Why One Response Can Support More Than One Explanation
A drug-associated change in a biomarker or clinical endpoint does not, by itself, reveal the biological route that produced it. For example, a lower biomarker concentration could reflect reduced production, increased clearance, or both. Two mechanistic models can therefore agree with the same observation while implying different responses to a new dose, an intervention on a pathway, or a measurement at a different time.
The task is not simply to ask which explanation sounds plausible. It is to make each explanation precise enough to generate predictions, then find a feasible observation or perturbation for which those predictions differ. This uses the basic distinction between a biological state and an observable: models propose how states change, while measurements reveal only selected consequences of those states.
A prediction must be tied to a defined context: the population or system, treatment and exposure, time of measurement, and any co-interventions. Otherwise, apparent disagreement may come from comparing different conditions rather than different mechanisms. Likewise, a model that fits existing data is not automatically supported over another model that fits those same data.
From Shared Fit to Different Predictions
Suppose two models predict the same response at the only exposure that has been studied. They are observationally indistinguishable with respect to that observation. This does not mean that the models are identical; it means that the available evidence has not tested a condition where their predictions separate. A useful next step is to ask what each model predicts beyond the shared observation, while keeping the comparison within conditions where its assumptions are credible.
The size of a predicted difference can be calculated as a contrast. For numerical predictions \(y_A\) and \(y_B\) on the same scale, define \(\Delta=y_A-y_B\). The sign indicates which model predicts the larger value; the magnitude gives the separation in the outcome's units. A large contrast is not automatically decisive: uncertainty, measurement error, and variability in the biological system can obscure it.
A short derivation illustrates why multiple measurements may be needed. In a simple one-compartment model with constant production rate \(P\) and first-order clearance rate constant \(k\), concentration \(C\) changes according to production minus removal. At steady state, the rate of change is zero:
Thus, one steady-state concentration identifies the ratio \(P/k\), not \(P\) and \(k\) separately. For example, \(P=5\) concentration units per hour and \(k=1\) per hour give the same \(C^*=5\) units as \(P=10\) units per hour and \(k=2\) per hour. A measurement of the concentration alone cannot distinguish those explanations. A measurement of tracer disappearance, which informs clearance, can add the missing information, provided the tracer follows the relevant clearance process.
Worked Example: Testing Direct Inhibition Against Compensation
Hypothetical data and models. Let \(E\) denote drug exposure on a normalized scale from 0 to 1, and let \(Y\) be a response expressed as a fraction of its untreated baseline. Two models are proposed:
Model A represents a direct, proportional effect. Model B represents stronger direct suppression partly offset by a nonlinear compensatory response. These are deliberately simple illustrative models, not claims about a particular drug or pathway.
At the studied exposure \(E=1\), the predictions agree:
The observation at \(E=1\) therefore does not distinguish the explanations. Suppose a second, lower exposure \(E=0.5\) can be studied under otherwise comparable conditions. Model A predicts \(Y_A(0.5)=1-(0.50)(0.5)=0.75\). Model B predicts \(Y_B(0.5)=1-(0.80)(0.5)+(0.30)(0.5^2)=1-0.40+0.075=0.675\).
The contrast is \(0.75-0.675=0.075\), or 7.5 percentage points of baseline response. If the response is measured precisely enough, the second exposure can distinguish the predictions. If measurement variability is larger than this separation, collecting the second exposure may still provide little practical discrimination. A selective perturbation of the proposed compensatory process could be more informative, but its predicted effect and specificity would need to be established before interpreting it.
Designing a Discriminating Measurement or Perturbation
A discriminating test can change what is measured, when it is measured, or what biological process is perturbed. The choice should follow from an explicit contrast between model predictions. Before treating a contrast as evidence, check that the proposed intervention changes the intended process, that it does not introduce a major unmodeled effect, and that the outcome is measured on the same scale and at the predicted time.
Describe the biological processes that differ, rather than giving two labels for the same fitted curve.
Specify the observations both explanations reproduce and the conditions under which they do so.
Compare predictions across a candidate exposure, time point, assay, or selective perturbation.
Consider assay precision, biological variability, feasibility, and whether the intervention is sufficiently specific.
Worked Example: Production Versus Clearance
Hypothetical data. A biomarker has a baseline steady-state concentration of 10 nanomoles per litre. During treatment its concentration is 5 nanomoles per litre. Two explanations are considered: treatment halves production while clearance remains unchanged, or treatment doubles clearance while production remains unchanged. Assume the simple steady-state model \(C^*=P/k\) applies and that treatment has lasted long enough for steady state.
For the production explanation, let baseline \(P=10\) nanomoles per litre per hour and \(k=1\) per hour. During treatment, \(P=5\) and \(k=1\), so \(C^*=5/1=5\) nanomoles per litre. For the clearance explanation, let \(P=10\) and \(k=2\), so \(C^*=10/2=5\) nanomoles per litre. Both match the treatment observation and the baseline values; the steady-state concentration alone does not separate them.
Now suppose a labeled tracer of this biomarker is administered during treatment, and its concentration is measured one hour later. Assume the tracer is removed by the same first-order clearance process, with no further tracer input. Its concentration satisfies \(C_{\mathrm{tr}}(t)=C_{\mathrm{tr}}(0)e^{-kt}\). If both groups start at 1 nanomole per litre, the production explanation predicts \(1e^{-1}=0.3679\) nanomoles per litre after one hour. The clearance explanation predicts \(1e^{-2}=0.1353\) nanomoles per litre. The predicted contrast is \(0.3679-0.1353=0.2326\) nanomoles per litre.
This tracer measurement is informative because the two explanations specify different clearance rates. The inference depends on the stated assumptions: the tracer must share the biomarker's relevant clearance, its assay must be reliable at these concentrations, and treatment must not alter tracer distribution in a way mistaken for clearance. If those conditions fail, different tracer trajectories would not cleanly identify the proposed mechanisms.
Interventions, Timing, and Independent Evidence
A perturbation can be more discriminating than another observation under the original conditions because it probes a causal link on which the explanations disagree. For instance, if one explanation requires pathway \(X\) and another predicts that a parallel pathway is sufficient, selectively perturbing \(X\) may produce different treatment effects. The key is to compare the treatment effect within each perturbation condition, rather than comparing raw outcomes across groups whose baselines may differ.
Worked Example: Testing Whether a Pathway Is Required
Hypothetical predictions. A drug produces a benefit equal to 0.30 of baseline disease score in an unperturbed system. Two explanations fit this result: mechanism A says the benefit requires pathway \(X\), while mechanism B says it is produced through a parallel route and does not require \(X\). Consider a selective intervention that blocks \(X\), and define benefit as the vehicle-adjusted change in disease score within each condition.
Both explanations predict a benefit of 0.30 without the blocker, so that condition does not separate them. With \(X\) blocked, mechanism A predicts no additional drug benefit, or 0.00, because its required route is unavailable. Mechanism B predicts a benefit of 0.30 because its parallel route remains available. The contrast is \(0.30-0.00=0.30\) of baseline disease score.
The comparison should use drug-versus-vehicle contrasts separately in the unblocked and blocked conditions. This adjustment matters if blocking \(X\) changes disease score by itself. The result would support one explanation over the other only if the intervention actually blocks \(X\), does not also disrupt the proposed parallel route, and the observed treatment contrasts are precise enough to distinguish 0.00 from 0.30.
Timing can also separate mechanisms. A transient direct effect and a delayed feedback response may yield the same endpoint at one selected time but different trajectories. Measuring an early proximal biomarker and a later downstream response can show whether the proposed sequence is consistent with the data. However, more measurements are not automatically more informative: several measurements that all reflect the same process under the same conditions may add less discrimination than one well-chosen perturbation.
Other approaches have complementary roles. Repeating the original measurement can improve precision about the shared response but cannot, by itself, resolve a structural ambiguity if both models continue to predict the same value. A new assay may help if it measures a state closer to the mechanisms in dispute. A selective perturbation can test whether a pathway is necessary, but may be difficult to implement or may have off-target effects. Model fitting can quantify agreement with available data, but a better fit to shared observations is not a mechanistic test unless the predictions differ and the comparison accounts for uncertainty.
Diagnostics, Pitfalls, and Limits
- Check prediction separation. Calculate the contrast on a common scale and with units. A tiny contrast relative to assay error is unlikely to discriminate well.
- Check the test conditions. If timing, exposure, or background conditions differ between model predictions and the proposed experiment, the comparison may not test the intended difference.
- Check perturbation specificity. If an intervention affects several pathways, the response may not identify which pathway caused the change.
- Check what was actually observed. A noisy, saturated, or indirect assay may fail to reflect the state that separates the models.
- Check whether predictions were specified in advance. Changing a mechanism after seeing the test result can make an apparent confirmation difficult to interpret.
Common mistake: “Both models predict the observed 50% response, so the data show that both mechanisms are correct.” The observation shows that both models are compatible with that response, not that either mechanism has been established. The correction is to identify a condition where predictions differ, quantify the expected separation, and judge the resulting evidence in light of measurement uncertainty and the assumptions behind the test.
Even a well-designed discriminating test has limits. A result may favor one explanation without proving it, because an unconsidered mechanism could produce the same outcome. Conversely, a test that fails to distinguish the predictions may be inconclusive rather than evidence that the mechanisms are equivalent. State the conclusion at the level justified by the measurement: which specified explanations are more or less consistent with the result, under the conditions tested.
Check Your Understanding
Answer each question, then compare with the brief solutions.
- Two mechanisms predict the same biomarker value at one dose. Does this establish that both mechanisms are correct? What should be examined next?
- At steady state, \(C^*=P/k\). What pairs of \(P\) and \(k\) produce \(C^*=4\) concentration units?
- Two models predict benefits of 0.22 and 0.14 of baseline score under a proposed perturbation. What is the contrast, and what additional information is needed to judge whether the test is useful?
- Why compare drug-versus-vehicle effects separately in blocked and unblocked conditions when a blocker may alter baseline disease score?
Solutions. (1) No. The observation shows compatibility, not confirmation; identify a measurement or perturbation where the predictions diverge. (2) Any positive pair satisfying \(P=4k\), such as \(P=8\) units per hour and \(k=2\) per hour; one steady-state concentration does not identify the two parameters separately. (3) The contrast is \(0.22-0.14=0.08\) of baseline score. Its usefulness depends on measurement precision, biological variability, feasibility, and whether the perturbation is sufficiently specific. (4) The blocker may shift the outcome on its own, so within-condition drug-versus-vehicle contrasts isolate the incremental drug effect more fairly than raw outcomes across conditions.