Why Parameter Provenance Matters
A QSP parameter is more than a number in a model file. It represents a quantity with a definition, a measurement scale, evidence behind it, and a particular role in the equations. If a simulation result is challenged, the team needs to be able to trace that number backward: which study informed it, what that study measured, what calculations were applied, and why the resulting value was used in this model. That chain is parameter provenance.
Traceability is especially important when evidence is incomplete or inconsistent. Two studies may report different values because they used different cell types, assays, exposure conditions, or parameter definitions. Recording both findings and the rationale for a model choice preserves that distinction; silently replacing them with a single number hides it. The aim is not to make all evidence agree, but to make the route from evidence to model use inspectable and reproducible.
What a Useful Parameter Record Contains
A record should be sufficiently specific that a second analyst can reconstruct the value without relying on memory or an undocumented spreadsheet. The evidence description should identify the source and the particular result used, not merely say “literature.” For a study, useful details include the experimental system, assay, conditions, reported statistic, and the location of the result in the source. The record should distinguish what the source reported from what the model uses.
- Definition and model role: what biological or mathematical quantity the parameter represents and where it enters the model.
- Selected value and unit: the value used in the stated model context, with enough unit information to interpret its scale.
- Uncertainty: the range or distribution used, what it represents, and whether it was reported by a source or constructed for modeling.
- Evidence: the source result, experimental context, and relevant limitations.
- Transformation: formulas, constants, conversions, and assumptions used to obtain the model value from the reported result.
- Decision and status: why this evidence was selected, how alternatives were handled, and whether the choice is provisional or unresolved.
A parameter table is often the most efficient summary, but a table cell should not become a substitute for an auditable calculation. If a value is derived, the record should point to the calculation or state it clearly. If different values are used for different model contexts, preserve that distinction rather than implying that one value is universally applicable.
| Field | Example entry | Traceability question answered |
|---|---|---|
| Parameter definition and model use | First-order loss rate for the modeled protein; used in the protein loss term | What does the value mean in these equations? |
| Value and unit | 0.03851 h-1 | What number is implemented, and on what scale? |
| Uncertainty | 0.03151–0.04951 h-1, transformed from a reported half-life range | What variation does the range represent? |
| Source and context | Hypothetical Study A; cultured cells under specified baseline conditions | What evidence supports the entry? |
| Transformation and decision | Rate calculated as ln(2) divided by half-life; Study A selected for context match | How was the value obtained, and why was it chosen? |
Transforming Evidence Without Losing Its Meaning
A source may report a quantity different from the parameter required by the model. A transformation must then be stated, and its assumptions checked. For a first-order loss process, the amount remaining follows exponential decay, \(C(t)=C_0e^{-kt}\), where \(k\) is the loss rate. The half-life \(t_{1/2}\) is the time at which \(C(t_{1/2})=C_0/2\). Substitution gives \(1/2=e^{-kt_{1/2}}\); taking natural logarithms and rearranging yields \(k=\ln(2)/t_{1/2}\). This relationship applies when the loss process is adequately represented as first order over the relevant conditions.
Because this transformation decreases as half-life increases, the endpoints of an interval must be reversed. The larger half-life produces the smaller rate. This is not a cosmetic detail: retaining the original endpoint order would report an invalid or misleading rate range.
Worked Example: Converting a Half-Life Range to a Rate Range
Suppose a hypothetical study reports a protein half-life of 18 h and gives a range from 14 to 22 h. The model represents loss with a first-order rate \(k\). Assume the study’s half-life definition and conditions are compatible with first-order loss in the model’s intended context. Calculate the point value and transform the range.
Point value. Use the reported half-life in \(k=\ln(2)/t_{1/2}\):
Range. The half-life endpoints are 14 h and 22 h. Since \(k\) decreases with \(t_{1/2}\), the lower rate comes from 22 h and the upper rate comes from 14 h:
The provenance entry should state that the rate and range were derived from half-life using the first-order relationship. It should also state what the source’s range means. If it is a confidence interval, the transformed interval is a confidence interval for the corresponding rate under that transformation; if it is an observed range or a plausible scenario range, it should retain that description instead. A transformation does not change the evidential meaning of the input range.
Working Through Conflicting Evidence
Before resolving a conflict, determine whether the studies estimate the same quantity in comparable conditions. A measured half-life in one cell type and a fitted loss rate in another are not automatically interchangeable, even if both can be expressed in inverse hours. Compare the biological system, assay, experimental conditions, endpoint definition, analysis, and uncertainty. Differences may reflect real context dependence, measurement procedures, or estimation limitations; a provenance record should not pretend to know which explanation is correct when the evidence does not establish it.
A defensible decision may select the study that best matches the intended model context, retain another study as an alternative scenario, or leave the value unresolved pending additional evidence. A weighted average is not a neutral default. It requires a justified basis for treating the estimates as measurements of a common quantity and a method that accounts for their uncertainty and comparability. If those conditions are not established, averaging can produce a precise-looking value with no clear interpretation.
Worked Example: Recording Two Conflicting Loss-Rate Estimates
Consider two hypothetical studies of a protein loss process. Study A, performed in a cell system selected for the model’s baseline context, reports a half-life of 18 h with a 95% confidence interval from 15 to 22 h. Study B, using a different cell system, directly reports a loss rate of \(0.060\ \mathrm{h}^{-1}\) with a 95% confidence interval from \(0.045\) to \(0.078\ \mathrm{h}^{-1}\). The model requires a first-order loss rate.
For Study A, transform the point estimate and interval using the first-order relationship:
The interval endpoints are reversed because the conversion is decreasing. Study B’s rate and interval are already on the model parameter scale: \(0.060\ \mathrm{h}^{-1}\) and \([0.045,\ 0.078]\ \mathrm{h}^{-1}\). The two point estimates differ. The reported intervals overlap only from \(0.045\) to approximately \(0.04621\ \mathrm{h}^{-1}\), a narrow interval of overlap. This observation alone does not establish that the estimates are equivalent, nor does it explain why they differ.
For a model explicitly representing Study A’s baseline cell context, a traceable decision could use \(0.03851\ \mathrm{h}^{-1}\) as the baseline value and record Study A’s transformed 95% confidence interval as its associated evidence-based uncertainty. Study B should remain in the evidence table, with its distinct system and reported result, and could be used as an alternative-context scenario if scientifically appropriate. The decision is based on context matching, not on a claim that Study A is universally more accurate. If the model is intended to represent both systems, the conflict signals that a single context-free value may be inadequate; the record should flag that limitation rather than conceal it with an unqualified average.
Derived Parameters and Explicit Uncertainty Rules
A derived parameter depends on one or more upstream quantities. Its provenance must include both the inputs and the rule used to combine them. Otherwise, an analyst can reproduce the final number only by guessing which inputs were selected. For simple deterministic transformations, show the formula and the substituted values. For uncertainty, state whether the reported range is a confidence interval, a plausible range, or a deliberately broad scenario envelope. These have different interpretations.
Worked Example: Tracing a Derived Binding Rate
Suppose a hypothetical source reports a dissociation constant \(K_D=8\ \mathrm{nM}\), and a second hypothetical source provides an association rate \(k_{\mathrm{on}}=0.025\ \mathrm{nM}^{-1}\mathrm{h}^{-1}\). The model uses a dissociation rate defined by \(k_{\mathrm{off}}=K_Dk_{\mathrm{on}}\). Assume the quantities refer to compatible binding definitions and conditions, and that the model uses this relationship.
Substitution gives:
The nanomolar units cancel, leaving inverse hours. Suppose, for illustration, the sources provide ranges \(K_D=6\)–\(10\ \mathrm{nM}\) and \(k_{\mathrm{on}}=0.020\)–\(0.030\ \mathrm{nM}^{-1}\mathrm{h}^{-1}\). If these are used as independent endpoint ranges, multiplying the smallest endpoints and the largest endpoints gives an outer range:
This is an endpoint scenario range, not automatically a 95% confidence interval for \(k_{\mathrm{off}}\). If the input ranges are statistically dependent, or represent confidence intervals, this endpoint calculation does not by itself provide a calibrated uncertainty interval for the product. The record should preserve the two sources, their ranges, the assumed relationship, and the specific interpretation of the resulting range.
Diagnostics, Alternatives, and Common Pitfalls
A provenance chain is useful only if it can be checked. Reviewers should be able to start from the implemented parameter and trace backward to the reported result, then forward through the stated transformation to the model use. Discontinuities in that chain are diagnostic warnings. Examples include a table value that does not match the source, an unexplained unit conversion, an uncertainty range whose endpoints violate the transformation’s direction, or a chosen study that measures a different quantity from the model parameter.
- Check arithmetic and direction: recompute each transformation, verify endpoint order, and retain enough digits internally to avoid rounding discrepancies.
- Check context: flag differences in biological system, assay, conditions, or endpoint that could limit transfer to the intended model use.
- Check uncertainty meaning: do not label a plausible scenario range as a confidence interval, or treat a confidence interval as the full range of biological variability.
- Check completeness: every selected value should have a source or an explicit statement that it is an assumption or modeler-defined scenario.
The main alternatives for handling competing evidence are selecting a context-matched source, retaining multiple values as distinct scenarios, or combining estimates when a common target quantity and a justified combination method are available. Selection is often clearest when one source closely matches the modeled context. Scenarios preserve meaningful differences without implying that a single value resolves them. Combining estimates can be useful, but only after comparability and the interpretation of the combined result have been established. The correct choice depends on the scientific question; provenance makes that choice visible.
Common mistake: “Both studies estimate a rate, so average their point estimates and report the result as the best-supported value.” This reasoning ignores context, differences in how the parameter was measured, and the uncertainty of each estimate. The correction is to document each result separately, determine whether the studies address the same quantity under sufficiently comparable conditions, and justify selection, scenario use, or combination. A transparent unresolved conflict is more informative than an unjustified single number.
Provenance cannot guarantee that a study is unbiased, that a transformation is biologically adequate, or that a parameter is identifiable from the available evidence. Nor does a well-documented value prove that the model is valid. It does make assumptions and decisions inspectable, supports focused review, and helps teams update a parameter when new evidence appears. Earlier material on calibration and predictive validation concerns model performance against data; parameter traceability complements those assessments by clarifying what evidence entered the model and how.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions.
- A study reports a first-order half-life interval of 10–20 h. In what order should the corresponding rate interval’s endpoints be calculated?
- One study reports a rate in a cell line matching the model context; another reports a different rate in a different species. Why is an unqualified average not automatically justified?
- A table gives a derived value but does not identify its inputs or formula. What essential provenance information is missing?
- A source reports a 95% confidence interval, while a modeler defines a wider range to explore plausible scenarios. Should both ranges be labeled as confidence intervals?
Solutions. (1) Calculate the lower rate from the longer half-life, 20 h, and the upper rate from the shorter half-life, 10 h, because \(k=\ln(2)/t_{1/2}\) decreases as half-life increases. (2) The studies may represent different contexts and may not estimate one interchangeable quantity; first assess comparability and document the context before selecting, scenario-testing, or combining them. (3) Record the source values, their units and uncertainty descriptions, the transformation formula, assumptions, and how the derived value is used. (4) No. Identify the source interval as a 95% confidence interval and label the modeler-defined range as a scenario or plausible range, with its rationale.