \(\epsilon(t)\) represents measurement error or other variation not captured by that rule. The function \(h\) might include a scale factor, a background signal, or a nonlinear response. In a model with several biological states, the assay can depend on more than one state.

An endpoint has its own definition and measurement procedure. It may be a continuous score at a visit, a count accumulated over a period, or an event time. It need not be a direct reading of any single state. A model could represent an endpoint as an additional output, but that does not make the endpoint identical to the biological states used to describe the system.

Worked Example: A Signaling State and a Blood Assay

Worked Example: An Assay Is Not a Clinical Endpoint

This distinction protects against a common overinterpretation: a biomarker can be measured reliably and can vary with a biological process without being a validated substitute for a clinical outcome. The current focus is on identifying the different quantities and their roles, not on assuming or establishing a causal or predictive link between them.

Why a Measurement May Not Identify a State

An assay can combine contributions from several biological states. Then different underlying biological situations can produce the same recorded value. For instance, let \(A\) and \(C\) be two hypothetical model states, both in model units, and suppose an assay reports their sum:

An assay result of \(10\) units is consistent with \(A=7,\ C=3\), and also with \(A=2,\ C=8\). The assay value alone cannot distinguish these possibilities. This is not necessarily an assay failure: it may simply be that the measurement captures a combined signal rather than each state separately.

The general lesson is that a measurement may inform a state without identifying it uniquely. Whether it does so depends on the measurement rule, the number of relevant states, the available data, and assumptions in the model. A model’s internal state values should therefore be interpreted as quantities conditional on that model structure and its evidence, not as direct assay readings.

Assumptions, Checks, and Failure Modes

Before interpreting a state-to-measurement or state-to-endpoint mapping, make its assumptions explicit. At a minimum, check that the quantities have clear definitions, the units are compatible, and the measurement rule is appropriate for the assay or endpoint being represented. Also ask whether the model’s time scale matches the observation schedule: a transient biological state and a monthly symptom count do not describe the same time window.

Useful checks include comparing predicted observables with measurements on the same scale, checking whether assay background or detection limits have been represented where relevant, and looking for systematic discrepancies across time or groups. If residual differences between predicted and measured values show a consistent pattern, the measurement mapping or the biological model may be inadequate. Agreement with one assay alone is not proof that every internal state has been correctly represented.

• Unit mismatch: Treating an amount in tissue as though it were a concentration in plasma without a defined conversion.
• State-observation confusion: Calling a predicted latent state an assay result, even though an observation rule is required to connect them.
• Endpoint substitution: Treating a biomarker change as though it were itself a change in symptoms or clinical benefit.
• False uniqueness: Inferring one precise combination of states from an assay that can be generated by several combinations.
• Time-scale mismatch: Comparing a momentary model state directly with an endpoint accumulated across a longer period without specifying how the time windows relate.

A more direct analysis may summarize the recorded assay or endpoint without representing internal biology. That can be suitable when the question concerns the measured quantity itself. A mechanistic-state representation can instead make biological quantities explicit, but requires assumptions about how those quantities evolve and how they relate to observations. Neither representation automatically supplies information that its data and assumptions do not support.

The introductory material on QSP gives the broad motivation for representing biological systems mathematically. The practical boundary emphasized here is that every model output should be named according to what it represents. A state is not automatically observable; an observable is not automatically a clinical endpoint; and similarity or association across these layers does not by itself establish that one can stand in for another.

Worked Example: Checking Units Before Interpreting a Prediction