Why Mechanism Inclusion Is a Decision, Not a Checklist
A biological mechanism may be real and well established yet unnecessary to represent explicitly in a particular quantitative systems pharmacology model. Conversely, a mechanism supported by limited evidence may warrant representation if plausible changes in it could alter the model output used for a development decision. The question is not simply whether a mechanism exists; it is whether representing it improves the model for its defined use.
The earlier material on tracing causal paths to set the mechanistic boundary helps identify candidate mechanisms connected to the target output. This tutorial applies a further screen to each candidate: does it affect that output, does it respond in relevant intervention scenarios, and can it plausibly vary enough to matter? The answer guides whether to retain an explicit state, use a simpler representation, or exclude the mechanism from the model.
Before applying the tests, specify the decision context. Let \(d\) denote a relevant intervention or intervention scenario, \(m\) represent the candidate mechanism, and \(y\) be the target model output. The output might be a predicted biomarker change, a clinical endpoint, or a quantity used to compare dose options. The tests presume that the output and the relevant decision have been named, that model comparisons hold other assumptions fixed as far as practicable, and that the plausible range of \(m\) is supported by biological or empirical knowledge. The tests do not replace evidence assessment: if the range or effect is poorly known, the result may be inconclusive.
Three Tests for a Candidate Mechanism
The tests are related but not interchangeable. A mechanism can affect an output at baseline without responding to the interventions being compared. It can respond strongly to an intervention yet have a negligible effect on the target output. It can also have a large effect in one assumed setting but little plausible variation in the context of use. Consider all three rather than treating any single positive test as sufficient.
- Effect on the target output: Holding the relevant context and other model components fixed, does representing plausible behavior of the mechanism change \(y\) by more than the precision or decision tolerance appropriate to the use?
- Intervention responsiveness: Across the interventions relevant to the decision, does the mechanism change, and does that change contribute to a material difference in \(y\) or its interpretation?
- Plausible variation: Can the mechanism plausibly differ across people, conditions, or intervention scenarios in a way that changes \(y\) or the decision? Distinguish such variation from mere uncertainty about a fixed quantity.
“Material” needs a context-specific meaning. It may refer to a prespecified output tolerance, such as a maximum acceptable prediction difference, or to crossing a decision threshold. A small output difference can matter if it changes which option is preferred; a larger difference may not matter to the decision if both options lead to the same action. There is no universal numerical cutoff for all models.
A useful comparison is between an explicit representation of \(m\) and a simpler alternative: for example, fixing it at a justified reference value or replacing its detailed behavior with an aggregate relationship. Compare predicted outputs under the same relevant conditions. If all plausible values lead to output differences below the accepted tolerance and leave the decision unchanged, an explicit state may not be warranted for that use. That conclusion is conditional on the scope and plausible values examined.
Worked Example: Does Output Impact Cross a Decision Threshold?
Suppose a hypothetical model is being used to compare a treatment with a decision threshold of at least \(20\%\) predicted reduction in a disease biomarker. In the same population and intervention scenario, a version that represents a candidate repair mechanism explicitly predicts a \(22\%\) reduction. A version that omits its changing behavior but retains the reference baseline predicts a \(17\%\) reduction. Assume both predictions use the same remaining model assumptions, and the comparison is intended as a screening calculation rather than a validated clinical prediction.
The output difference is
The explicit-mechanism prediction exceeds the \(20\%\) threshold by \(2\) percentage points, while the simplified prediction falls below it by \(3\) percentage points. Thus, the two representations imply different decisions in this hypothetical case. The mechanism passes the output-impact screen for this decision context. This does not establish that the explicit version is correct: evidence for the mechanism’s behavior and the predictions still need assessment. It does show that omission cannot be justified merely by calling the mechanism biologically secondary.
From a Positive Screen to a Representation Choice
Passing a screen does not automatically require a separate dynamic state. Representation should be no more detailed than needed to preserve decision-relevant behavior. Use an explicit state when the mechanism’s own behavior must be tracked to distinguish relevant scenarios and a simpler representation cannot preserve the needed output differences. Use a simplified representation when an aggregate effect or fixed relationship captures those differences adequately. Exclusion is reasonable when the tests indicate no material contribution within the stated scope and plausible range.
One way to justify a screen is to bound the output change associated with a candidate mechanism. Suppose, for this local comparison, that the output can be written as a differentiable function \(y=f(m)\), with the other context held fixed. Suppose also that over the plausible interval from \(m_0\) to \(m\), the sensitivity satisfies \(\left|df/dm\right|\leq L\). By the mean value theorem, for some value between \(m_0\) and \(m\),
If plausible variation is bounded by \(|m-m_0|\leq\delta\), then the output difference is at most \(L\delta\). If that bound is smaller than a prespecified output tolerance, this supports using the reference representation for that output under the stated assumptions. The argument is only as reliable as its sensitivity bound and plausible range. It also cannot settle a decision that depends on crossing a threshold unless the bound is considered relative to that threshold.
Worked Example: Does the Mechanism Respond to the Intervention?
In a hypothetical dose-comparison model, a candidate pathway activity \(m\) is scaled from \(0\) (inactive) to \(1\) (reference activity). A low intervention is predicted to leave \(m=0.8\), while a high intervention is predicted to reduce it to \(m=0.2\). For this screening example, assume the model’s predicted biomarker benefit is \(y=18+16(1-m)\%\), with other inputs unchanged. The decision threshold is \(25\%\) benefit.
At the low intervention:
At the high intervention:
The predicted pathway change is \(0.8-0.2=0.6\) activity units, and the output difference is \(30.8\%-21.2\%=9.6\) percentage points. The low-intervention prediction is below the decision threshold; the high-intervention prediction is above it. Under these hypothetical assumptions, the mechanism responds to the intervention and that response contributes to a decision-relevant output difference. An explicit representation is warranted unless a simpler model can preserve this contrast. The calculation does not show that the intervention causes a clinical benefit; it reports the implications of the stated model.
Plausible Variation, Not Just Possible Variation
A mechanism’s mathematical ability to vary is not evidence that it varies meaningfully in the modeled context. Plausible variation should be supported by relevant biological knowledge, observations, or a clearly stated range of conditions. It may reflect differences between individuals or conditions of use; it should not be confused with uncertainty about a parameter that is assumed fixed. Both may affect predictions, but they support different conclusions about whether the model needs to represent variation in the mechanism itself.
Worked Example: When Plausible Variation Is Too Small to Matter
Suppose a candidate mechanism has a plausible activity range \(m=0.2\) to \(m=0.8\) in the modeled population. A screening model gives the target output, in percentage points of biomarker benefit, as \(y=20+2m\). The decision threshold is \(25\%\), and the team has prespecified an output tolerance of \(2\) percentage points for this comparison. Assume the range and output relationship are credible for this hypothetical use.
At the lower end of the range, \(y=20+2(0.2)=20.4\%\). At the upper end, \(y=20+2(0.8)=21.6\%\). The full output range attributable to plausible mechanism variation is
This is below the \(2\)-percentage-point tolerance, and neither end reaches the \(25\%\) decision threshold. For this use, representing the variation explicitly is not supported by the screen. A simplified representation could use the midpoint \(m=0.5\), giving \(y=20+2(0.5)=21.0\%\), or the mechanism could be excluded if retaining its reference contribution is also unnecessary. The conclusion is limited to this output, range, and decision; it does not establish that the mechanism is unimportant in other contexts.
A Practical Inclusion Workflow
Name the modeled population, relevant interventions, target output, decision threshold if applicable, and an acceptable output tolerance. Without these, “important” has no operational meaning.
State what the mechanism represents and what changes in it are plausible for the context. Separate supported variation from uncertainty or speculation.
Compare explicit and simpler representations under the relevant scenarios. Check both the size of output differences and whether they alter interpretation or the decision.
Retain an explicit state if its behavior matters and cannot be preserved adequately by a simpler representation. Simplify when an aggregate representation preserves what the decision needs. Exclude only when the tests support negligible relevance within the stated scope.
A useful comparison of the choices is shown below. These are modeling decisions for a defined use, not biological classifications.
| Choice | When it is supported | What to check |
|---|---|---|
| Explicit state | Its plausible behavior materially changes the target output or decision across relevant scenarios, and simpler alternatives do not preserve that behavior. | Whether the state’s behavior is supported well enough for the intended comparison. |
| Simplified representation | The mechanism contributes, but a compact relationship or reference contribution preserves the decision-relevant output adequately. | Whether simplification retains the relevant intervention contrasts and output accuracy. |
| Exclusion | Its plausible contribution is negligible for the named output and decision, within the stated context and tolerance. | Whether the conclusion depends on a narrow or poorly supported range. |
Diagnostics, Alternatives, and Limitations
Several checks help detect a misleading inclusion decision. First, verify that the output being compared is actually the one used for the decision; a mechanism can be unimportant for one output and important for another. Second, inspect predictions across all relevant intervention scenarios, not only a reference case. Third, ask whether the plausible range is evidence-based and wide enough to include relevant conditions. Finally, check the decision itself: a small numerical change can matter near a threshold, while a larger change may leave the chosen action unchanged.
The main alternatives are not simply “detailed” and “absent.” An explicit state preserves the mechanism’s separate behavior; a simplified representation preserves selected consequences without modeling all its detail; exclusion removes it from the model boundary. A fixed contribution can be appropriate when the mechanism matters at baseline but does not need to vary for the use. The choice should reflect the output and intervention contrasts that must be represented.
A negative screen is not proof of biological irrelevance. The mechanism may matter for an unmodeled output, a different population, or an intervention outside the specified comparison. A positive screen is not proof that the mechanism is identified or that its model form is correct. If evidence cannot establish a credible plausible range or whether the output effect is material, record the screen as inconclusive rather than treating missing information as evidence for exclusion.
Common pitfall: “The mechanism has little effect at the reference condition, so it can be excluded.” This reasoning ignores intervention responsiveness and plausible variation. A mechanism can have little baseline effect yet change substantially under an intervention and move the target output across a decision threshold. The correction is to compare plausible mechanism behavior across the relevant scenarios, then assess output and decision impact—not just its reference-condition contribution.
Check Your Understanding
Answer each question using the stated context and tolerances.
- A mechanism changes a predicted output by \(1\) percentage point, but the decision threshold is \(0.5\) percentage points away. Is the change necessarily immaterial? Explain.
- A candidate mechanism responds strongly to two interventions, but its modeled effect on the target output is negligible throughout the plausible range. Which test is positive, and what else is needed before including it explicitly?
- For a hypothetical mechanism, plausible values are \(m=0.1\) to \(m=0.5\), and \(y=10+5m\%\). Calculate the output range and its width.
- Why is a poorly known plausible range not, by itself, evidence that a mechanism can be excluded?
Solutions. (1) No. A \(1\)-percentage-point change could cross the threshold, so it may change the decision even if it is small relative to other output differences. (2) The intervention-responsiveness test is positive, but output impact appears negligible. Explicit inclusion is not justified by responsiveness alone; assess whether the output tolerance and decision remain unchanged, and whether a simpler representation is needed. (3) At \(m=0.1\), \(y=10+5(0.1)=10.5\%\). At \(m=0.5\), \(y=10+5(0.5)=12.5\%\). The range is \(10.5\%\) to \(12.5\%\), with width \(12.5\%-10.5\%=2.0\) percentage points. (4) The screen depends on plausible behavior. If that range is not established, the mechanism’s potential output impact is unresolved; the appropriate conclusion is inconclusive, not automatically negligible.