Why a Mechanistic Prediction Needs a Link-by-Link Audit
A mechanistic model can connect a drug intervention to a clinical outcome through several biological states. That connection is useful when it allows a prediction beyond the observations already available, such as a new dose, schedule, species, or patient group. But a coherent chain of mechanisms is not, by itself, evidence that every link in the chain will hold in the new setting.
For example, a prediction that inhibiting a target will improve symptoms might rely on evidence about drug exposure and target engagement, then on a proposed effect on a pathway, and finally on a relationship between that pathway and symptoms. The first links may have been tested directly, while the final link may be inferred from a different population or simply assumed. An audit makes those differences visible before the prediction is used to inform a decision.
“Supported” does not mean proved for every patient, dose, or time. It means that the available evidence addresses the link in a sufficiently relevant setting to support the stated use. Conversely, an assumption may be reasonable and necessary for a prediction, but it should remain identifiable as an assumption. “Untested” does not mean false; it means the claim is not established by the evidence being considered.
How to Trace the Links
Start with the prediction as a specific claim: what intervention is being changed, what outcome is predicted, and in what context? Then write the shortest plausible causal chain connecting them. Use states and observables carefully. A measured biomarker may indicate a change in a biological state, but a measured change in that biomarker does not automatically establish that the state causes a clinical benefit.
For each proposed link, record what evidence bears on it and whether that evidence matches the predicted context. Context includes such features as species, experimental system, dose range, schedule, disease state, population, and observation period. Evidence can be strong for a link in one context yet insufficient to support carrying that link into another.
Name the intervention, outcome, and new context. Avoid broad claims such as “the model predicts efficacy” without specifying what is predicted and for whom.
List the intervening exposures, targets, biological states, biomarkers, and outcomes needed to connect the intervention to the prediction.
Identify whether evidence directly tests the link in the intended setting, whether it must be transferred from another setting, or whether the link lacks adequate evidence.
State which assumptions or untested links the prediction relies on, and avoid presenting the full chain as equally established.
A chain can be represented schematically as \(D \rightarrow X_1 \rightarrow X_2 \rightarrow Y\), where \(D\) is an intervention, \(X_1\) and \(X_2\) are intermediate states or measurements, and \(Y\) is the outcome. The arrows are causal claims, not merely associations or a sequence of measurements. If the prediction requires the chain to operate in a new context, then evidence for the links in the original context does not automatically validate the transferred chain.
Worked Audits of Proposed Extrapolations
Worked Example: Predicting a Higher Dose
Hypothetical claim: A model predicts that increasing a drug dose from 10 mg to 20 mg once daily will reduce weekly symptom episodes in adults with a particular inflammatory disease. In a small study in that patient group, 10 mg once daily produced measurable drug exposure, 70% target occupancy, and a reduction in a pathway biomarker from 100 to 80 arbitrary units over four weeks. No one in the study received 20 mg, and symptom episodes were not measured.
Trace the claim. The proposed chain is: dose \(\rightarrow\) drug exposure \(\rightarrow\) target occupancy \(\rightarrow\) pathway biomarker \(\rightarrow\) symptom episodes. The study gives direct observations of exposure, occupancy, and biomarker change at 10 mg. It does not directly measure the response to 20 mg or the link from biomarker change to symptom episodes.
Label the links. The 10 mg dose-to-exposure and exposure-to-occupancy links are supported at 10 mg in this patient group, within the conditions studied. A prediction at 20 mg requires extending those dose-related links beyond the observed dose: that extension is an assumption, not a result directly established by the 10 mg observations. The occupancy-to-biomarker link has evidence at 10 mg, but its behavior at higher occupancy is also not established. The biomarker-to-symptom link is untested in this study.
Check the numerical statement. The biomarker changed by \(80-100=-20\) arbitrary units, or \(100(80-100)/100=-20\%\). This establishes a 20% decrease in the measured biomarker in the stated study; it does not establish a 20% decrease, or any decrease, in symptom episodes. The final clinical prediction depends on a link not measured here.
Conclusion. The model may generate a 20 mg prediction, but the audit should report that the prediction relies on extrapolation beyond the observed dose and on an untested biomarker-to-symptom link. The 10 mg observations do not validate those additional steps.
Worked Example: Translating a Mechanism Across Species
Hypothetical claim: A QSP model predicts that a human drug will slow a chronic lung disease because target inhibition reduces tissue inflammation. In mice, a genetic intervention that reduces target activity was followed by lower inflammation and less tissue damage over six weeks. In human cells in culture, the drug bound the target, but inflammation and tissue damage were not measured in those cells or in treated patients.
Trace the claim. The chain is: human drug \(\rightarrow\) target inhibition in human lung tissue \(\rightarrow\) reduced inflammation \(\rightarrow\) less tissue damage \(\rightarrow\) slower disease progression. The mouse evidence addresses consequences of reducing target activity in mice. The human-cell result addresses drug binding in a culture system.
Label the links. Drug binding to the target is supported in the human-cell system described. The claim that this binding produces the required degree of target inhibition in human lung tissue is assumed, because the culture measurement does not establish target inhibition in patient tissue. The mouse results support links between reduced target activity, inflammation, and tissue damage in the mouse setting. Carrying those links to human disease is an assumption about translation across species and disease contexts. The prediction of slower progression in patients is untested by the evidence described.
Separate evidence from interpretation. The mouse intervention and the human-cell binding result are not interchangeable tests. One provides evidence about downstream biology in mice; the other provides evidence about binding in human cells. Neither alone demonstrates the entire human causal chain.
Conclusion. The prediction can be a mechanistic hypothesis worth evaluating, but it should not be described as a demonstrated human treatment effect. The species transfer, the connection from binding to inhibition in patient tissue, and the final clinical consequence remain assumptions or untested links.
Worked Example: Interpreting a Biomarker Link
Hypothetical claim: A model predicts that a treatment will improve physical function because it reduces a blood biomarker. In a 12-week study of 40 patients, the average biomarker concentration decreased from 50 to 40 ng/mL. The study did not include a measure of physical function. A separate observational analysis found that patients with lower biomarker concentrations tended to have better function.
Trace the claim. The proposed chain is: treatment \(\rightarrow\) biomarker reduction \(\rightarrow\) improved physical function. The treatment study directly measured the biomarker change, while the separate analysis measured an association between biomarker concentration and function.
Check the arithmetic. The average biomarker change was \(40-50=-10\) ng/mL. Relative to the starting average, this is \(100(-10/50)=-20\%\), a 20% decrease. These calculations describe the observed biomarker averages; they do not calculate a treatment effect on function.
Label the links. The treatment-to-biomarker link is supported for the measured study and its conditions. The biomarker-to-function link is associated with function in the separate observational analysis, but association alone does not establish that changing the biomarker causes function to improve. For this treatment and patient group, the causal biomarker-to-function link is therefore assumed or untested, depending on whether the model explicitly adopts it as a premise. The treatment study itself does not test that link because it did not measure function.
Conclusion. The data support a biomarker prediction, not a clinical-function prediction. The latter requires a causal bridge that the stated evidence does not establish.
Diagnostics, Alternatives, and Limitations
Several warning signs suggest that an extrapolation needs a more careful audit. A long chain may contain a clinical outcome that is never measured in the evidence being cited. A relationship established in animals or cells may be described as if it were established in patients. A biomarker association may be treated as proof that intervention on the biomarker changes the outcome. A study at one dose may be used to support a different dose or schedule without identifying the transfer as an assumption. In each case, the problem is not necessarily that the prediction is wrong; it is that the strength and scope of its evidence have been overstated.
A mechanistic extrapolation is not the only way to predict beyond observed data. An empirical exposure–response relationship can summarize an observed relationship and may be useful when the decision stays near the studied range. Direct measurement in the target context can test the outcome more directly, though it may require a new experiment or clinical study. A mechanistic model can connect evidence across scales and generate predictions where direct observations are sparse, but it adds causal claims that must themselves be assessed. These approaches answer different questions; a mechanistic explanation should not be treated as automatically more secure than an empirical fit.
This audit has limits. Labels depend on the prediction and evidence under review: a link supported for one dose or population may be an assumption for another. Evidence may be mixed rather than decisive, and a single label can hide differences in quality, precision, or relevance. The audit does not quantify uncertainty or prove that a model is correct. Its purpose is to make the evidential dependencies explicit so that a decision-maker can distinguish what was observed from what was carried forward or left unresolved.
Common Mistakes
- Calling the whole chain supported because one link is measured. Evidence for drug binding does not establish downstream pathway change or clinical benefit.
- Treating a plausible biological story as direct evidence. Plausibility can justify an assumption to investigate; it does not change the label to supported.
- Confusing a measured association with a causal link. A biomarker that tracks with an outcome may not cause that outcome, and changing the biomarker may not change the outcome.
- Ignoring the context of the evidence. A link supported in mice, at one dose, or over a short time does not automatically hold in patients, at another dose, or over a longer period.
- Calling an untested link false. Lack of evidence makes the link unresolved, not disproved.
Check Your Understanding
For each item, identify the relevant evidence label and explain what the evidence does and does not establish. Solutions follow the questions.
- A drug reduces a target’s activity in cultured human cells. A model predicts a reduction in patient hospitalizations. The cells were not used to measure inflammation or hospitalization. Which links are supported, assumed, or untested?
- A study in adults with a disease shows that a 5 mg dose changes a biomarker. A prediction is made for 10 mg in the same population. What must be identified in the audit?
- An observational dataset shows that patients with lower biomarker values tend to have better function. Does this establish that lowering the biomarker will improve function?
- A causal link is supported in mice but has not been measured in patients. How should it be labelled for a prediction about patients?
Solutions. (1) Target inhibition is supported in the cultured-cell system. The links from inhibition to any unmeasured inflammatory change and from that change to hospitalization are untested in the stated evidence; transferring the cell result to patient biology is an assumption. (2) The observed biomarker response is supported at 5 mg, but the dose-to-exposure and downstream response at 10 mg require extrapolation. The audit should label that extension as an assumption unless relevant 10 mg evidence is available. (3) No. The dataset establishes an association, not that intervention to lower the biomarker causes better function. That causal link remains assumed or untested for the prediction. (4) For a patient prediction, the mouse result supports the link in mice. Carrying it to patients is an assumption unless patient evidence supports the link in the relevant context.