From a Decision to a Model Target
A QSP model is useful for a development decision only if its predictions can be connected to the quantity on which that decision depends. “Does intervention A improve disease control?” is not yet a model target: improvement could mean a change in a biomarker, a clinical endpoint, or a summary over time, and “compared with” could refer to several possible alternatives. A model may simulate many biological quantities, but only some are needed to answer the stated question.
The first task is therefore to define the requested prediction before deciding what to represent. Specify what quantity is being compared, which scenarios are being contrasted, and when or over what period the quantity is evaluated. Then identify the model outputs needed to calculate it and the biological processes that must be represented for those outputs to be credible under the comparison.
For a scenario \(a\), let \(Y_a(t)\) denote the output of interest at time \(t\). A simple decision target might be the difference between two scenarios at a specified time \(t^*\):
The subtraction is meaningful only when \(Y_a\) and \(Y_b\) represent the same outcome, on the same scale and with the same units. If the decision instead concerns a relative change, an accumulated amount, or whether a threshold is met, that quantity must be stated explicitly. Different summaries of the same simulated trajectory can lead to different decisions.
Define the Target Before the Output List
A useful target specification is a short verbal statement that can be translated into a calculation. It should name the endpoint or model-derived quantity, the alternative scenarios, and the evaluation time or interval. It should also make clear whether the quantity is an absolute difference, a relative difference, a threshold comparison, or another specified summary.
The output is not necessarily a directly measured clinical endpoint. It may be a model state used as a surrogate, or a calculated endpoint derived from multiple states. In either case, the mapping from model quantities to the requested target must be stated. For example, a model may track a circulating biomarker as a state, while the decision target is the percentage reduction from its baseline value at day 28.
Worked Example: Biomarker Difference at a Specified Time
Hypothetical decision. A development team wants to compare candidate A with standard treatment on a disease biomarker at day 28. The model predicts biomarker concentrations of 42 ng/mL under candidate A and 50 ng/mL under standard treatment at that time. Assume both predictions use the same biomarker definition and baseline conditions.
Target and calculation. Define the target as candidate A minus standard treatment at day 28:
The predicted concentration is 8 ng/mL lower under candidate A at day 28. This is an absolute difference, not an 8% reduction. If the decision were instead framed as a relative reduction from standard treatment, the calculation would be \(100(50-42)/50=16\%\). Those are distinct target quantities and should not be interchanged.
Output boundary. At minimum, the model must provide the biomarker output at day 28 in both scenarios. It must also represent the processes needed for the model to predict how the two scenarios produce that biomarker value. A model that predicts only a related upstream signal, without a defined mapping to biomarker concentration, does not directly answer this target.
Map the Target to Outputs and Processes
A model often represents internal quantities that are not themselves the decision endpoint. For example, a system of differential equations may describe biological states \(x(t)\), with scenario inputs \(u_a(t)\) and parameters \(\theta\). A reported output is then calculated from those quantities, such as \(Y_a(t)=h(x_a(t),u_a(t),\theta)\). The target is a further specification of which output values or summaries are compared.
This notation separates three questions. The state equation describes the modeled system. The output function says how model quantities become a reported quantity. The target definition says which reported quantities are used to make the comparison. Confusing these levels can produce a model that is biologically detailed but not connected to the decision endpoint.
Once the target is explicit, ask which outputs are needed to calculate it, and which represented processes are needed to predict those outputs under each scenario. This is not a request to include every process that might be biologically relevant. It is a requirement that the model’s represented system and output mapping support the stated prediction. If a required quantity is absent, the gap may be addressed by defining a justified mapping or by revising the target; it should not be hidden by relabeling a different output as the requested endpoint.
Name the endpoint or summary, including its units or scale where applicable.
Identify the scenarios and the time point or interval at which the target is calculated.
Write the calculation that converts model-predicted outputs into the target.
Ask whether the model represents the quantities and processes needed to predict those outputs under the specified comparison.
Why the Comparison Definition Matters
A comparison is more than a pair of scenario labels. Its interpretation depends on the target scale, denominator, and evaluation time. For instance, a difference of 8 ng/mL and a relative reduction of 16% describe the same two predictions but answer different questions. A threshold target also needs an explicit rule: whether the output must be strictly below a threshold, at or below it, or within a specified range.
The comparison must also be internally consistent. Both scenarios should use the same output definition and evaluation convention. If one value is measured at day 28 and the other is averaged over days 21–28, the subtraction is not a well-defined contrast unless that mismatch is intentionally part of the target. Likewise, an endpoint reported in one unit must be converted consistently before comparing it with another.
Worked Example: Comparing Trajectory Summaries
Hypothetical decision. Compare two dosing schedules using the total modeled biomarker concentration over days 0–14. The values below are predicted at equally spaced times, in ng/mL. Define the target as the trapezoidal area under the concentration-time curve over the interval; its units are ng·day/mL.
| Time (days) | Schedule A (ng/mL) | Schedule B (ng/mL) |
|---|---|---|
| 0 | 10 | 10 |
| 7 | 8 | 9 |
| 14 | 6 | 8 |
For equally spaced observations with interval width 7 days, the trapezoidal area is the sum of the average endpoint values multiplied by the interval width. Thus:
The target, Schedule A minus Schedule B, is \(112-126=-14\ \mathrm{ng\cdot day/mL}\). Under this specified summary, Schedule A has a lower predicted total biomarker concentration over days 0–14 by 14 ng·day/mL. This result does not establish that Schedule A is preferable for every other outcome or at every other time.
The required outputs are the concentration trajectories at the times used in the calculation, or a model output that supports the same area calculation. The relevant processes must support prediction of those trajectories under both schedules. A single day-14 output would not by itself determine this interval summary.
Justification: A Contrast Requires Comparable Outputs
The arithmetic of a contrast is simple, but its validity depends on what is being subtracted. Suppose the reported outputs are \(Y_a(t^*)\) and \(Y_b(t^*)\). If they measure the same quantity in the same units, their difference has a defined interpretation in those units. If they instead measure different quantities, or use different transformations, subtraction does not produce the stated target.
This can be seen directly from a simple unit conversion. If one scenario’s concentration is reported in ng/mL and the other in µg/mL, the numerical values cannot be subtracted as written. After converting \(1\ \mathrm{\mu g/mL}=1000\ \mathrm{ng/mL}\), both values share a unit and the contrast can be calculated. More generally, the output mapping must be held consistent across scenarios so the difference reflects the scenario comparison rather than a change in definition.
A relative reduction has an additional condition: its denominator must be defined and nonzero. For a reference value \(Y_b(t^*)\), the reduction from that reference is:
This formula is undefined if \(Y_b(t^*)=0\), and its interpretation depends on the choice of reference scenario. Therefore, an absolute difference and a relative reduction are not interchangeable summaries. The target definition determines which calculation is appropriate; the model does not choose it automatically.
Worked Example: A Threshold-Based Decision Target
Hypothetical decision. A candidate is considered to meet a biomarker criterion if its predicted day-28 concentration is at most 45 ng/mL. Candidate A is predicted at 42 ng/mL and comparator B at 47 ng/mL. Define the target as whether each scenario meets the criterion, rather than as a difference alone.
For candidate A, \(42\ \mathrm{ng/mL}\leq45\ \mathrm{ng/mL}\), so the predicted output meets the stated criterion. For comparator B, \(47\ \mathrm{ng/mL}>45\ \mathrm{ng/mL}\), so it does not meet the criterion. The difference is \(42-47=-5\ \mathrm{ng/mL}\), but the decision rule is the threshold comparison, not merely whether the difference is negative.
The model must provide the same biomarker output at day 28 under both scenarios, with units compatible with the threshold. If the model instead reports a different biomarker, a threshold in concentration units cannot be applied without a justified mapping. Meeting this modeled criterion is not, by itself, proof of clinical benefit; it answers only the specified model-based target.
Diagnostics, Alternatives, and Limitations
Several checks help reveal when the output boundary is not aligned with the decision. A mismatch in units, time points, endpoint definitions, or baseline conventions is a warning that the calculated contrast may not mean what its label suggests. Another warning is that the model reports a convenient proxy while the decision depends on a different endpoint and no output mapping has been specified. In these cases, the calculation may be precise but still fail to answer the decision question.
The appropriate level of output detail depends on the decision. A single value at a specified time may be sufficient for a point-specific comparison. A trajectory summary is needed when the target concerns accumulated exposure or a period-wide response. A vector of outputs may be necessary when the decision explicitly requires more than one outcome, such as meeting an efficacy threshold while remaining below a safety threshold. Combining distinct outcomes into one score requires a separately specified rule and should not be done implicitly.
An alternative is to use an empirical model that predicts the target output directly rather than representing the underlying biological processes in detail. That can be adequate when the decision only requires prediction within a supported setting. A mechanistic QSP representation becomes relevant when the decision requires predictions across interventions or conditions that depend on biological relationships. Neither approach repairs an undefined target: both require a clear outcome, comparison, and evaluation.
The target definition also does not guarantee that the model can predict it reliably. A model may include a suitable output but have inadequate evidence for a process that controls that output under the comparison. Model predictions remain conditional on model structure, assumptions, and parameterization. The purpose of specifying the boundary is to expose what must be predicted, not to certify that the available model already does so.
Common Pitfall
A related mistake is to report a favorable value without specifying its comparator or time. “The predicted biomarker is lower” is incomplete unless it says lower than what, at what time or over which interval, and on what scale. A fully specified target makes these choices visible and allows reviewers to check whether the model output truly supports the decision.
Check Your Understanding
For each question, first identify the target definition that is needed. Solutions follow the questions.
- A model predicts 30 ng/mL for treatment and 36 ng/mL for comparator at day 21. What is the treatment-minus-comparator difference, and what does its sign mean?
- Why is a day-14 concentration alone insufficient to calculate a total concentration-time area from day 0 to day 14?
- A threshold is defined as strictly less than 20 units. Does a prediction of exactly 20 units meet it?
- One scenario’s output is in mg/L and another’s is in µg/L. What must be done before calculating their difference?
- A team wants a relative reduction from a reference value of zero. Why is this target not defined by the usual relative-reduction formula?
Solutions. (1) The difference is \(30-36=-6\ \mathrm{ng/mL}\). The negative sign means the treatment prediction is 6 ng/mL lower than the comparator at day 21. (2) A single endpoint value does not determine the intervening trajectory, and the area depends on values across the interval or an equivalent model-derived summary. (3) No. “Strictly less than 20” excludes equality, so a value of 20 does not meet the criterion. (4) Convert both values to the same unit, using \(1\ \mathrm{mg/L}=1000\ \mathrm{\mu g/L}\), then subtract in that common unit. (5) The reference value is the denominator, and division by zero is undefined; specify a different target, such as an absolute difference, if appropriate.