Why Ask a Counterfactual Question?
Mechanistic models are useful not only for describing what was observed, but also for asking what might happen under a different intervention or biological condition. For example: if a drug were present rather than absent, how might a pathway marker change? If a target were reduced, what might happen to a downstream signal? These are counterfactual questions: they compare an outcome under one specified condition with the outcome that would occur under an alternative condition.
The word “might” matters. A mechanistic model can calculate a prediction under its equations and assumptions. That calculation does not by itself demonstrate that the intervention will produce the predicted outcome in an experiment or in patients. The earlier material on evidence chains helps identify the biological links a model relies on; here the focus is how to formulate the comparison and keep its interpretation within bounds.
Specify the Comparison Before Simulating
A useful counterfactual question names four things: the intervention, the comparator, the biological context, and the outcome (including its measurement time or horizon when relevant). “What if the target were inhibited?” is not yet precise enough. It does not specify how inhibition is represented, which system is being considered, what the alternative is, or what outcome should be compared.
Let \(I\) denote an intervention, \(C\) the biological context, and \(Y\) the outcome. A mechanistic model with parameters \(\theta\) produces an outcome prediction \(Y_M(I;C,\theta)\). The model-based contrast between interventions \(I_1\) and \(I_0\) is:
The subscript \(M\) emphasizes that this is a contrast calculated by the model. The two predictions hold the stated context and parameter values fixed while changing the intervention. If the outcome is a biomarker concentration in nanomoles per litre, the contrast has the same units. A negative contrast means the predicted outcome is lower under \(I_1\) than under \(I_0\); it does not automatically mean that the change is beneficial.
A counterfactual contrast is not simply a comparison between two unrelated model runs. For the comparison to answer the intended question, the runs should differ in the specified intervention, not silently in baseline conditions, parameters, or outcome definition. If the question instead concerns a change in biological condition, state that change explicitly and treat it as the intervention or scenario being compared.
Worked Counterfactuals
Worked Example: Drug Present Versus Drug Absent
Hypothetical model. A pathway signal \(S\), in arbitrary signal units, is reduced by drug occupancy \(O\), a fraction between 0 and 1. The model is \(S=100(1-0.8O)\). A downstream marker \(Z\), in nanograms per millilitre, is predicted by \(Z=20+0.5S\). Compare the drug-absent condition with a specified drug-present condition having \(O=0.70\), in the same hypothetical biological context.
For the comparator, drug absent means \(O=0\). Thus \(S_0=100(1-0.8\times0)=100\) signal units, and \(Z_0=20+0.5\times100=70\) ng/mL.
For the drug-present scenario, \(S_1=100(1-0.8\times0.70)=100(0.44)=44\) signal units. Therefore \(Z_1=20+0.5\times44=42\) ng/mL.
The model-based contrast is \(\Delta_M=42-70=-28\) ng/mL. Under this model and the specified occupancy, the predicted marker concentration is 28 ng/mL lower with drug present than with drug absent. This calculation does not establish that the drug achieves 70% occupancy in a particular system, that the equations are adequate, or that the marker change improves a clinical outcome.
Worked Example: Reducing a Target
Hypothetical model. Suppose target abundance \(T\) is expressed as a fraction of its untreated level, and a pathway output \(P\), in arbitrary units, is \(P=30+70T\). Compare untreated target abundance \(T=1.00\) with a hypothetical intervention that reduces target abundance to \(T=0.60\), holding the rest of the context fixed.
At baseline, \(P_0=30+70(1.00)=100\) units. With reduced target abundance, \(P_1=30+70(0.60)=72\) units. Hence \(\Delta_M=72-100=-28\) pathway-output units.
The model predicts a 28-unit decrease for the specified change in target abundance. This is not automatically a prediction for any particular drug: a drug may not reduce target abundance, and its effects may depend on other mechanisms represented—or omitted—in the model. Evidence is needed to connect a real intervention to the assumed change in \(T\), and to assess whether the relationship between \(T\) and \(P\) is supported.
Worked Example: Changing the Biological Condition
Hypothetical model. A biomarker \(B\), in micromoles per litre, is at steady state with \(B=R/k\), where \(R\) is its production rate in micromoles per litre per hour and \(k\) is its elimination rate in inverse hours. A treatment reduces production by 25%, while \(k\) remains 2 per hour. Compare the treatment contrast in a baseline condition with \(R=8\) and in an inflammatory condition with \(R=12\), both in micromoles per litre per hour.
In the baseline condition, untreated \(B_0=8/2=4\) micromoles per litre. Treatment changes production to \(8(1-0.25)=6\), so \(B_1=6/2=3\) micromoles per litre. The contrast is \(3-4=-1\) micromole per litre.
In the inflammatory condition, untreated \(B_0=12/2=6\) micromoles per litre. Treatment changes production to \(12(1-0.25)=9\), so \(B_1=9/2=4.5\) micromoles per litre. The contrast is \(4.5-6=-1.5\) micromoles per litre.
In this model, the treatment contrast depends on the biological condition: the predicted decrease is 1 micromole per litre at baseline and 1.5 micromoles per litre in the inflammatory condition. The difference follows from the specified model and inputs. It does not establish that real inflammation changes treatment response in this way; that requires evidence for the relevant production, elimination, and treatment-effect assumptions.
Why the Contrast Has a Causal Interpretation Only Conditionally
In causal reasoning, the counterfactual outcomes refer to outcomes under alternative interventions for a defined system or population. In an experiment, a given unit is observed under one assigned condition, not simultaneously under both alternatives. A mechanistic model can compute both outcomes for the same specified starting state, but the unobserved alternative remains a prediction. The model’s ability to calculate both is not itself evidence that its representation of the biology is correct.
The basic contrast has a short algebraic justification. In the first worked example, substituting the pathway equation into the marker equation gives:
Therefore, for drug-absent \(O_0=0\) and drug-present \(O_1=0.70\):
The result follows from the model equations and the stipulated values. The derivation does not show that the equations or values match a real system. That distinction separates a mathematical consequence of a model from a biological conclusion supported by evidence.
Interpreting a simulation as a prediction about an intervention requires, at minimum, that the intervention be well defined, the relevant causal pathways be represented adequately for the outcome, and the model inputs be appropriate for the stated context. Evidence must support the links from the intervention to its mechanistic effects and from those effects to the outcome. If an important pathway is missing, an input is poorly supported, or the intervention is unlike conditions for which the model has been assessed, the numerical output may still be computable while its interpretation is weak.
From Model Output to Evidence-Based Conclusion
It is useful to keep three claims separate:
| Claim | What supports it? | What it does not establish alone |
|---|---|---|
| “The model predicts a lower marker under intervention \(I_1\) than \(I_0\).” | The model equations, inputs, context, and calculation. | That the intervention will produce this change in a real system. |
| “The intervention changed the marker in an observed experiment.” | Relevant experimental observations and a suitable comparison. | That the change follows the model’s proposed mechanism. |
| “The intervention improves a clinical outcome through this mechanism.” | Evidence connecting intervention, mechanism, and clinical endpoint, with an appropriate study design. | More than the model can warrant if key links remain untested. |
Model predictions and empirical evidence can inform one another. A prediction can help identify a useful experiment, while new observations can challenge or support parts of the model. But agreement between a simulation and one observed result does not uniquely establish the mechanism: different mechanisms can sometimes produce similar outputs. Conversely, disagreement is a signal to examine the assumptions, inputs, measurements, and model structure rather than to treat the model output as an observed fact.
The main alternatives serve different purposes. An empirical comparison summarizes outcomes observed under different conditions, but may not explain how the difference arose or predict a new mechanistic scenario. A mechanistic simulation can trace a proposed pathway and calculate an internally consistent counterfactual, but its result depends on whether that pathway and its inputs are credible. Direct experimental evidence can test a specific intervention in a specified system, while still leaving uncertainty about other contexts or clinical outcomes. These approaches complement one another; none removes the need to match the claim to the evidence.
Diagnostics, Limitations, and Common Pitfalls
Before interpreting a counterfactual result, check whether it is answering the question actually posed. Verify that the comparator is explicit, the two scenarios use a common context and outcome definition, and every changed input corresponds to the intended intervention or condition. Inspect intermediate states and observables: an implausible change in a mechanistic state may reveal a unit error, an inconsistent input, or a structural assumption that deserves attention. Also check whether the result is driven by a parameter or pathway with little supporting evidence.
A model can fail as a guide to a counterfactual even when its calculations are correct. The equations may omit a relevant pathway; available data may not identify the parameters needed for the prediction; or the intervention may affect the system in a way not represented by the model. These limitations make the prediction uncertain or out of scope. Numerical precision does not resolve them.
A common mistake is to say, “The simulation predicts a marker decrease, so the drug causes that decrease in patients.” The first clause describes a model result; the second asserts a real-world causal conclusion. The correction is to state the result conditionally: “Under the model structure, inputs, and specified context, the predicted marker is lower by the calculated amount. Whether the drug produces that change in patients requires relevant evidence.” Likewise, a lower biomarker is not automatically a clinical benefit unless the relationship between that biomarker and the clinical outcome is supported.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions below.
- A model predicts a pathway signal of 80 units under a comparator and 65 units under an intervention. What is the model-based contrast, and what does its sign mean?
- A simulation compares a drug condition in one biological context with a no-drug condition in a different context. Why is that comparison insufficient to isolate the drug counterfactual?
- In the steady-state model \(B=R/k\), an intervention changes \(R\) from 10 to 7 micromoles per litre per hour while \(k=2\) per hour. Calculate the predicted contrast and its units.
- A model predicts a lower biomarker after a target intervention. Name one additional kind of evidence needed before concluding that the intervention improves a clinical outcome.
Solutions. (1) The contrast is \(65-80=-15\) units, so the model predicts a 15-unit lower signal under the intervention. (2) Both intervention and context changed, so the difference cannot be attributed to the intervention alone; use the same defined context for both scenarios or explicitly ask a joint scenario question. (3) The comparator is \(B_0=10/2=5\) micromoles per litre and the intervention prediction is \(B_1=7/2=3.5\) micromoles per litre. Thus \(\Delta_M=3.5-5=-1.5\) micromoles per litre. (4) Evidence must connect the intervention to a clinically relevant outcome—for example, appropriate experimental or clinical evidence assessing that outcome and supporting the proposed mechanistic link. A biomarker prediction alone does not establish clinical benefit.