Constructing Coupled State Models for Two-Drug Regimens

From Two Drug Inputs to a Dynamic Model

A two-drug regimen is not represented adequately by naming both compounds and writing down a response equation. A mechanistic model must specify what each drug reaches, which molecular or cellular state it changes, how those states influence one another, and how the measured biomarker is produced or removed. The same two exposures can therefore imply different models: the drugs may affect separate pathways that converge on one outcome, bind the same target, or act at different points in a sequential pathway.

The aim here is to build that state structure explicitly. A state is a quantity whose value changes over time and whose current value helps determine its future rate of change. Drug concentrations may be modeled as states when their kinetics matter, or supplied as time-varying inputs when pharmacokinetics have already been characterized. Target activity, pathway activity, cell abundance, and biomarker concentration can also be states. The model should include only states needed to express the proposed mechanism and connect it to observations.

A useful construction starts by drawing the causal sequence at the biological level, then writing equations for each state. For drug \(i\), a simple saturable effect function is \(E_i(C_i)=C_i/(EC_{50,i}+C_i)\), where \(C_i\) is concentration and \(EC_{50,i}\) is the concentration producing half of the maximum effect under this function. If the drug inhibits a process with maximum fractional inhibition \(I_{\max,i}\), its fractional inhibition is \(I_i(C_i)=I_{\max,i}E_i(C_i)\). This effect function is a modeling assumption, not a universal law; data or prior knowledge must support its shape and parameter interpretation.

Building Equations From State Flows

For each state, list the processes that add to it and the processes that remove or deactivate it. The simplest turnover equation has a constant input and first-order loss. If \(X\) is a normalized pathway activity, \(k_{\mathrm{in}}\) is its production rate, and \(k_{\mathrm{out}}\) is its loss-rate constant, then input minus loss gives its rate of change:

The units must agree: if \(X\) is dimensionless and time is measured in hours, both \(k_{\mathrm{in}}\) and \(k_{\mathrm{out}}X\) have units of per hour. If a drug inhibits production, replace the input with a drug-modulated input, such as \(k_{\mathrm{in}}[1-I(C)]\). If it increases loss, the drug effect belongs in the loss term instead. Those alternatives describe different mechanisms and should not be chosen merely because they fit a response curve.

For two distinct pathways, one possible model has separate drug-modulated equations, with both pathways contributing to a shared biomarker. The pathways are dynamically independent of one another in this structure, even though the biomarker depends on both. For a shared target, by contrast, both concentrations enter one target-engagement expression. In a serial pathway, an upstream state helps determine downstream production, so the downstream rate depends on the upstream state.

Parallel Effects on Distinct Pathways

Consider a model with two normalized pathway activities, \(X_A\) and \(X_B\). Drug A inhibits production of \(X_A\), and drug B inhibits production of \(X_B\). A biomarker \(Y\) is produced at a baseline rate, receives contributions from both pathways, and is cleared with first-order kinetics:

This model assumes that each drug changes only its assigned pathway production rate; the pathways do not alter one another; and the biomarker has the stated production and clearance processes. The coefficients \(\beta_A\) and \(\beta_B\) convert pathway activity into biomarker production. A fitted model cannot establish those biological assumptions by itself.

Shared Targets and Coupled Pathways

When two drugs bind the same target at mutually exclusive sites, a simple rapid-equilibrium model can represent their competition for target occupancy. Let \(C_A/K_{d,A}\) and \(C_B/K_{d,B}\) be dimensionless concentration-to-dissociation-constant ratios. If free concentrations are treated as fixed and target binding equilibrates quickly relative to the modeled response, the fractional occupancies are:

The remaining unoccupied fraction is \(1-\theta_A-\theta_B\). If both ligands fully inhibit target activity while bound, an activity state can be driven by this unoccupied fraction. If their efficacies differ, occupancy alone is insufficient: the model must specify the activity associated with each bound state. The rapid-equilibrium assumption also fails when binding kinetics are slow enough to influence the observed time course.

A distinct form of coupling occurs when the drugs act at different locations in a serial pathway. For example, drug A may inhibit formation of an upstream state \(X\), while drug B inhibits conversion of \(X\) into a downstream state \(Z\). A minimal model is:

Here \(X\) affects the production rate of \(Z\), so the states are coupled through the pathway's flow. The product \(X[1-I_B(C_B)]\) expresses that downstream production requires both available upstream activity and uninhibited conversion. It is a consequence of this proposed mechanism, not an additional parameterized interaction term. A dynamical dependence between states should not automatically be interpreted as a statistical claim about combination interaction.

Deriving and Checking a Steady-State Result

For a first-order turnover state with constant input \(u\), \(\frac{dX}{dt}=u-kX\), a steady state is a value \(X_{\mathrm{ss}}\) for which the rate of change is zero. Setting the derivative to zero gives \(0=u-kX_{\mathrm{ss}}\), and, when \(k>0\), solving yields \(X_{\mathrm{ss}}=u/k\). The condition \(k>0\) matters: it ensures that the state is removed and that the equilibrium is stable for this one-state model. To check stability, write \(X=X_{\mathrm{ss}}+\delta\); then \(d\delta/dt=-k\delta\), so \(\delta(t)=\delta(0)e^{-kt}\), which approaches zero as time increases.

For coupled systems, a steady state requires all state derivatives to equal zero simultaneously. Solving one equation while treating a changing upstream state as fixed can produce a false equilibrium. Further, a mathematically valid steady state may be biologically impossible, for example if a nonnegative abundance state is predicted to be negative. Such predictions indicate a structural or parameterization problem, or that the model is being used outside its intended range.

Diagnostics, Alternatives, and Limitations

A coupled state model is useful when timing and mechanism matter: it can predict delayed responses, distinguish a rapid target effect from slower biomarker turnover, and expose how an upstream state propagates to a downstream measurement. But additional states and links add assumptions and parameters. A model with more detail is not automatically more informative, especially when available data measure only the final biomarker.

Several checks help identify a model that is going wrong:

• Check units and signs. Each term in a state equation must have units of state per time. Production should not be negative unless the equation explicitly represents a signed flux or another appropriate quantity.
• Check limiting cases. At zero drug concentration, a drug-effect function should return its defined baseline effect. At high concentration, it should approach its stated maximum effect rather than diverge.
• Check positivity and plausible bounds. Concentrations, abundances, and many activity measures cannot be negative. A normalized activity may also be expected to remain between zero and one.
• Inspect time-course patterns. A model that matches a final response but predicts the wrong onset, peak, or recovery may have an incorrect turnover or pathway structure.
• Check what the data actually inform. If only \(Y\) is observed, multiple combinations of pathway rates and biomarker coefficients may produce similar predictions. Do not interpret a parameter as uniquely determined without adequate information.

A parallel-pathway model is a reasonable alternative when there is evidence that two pathways respond separately and only converge at the measured outcome. A shared-target model is appropriate when both compounds engage the same target and binding assumptions are credible. A serial model is needed when an upstream state controls downstream production. In contrast, an empirical response model can summarize observed outcomes with fewer mechanistic claims, but it may not represent internal states or predict unmeasured pathway quantities. The choice should reflect the scientific question and the available measurements, not just which model has the most terms.

A final limitation is that a plausible diagram does not establish causality or identify every rate. State relationships should be grounded in pharmacology and checked against time-course and target-engagement data where possible. If a proposed mechanism is uncertain, compare defensible alternative structures and report which predictions are robust to that choice.