Why Parameter Bounds Need a Biological Argument
A parameter range can look reasonable in software and still be indefensible for the patients a model is meant to represent. A bound that is too narrow may exclude people who were observed in the target population; a bound that is too broad may permit values with no credible biological interpretation. In virtual-patient generation, bounds define which values are eligible for consideration. They do not, by themselves, describe how common each value is or guarantee that a selected parameter vector is physiologically coherent.
The starting point is the parameter’s meaning: what quantity does it represent, in which units, and for which population and conditions? Those details determine whether a reference interval is relevant, whether a value can be compared across sources, and whether a proposed extreme is meaningful. The earlier material on virtual patients treats a patient as a mechanistic parameter vector; here the focus is on defining defensible ranges for its components.
It is useful to distinguish three kinds of statements. A hard constraint follows from the definition of the quantity or a well-established physical or biological limit. A plausible range describes values considered credible for a specified target population and use, while allowing for uncertainty and uncommon cases. A reference interval summarizes measurements in a defined reference population, often healthy people under specified conditions. A reference interval is not generally a hard boundary: a substantial fraction of healthy individuals can fall outside a central 95% reference interval, and patients with disease may reasonably have values beyond it.
Build Bounds From Meaning, Units, and Evidence
Before choosing numbers, write down the parameter’s definition and unit, the population to be represented, and the circumstances under which it is defined. For example, a concentration measured in plasma cannot be compared uncritically with a concentration in tissue, and a rate reported per day cannot be inserted as though it were per hour. If the parameter is an estimated or latent mechanistic quantity rather than a direct measurement, identify how it relates to observed measurements and what assumptions that relationship uses.
Evidence for a range may come from biological constraints, reference data, clinical measurements, published parameter estimates, or model-specific requirements. Use sources that match the intended population as closely as possible. Differences in age, disease status, treatment, assay, sampling time, or units can explain apparent discrepancies. When sources do not match exactly, record the mismatch and treat the range as uncertain rather than silently treating the values as interchangeable.
Observed extremes are valuable checks, but they are not automatically the true limits of a population. The smallest and largest recorded values depend on sample size and may also reflect measurement error, data-processing errors, or a population that differs from the intended one. Review extreme records in context before excluding them. If an extreme is credible and belongs to the target population, bounds that prohibit it need a specific scientific justification.
Worked Example: Setting a Body-Weight Range Without Excluding Observed Adults
Suppose a hypothetical QSP model represents adults in a defined clinical program. A hypothetical data extract contains 240 eligible adults with recorded body weights from 46 kg to 138 kg. The records have been checked for unit errors and implausible entries. Propose preliminary bounds for body weight.
The parameter is body weight in kilograms, so its admissible range must also be in kilograms. The observed range is \(46\ \text{kg}\) to \(138\ \text{kg}\). If the target population is the same adult program population and the observations are credible, choosing bounds of 50 kg to 120 kg would exclude observed patients. That choice would require a defensible reason to exclude those patients; convenience alone is not sufficient.
As a provisional range, suppose the modeling team selects \(40\ \text{kg} \leq \theta \leq 150\ \text{kg}\). This adds margins of \(46-40=6\ \text{kg}\) below and \(150-138=12\ \text{kg}\) above the observed extremes. These margins are modeling choices, not estimates of population tail probabilities. They are defensible only if the broader target population and available clinical knowledge support including those values.
The bounds contain every checked observation because \(40 \leq 46\) and \(138 \leq 150\). This is a useful initial audit, not proof that the range captures every plausible adult. If the intended population later includes a group not represented in the extract, the bounds must be reassessed against evidence for that group.
A simple probability calculation explains why observed minima and maxima should not be mistaken for population limits. If \(n\) independent observations come from a continuous population with cumulative distribution function \(F\), then for a proposed lower bound \(L\), the probability that every observation exceeds \(L\) is
This follows because one observation exceeds \(L\) with probability \(1-F(L)\), and independence makes the probability that all \(n\) observations exceed \(L\) the product of those probabilities. Thus a finite sample can miss values below its minimum, particularly when such values are uncommon. The calculation does not identify a correct bound by itself; it shows why a sample extreme is evidence about the population, not a guaranteed population endpoint.
Translate Biological Extremes Into Parameter Units
A range is easier to defend when it is connected to an interpretable biological quantity. A common example is a first-order rate constant \(k\), measured in inverse time, related to a half-life \(t_{1/2}\) by \(t_{1/2}=\ln(2)/k\). Because the relationship is decreasing, longer half-lives correspond to smaller rate constants. The conversion should be done explicitly, including the time unit.
Worked Example: Converting Half-Life Limits to Rate-Constant Bounds
Consider a hypothetical receptor-internalization rate constant \(k\) in \(\text{h}^{-1}\). Suppose a mechanistic review supports plausible internalization half-lives from 2 hours to 20 hours for the intended setting. Assume the first-order half-life relation applies and that the cited half-life range is relevant to this model quantity.
Using \(k=\ln(2)/t_{1/2}\), the slowest rate corresponds to the longest half-life:
The fastest rate corresponds to the shortest half-life:
The resulting provisional range is \(0.03466 \leq k \leq 0.3466\ \text{h}^{-1}\). Its order is important: the 20-hour half-life gives the lower rate bound, not the upper one. If a credible patient-specific estimate were \(k=0.025\ \text{h}^{-1}\), its implied half-life would be \(0.6931/0.025=27.72\ \text{h}\), outside the proposed range. The observation should be investigated, not discarded automatically: the half-life evidence may not apply to that patient or model, or the range may need revision.
A derivation like this makes the assumptions visible. The conversion assumes a first-order process with a single rate constant; it would not directly justify bounds for a multi-compartment or time-varying process. More generally, an interpretable transformation can help set limits, but its assumptions must match the parameter’s definition in the model.
Reference Ranges and Credible Outliers
A clinical reference interval can inform a parameter bound when the model parameter is directly comparable with the measured quantity and the reference population matches the intended use. Even then, a central 95% reference interval is a summary of a distribution, not a guarantee that values outside it are impossible. For disease biomarkers, treatment targets and healthy reference intervals answer different questions. The target population may intentionally include values beyond the healthy interval.
Worked Example: Auditing Bounds for a Baseline Biomarker
Suppose a hypothetical model uses baseline plasma biomarker concentration \(C\), measured in \(\text{pg}/\text{mL}\). A hypothetical healthy reference source reports a central 95% interval of 1.2 to 8.0 \(\text{pg}/\text{mL}\). The intended virtual population, however, is a clinical population that may include elevated biomarker levels. In a checked hypothetical sample of 180 eligible patients, concentrations range from 0.7 to 10.5 \(\text{pg}/\text{mL}\).
Using 1.2 and 8.0 \(\text{pg}/\text{mL}\) as hard bounds would exclude the observed values 0.7 and 10.5 \(\text{pg}/\text{mL}\). The healthy reference interval is not an appropriate hard support for this clinical population. Suppose clinical evidence relevant to the target setting supports a provisional range of 0.5 to 12 \(\text{pg}/\text{mL}\). This range includes the checked observations and leaves margins of \(0.7-0.5=0.2\ \text{pg}/\text{mL}\) and \(12-10.5=1.5\ \text{pg}/\text{mL}\).
The proposed limits are defensible only if the clinical evidence supports them and the measurements are comparable in specimen type, assay, and timing. The observed values alone do not establish that 0.5 and 12 are population limits. If the value 10.5 were instead traced to a unit-conversion error, it would not be evidence for widening the range; the record and its provenance would need correction.
Audit, Diagnose, and Compare Bound Choices
For each parameter, keep a short record of its definition, units, target population, evidence source, proposed lower and upper limits, and reasons for the selected margins. Then compare the bounds with available patient-level observations. Report how many credible observations fall below or above the proposed interval, and inspect those records. A single excluded record may be a data error, but it may also reveal a genuine subgroup or a mistaken assumption about the target population.
Several warning signs indicate that a range needs review:
- Unit inconsistency: a rate in \(\text{day}^{-1}\) is compared directly with a bound in \(\text{h}^{-1}\), or a concentration uses a different unit or specimen type.
- Unexplained exclusions: credible observations from the intended population lie outside the range, with no stated scientific or eligibility reason.
- False precision: bounds are reported with many digits despite uncertain evidence for the endpoints.
- Uninterpretable breadth: a range spans orders of magnitude without a biological explanation, or is so narrow that ordinary measurement and population variation would violate it.
- Population mismatch: a healthy reference interval is treated as a hard boundary for a patient population with a different disease or treatment context.
Three common approaches provide different kinds of evidence. Biological or mechanistic limits are strongest when a genuine constraint is known, but may be unavailable or may apply only under specific conditions. Reference and clinical data anchor the range to measured populations, but require careful checks of population, assay, and sampling context. Optimization bounds can stabilize parameter estimation or prevent numerical searches in unhelpful regions, but they are not automatically biological bounds. An optimizer’s convenient search interval should not be presented as evidence of what patients can have.
These approaches can be combined, but the resulting interval remains a modeling choice subject to evidence and revision. Bounds alone do not encode the frequency of values within the range, the relationships among patient traits, or whether a full parameter vector is physiologically feasible. Those are separate questions. In particular, checking each parameter individually does not establish that every combination of individually admissible values is biologically credible.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions below.
- A parameter is reported in \(\text{day}^{-1}\), but its proposed bounds are documented in \(\text{h}^{-1}\). What must be done before using the bounds?
- A healthy-population central 95% reference interval is proposed as a hard range for a disease population. State one reason this may be inappropriate.
- A hypothetical rate constant has a plausible half-life range of 4 to 16 hours. Calculate the corresponding rate-constant bounds in \(\text{h}^{-1}\).
- A credible observation from the intended population falls outside a proposed bound. Give two checks to make before revising the bound or excluding the observation.
Solutions. (1) Convert the units consistently and verify the parameter definition. Since \(1\ \text{day}=24\ \text{h}\), a rate per day and a rate per hour differ by a factor of 24; the conversion direction must be checked rather than assumed. (2) The disease population may include values outside the healthy interval, which summarizes a reference group and is not generally a hard biological limit. (3) The longest half-life gives the lower rate: \(k_{\min}=0.6931/16=0.04332\ \text{h}^{-1}\). The shortest gives the upper rate: \(k_{\max}=0.6931/4=0.1733\ \text{h}^{-1}\). (4) Check the record’s units, assay, timing, and data quality; also check whether the patient meets the target-population definition and whether the evidence used to set the bounds applies to that population.