Separating Parameters, Inputs, and Model State

Why Separate the Model’s Three Roles?

An ODE model can be written mathematically as a rule for how its state changes over time. In software, that rule is evaluated repeatedly at times and states chosen by the numerical solver. The implementation is easier to reason about when it keeps three roles distinct: parameters describe fixed model quantities, inputs provide externally specified values that may vary with time, and states are the quantities whose changes the ODE defines.

This separation is more than a matter of tidy code. If an input is accidentally held fixed, or a state is mistakenly treated as a parameter, the program may calculate a different model from the one intended. A clear interface makes it possible to inspect what affects a derivative and to test each dependency without changing the model’s other components.

The role is determined by how a quantity enters the model, not by its biological name. Drug concentration can be an externally supplied input in a pharmacodynamic model, or it can be a state in a pharmacokinetic model that calculates its own concentration trajectory. A quantity’s role should be consistent throughout the equations and the implementation.

An Exposure-Driven Turnover Model

Consider a response amount or concentration \(R(t)\) with constant production and exposure-enhanced loss. Let \(C(t)\) denote drug exposure supplied to the turnover model. One possible model is

Here, \(R(t)\) is the state. The parameters are \(k_{\mathrm{in}}\), \(k_{\mathrm{out}}\), \(E_{\max}\), and \(EC_{50}\). The time-varying input is \(C(t)\). The fraction \(C/(EC_{50}+C)\) describes a saturable exposure effect: for nonnegative exposure and positive \(EC_{50}\), it is zero at zero exposure and approaches one as exposure increases.

For dimensional consistency, if \(R\) has units of amount per volume and time is measured in days, then \(k_{\mathrm{in}}\) has units of amount per volume per day, while \(k_{\mathrm{out}}\) has units of inverse days. \(E_{\max}\) is dimensionless, and \(C\) and \(EC_{50}\) must have the same units. The parenthetical multiplier is dimensionless, so the loss term has the same units as production: amount per volume per day.

The corresponding software interface should make all three roles visible. In words, the right-hand-side function receives the current time, the current state vector, the fixed parameters, and the input value or input function evaluated at that time; it returns the derivative vector. For this one-state model, it computes \(f(t,R,\theta,C(t))\) and returns \(dR/dt\). It should calculate from the supplied arguments rather than silently retrieving exposure from an unrelated global variable or changing the state in place.

Derivatives Reveal Which Role a Quantity Plays

A useful justification for the separation is to inspect how the right-hand side depends on its arguments. In the turnover model, \(R\) appears as the current state, the four fixed quantities enter as parameters, and \(C(t)\) is an input. Holding the state and parameters fixed while evaluating the function at another exposure changes only the input argument.

For this model, the change in the derivative with exposure can be expressed by a partial derivative. Differentiating the saturable fraction gives

Thus, under the stated assumptions,

This derivative is nonpositive. It is strictly negative when \(R\ne 0\), \(E_{\max}\ne 0\), \(k_{\mathrm{out}}>0\), and \(EC_{50}>0\). Consequently, increasing exposure then decreases the derivative, because exposure increases the loss term. If \(R=0\), however, the exposure-dependent loss term is zero, so changing exposure does not change the derivative. The dependency test must account for the current state and parameters, not just the formula’s appearance.

A related check is to vary one role at a time. At fixed state and parameters, changing exposure should change the derivative when \(E_{\max}\ne0\), \(R\ne0\), and the exposure change changes the saturable fraction. At fixed input and parameters, changing \(R\) changes the loss term in proportion to \(R\). Changing a parameter changes the model’s rule, not the current state or the external exposure value. These are dependency checks, not universal claims that every change must alter every derivative: for example, at \(R=0\), the loss term vanishes for any exposure.

Implementation Checks and Common Failure Modes

A practical right-hand-side evaluation can be checked in stages. First confirm that the input is evaluated at the requested time and that parameters are passed as fixed values. Then confirm that the state argument—not a separately stored or previously calculated response—is used in the derivative. Finally, check that the returned value is a derivative with the correct units, rather than a new state value.

A common mistake is to store exposure in a global variable and let the ODE function read it implicitly. This hides a model dependency: the same stated arguments can then produce different derivatives depending on external program state. Pass exposure explicitly, or pass a clearly defined input function that can be evaluated at the requested time.

Another mistake is to update the state inside the right-hand-side function. The right-hand side should evaluate the derivative at the solver’s supplied state; it should not advance the state itself. The solver uses repeated derivative evaluations to perform that advancement. Mutating the supplied state can make the calculation depend on evaluation order and can interfere with the solver’s control of the integration.

A third pitfall is treating every biologically measured quantity as a parameter. If a value changes with time and is supplied from outside this ODE, it is an input here. Conversely, if exposure is generated by pharmacokinetic equations in the model, it belongs in the state vector, with its own derivative; that is a different model structure, not merely another way to pass the same external input.

Choosing the Representation and Its Limits

The explicit-input formulation is appropriate when exposure is supplied as known or externally generated information and the aim is to calculate the response conditional on that exposure. Representing exposure as a state is preferable when its trajectory is to be generated by the model’s own dynamic equations and coupled to the response. Treating a time-varying exposure as one fixed parameter is only appropriate if it truly remains constant over the integration interval.

This separation does not establish that the exposure-effect relationship is biologically correct, identify parameter values, or account for uncertainty in the supplied input. Nor does it determine how an input should be represented when it changes abruptly. Those are separate modeling and numerical questions. The present structure makes the chosen roles explicit so such decisions can be made without confusing them with state updates or parameter values.