Worked Model Comparison for a Feedback-Driven Drug Combination

Why Compare Mechanistic Models?

A combination can appear to outperform either single agent at one measurement time without establishing why it does so. An explicit interaction term can describe a combination response that differs from a chosen reference model, but it does not by itself identify a biological mechanism. In contrast, a feedback-driven model represents an adaptive change in pathway activity over time. Measurements from single agents and their combination, collected at more than one time, can therefore test whether the time course is consistent with feedback or is adequately summarized by a time-invariant interaction.

The comparison below uses a deliberately simulated example. Inhibitor A suppresses a primary disease-driving pathway, while inhibitor B blocks a compensatory pathway activated by A. The outcome \(E\) is disease-signal suppression in percentage points relative to untreated baseline: larger \(E\) indicates greater suppression. The inputs \(a\) and \(b\) are fractional pathway inhibition, between 0 and 1, at the selected doses. They are treated as known from separate measurements; this example does not estimate pharmacokinetics or dose-to-occupancy relationships.

Models and Assumptions

For the feedback model, let \(F(t)\) denote compensatory pathway activity induced by inhibitor A. Suppose that after treatment begins it follows first-order adaptation toward a steady state proportional to A's inhibition:

Here \(\tau\) is the adaptation time constant, in hours. This equation assumes a constant drug input over the observation period, a fixed inhibition level \(a\), and a single adaptation timescale. B blocks the contribution of this compensatory activity in proportion to \(b\). With direct response coefficients \(c_A\) and \(c_B\), and feedback magnitude \(c_F\), the response model is:

Solving the state equation gives \(F(t)=a(1-e^{-t/\tau})\), so the response can be written directly as:

This formulation assumes that feedback reduces the measured suppression, that B blocks feedback proportionally, and that the direct effects are constant over the observation window. It is an approximation: real pathways may have multiple feedback loops, delays, or saturation. The model is useful here because its biological claim makes a time-course prediction that can be checked against observations.

For comparison, consider a time-invariant explicit-interaction model:

The coefficients \(\beta_A\) and \(\beta_B\) describe the single-agent contributions on this response scale; \(\beta_{AB}\) describes departure from their additive combination on that scale. This model assumes that the coefficients do not change with time. A nonzero \(\beta_{AB}\) is not, by itself, evidence for a specific feedback mechanism. Both models assume measurements are on a common scale, inputs are accurately known, the untreated baseline is comparable across conditions, and the sampled time points are correctly recorded.

Worked Simulated Data

Suppose a hypothetical experiment measures suppression at 0, 8, and 24 hours. The chosen doses produce \(a=0.8\) for A and \(b=0.6\) for B. For this illustration, \(\tau=8\) hours is prespecified from separate information, rather than estimated from these nine observations. The simulated feedback model uses \(c_A=50\), \(c_B=30\), and \(c_F=40\), with coefficients expressed in percentage points per unit of fractional inhibition.

Fit the Explicit-Interaction Alternative

Fit the time-invariant model by ordinary least squares: choose the coefficients that minimize the sum of squared differences between observed and predicted suppression. Because the model has no time term, the three measurements for a given regimen receive the same prediction. The vehicle observations are zero and do not affect the coefficient estimates.

Interpret the Comparison, Then Choose a Regimen

A residual sum of squares measures disagreement between observed and predicted values on the chosen response scale. It is meaningful to compare such measures when the models are fitted to the same observations and scale. Here the striking fit difference illustrates the consequence of omitting time dependence. It does not imply that a feedback model will always win in experimental data: real measurements contain noise, and a more flexible model can fit chance variation. In practice, inspect residuals against time and regimen, assess whether predictions remain plausible, and evaluate prediction on data not used for fitting when such data are available.

Suppose the decision criterion in this hypothetical example is to choose a regimen predicted to achieve at least 40% suppression at 24 hours, while keeping a separately assessed tolerability burden at or below 1.5 arbitrary units. Suppose the simulated burden is 0.8 units for A, 0.5 for B, and 1.4 for A + B. The feedback model predicts 9.59% suppression for A alone, 18.00% for B alone, and 45.84% for the combination at 24 hours. Only A + B meets both stated criteria. This is a conditional choice for the specified simulated inputs and criteria, not a clinical recommendation. If tolerability estimates, exposure, or uncertainty changed, the choice could change too.

Diagnostics, Alternatives, and Limitations

Model checking should focus on patterns that expose the assumptions, not only on one overall fit statistic. A feedback model is questionable if its residuals show a consistent time trend, if the observed late response rebounds in the opposite direction, or if feedback appears equally strong with and without B despite the assumed blocking mechanism. The explicit-interaction model is inadequate when its residuals systematically vary with time or dose. Unexpectedly large or implausible fitted coefficients can also indicate a scale mismatch, poor input measurements, or an omitted process.

The main alternative depends on the question. A time-invariant interaction model is simpler and can be preferable when the endpoint is defined at one prespecified time and no dynamic mechanism is needed. A dynamic feedback model is useful when adaptation over time matters to the decision. Other possibilities include fitting a reference-null model for a specified additivity question, or extending the dynamic model to represent measured drug exposure and more than one pathway state. Those choices change the scientific question and assumptions; a better numerical fit alone does not decide which is appropriate.

This comparison has important limits. The inputs \(a\) and \(b\) are treated as known and constant, feedback is represented by one first-order state, and the feedback timescale is fixed. The example has no measurement error, between-unit variability, or uncertainty in tolerability. A model that fits these data may still make poor predictions at different doses, schedules, or observation times. Conversely, failure to distinguish models with sparse measurements is not proof that their mechanisms are equivalent. The next experimental design should include informative time points and single-agent controls, as well as combination conditions that test whether B modifies the proposed feedback.

Common Pitfall

A common mistake is to call a positive fitted \(\beta_{AB}\) proof that B blocks feedback, or proof of synergy in every possible sense. The coefficient only describes a departure from additivity for the stated equation, outcome scale, doses, and time handling. In this example, the interaction fit averages over time and cannot represent the observed adaptation. The correction is to compare the models' predictions against the full time-course data and describe a mechanistic interpretation as supported only to the extent that its assumptions and predictions are consistent with those data.