Building a DDI Evidence and Parameter Map

Why Build an Evidence and Parameter Map?

A drug–drug interaction (DDI) model is only as reviewable as the evidence behind its inputs. A value copied into a model file without its source, experimental context, or intended role can look precise while being difficult to interpret or update. For a CYP-mediated interaction, the model may need information about the victim substrate, the perpetrator, and the evidence used to connect them. These inputs can come from different kinds of evidence: in vitro enzyme assays, clinical interaction studies, and product information.

An evidence and parameter map is a structured record linking each model parameter to the evidence that supports it. It is not a new parameter-estimation method, and it does not make unlike evidence interchangeable. Its purpose is to make clear what a value represents, how it was obtained or reported, what uncertainty remains, and where the value is intended to be used in the model.

Definition: A DDI evidence and parameter map is a parameter-level register that links each DDI model input to its source evidence, relevant context, uncertainty, any transformation or assumption used to obtain the input, and its intended model use. A parameter-level record refers to a specific quantity, not merely to a drug or study in general.

The earlier material on assigning victim and perpetrator roles distinguishes the drug whose exposure is affected from the drug producing the enzyme effect. The map builds on that distinction: it should identify which drug and model role a parameter belongs to. For example, a CYP3A fraction metabolized for a victim and a perpetrator’s in vitro inhibition potency are different quantities, even if both are relevant to the same interaction.

What Each Record Should Preserve

A useful map makes it possible to answer, without searching through undocumented model code: What is this parameter? Which evidence supports it? Under what conditions was that evidence produced? How uncertain is its application to this model? What calculation or assumption connects the evidence to the input? A compact table can include the fields below.

Field	What to record	Why it matters
Parameter identity	Name, symbol or model label, drug, enzyme, and victim or perpetrator role	Prevents confusion between quantities with similar names or different roles
Value and units	Reported or model-entered value, units, and whether it is a point estimate, range, or category	Makes the input interpretable and supports unit checking
Source and provenance	Study or product-information source, version or date, and the specific table, result, or statement	Allows reviewers to locate and verify the evidence
Evidence context	Assay system and conditions, clinical regimen and population, or product-information context, as applicable	Shows what the evidence directly describes and where it may not transfer
Uncertainty and limitations	Reported uncertainty, plausible range if justified, unresolved variability, and relevant limitations	Distinguishes a well-supported input from an assumption or uncertain translation
Model use	Model component, intended simulation or check, and whether the value is used directly or after a stated transformation	Links the evidence to the calculation it is meant to inform
Status and alternatives	Selected, provisional, contextual, or rejected; competing evidence and the reason for the decision	Keeps disagreement visible and makes revisions auditable

The record should preserve the evidence as reported, separately from the model input if the two differ. For example, an assay may report an IC₅₀ under specified conditions, while a model uses a different inhibition parameter. The map should retain the reported IC₅₀, document the assumptions and method used for any conversion, and identify the model input separately. It should not relabel the assay result as though it directly measured the model input.

Uncertainty also needs a meaningful description. A confidence interval reported by a study, a range across laboratories, a plausible range chosen for sensitivity analysis, and a qualitative product statement are not the same kind of uncertainty. Label each according to what it represents. If evidence does not supply a numerical uncertainty, record that fact rather than inventing a range. A map can still mark the input as uncertain and explain why.

From Evidence to Model Use: A Justification

The reason to keep provenance and model use together is that uncertainty in an input can affect predictions, but the extent depends on how the model uses that input. Suppose a model output \(y\) is a differentiable function \(g\) of parameter values collected in \(\boldsymbol{\theta}\). For a small change \(\Delta\boldsymbol{\theta}\), a first-order Taylor expansion gives:

\Delta y = g(\boldsymbol{\theta}+\Delta\boldsymbol{\theta})-g(\boldsymbol{\theta}) \approx \nabla g(\boldsymbol{\theta})^\mathsf{T}\Delta\boldsymbol{\theta}

The approximation follows by expanding \(g\) around \(\boldsymbol{\theta}\) and omitting terms of second and higher order in the parameter changes. It is most informative for small changes where the model is locally smooth. It does not quantify uncertainty by itself: the parameter changes and the model’s response must still be assessed. But it explains why recording uncertainty without the parameter’s intended model use is incomplete. The same uncertain value may have little effect in one model context and a substantial effect in another. For large changes or nonlinear behavior, this local approximation may not be adequate.

Worked Example: Recording a Victim Substrate Parameter

Suppose a hypothetical in vitro metabolism study reports that the victim drug’s fraction metabolized by CYP3A, \(f_m\), is 0.65, with a reported range of 0.45 to 0.80. The study used a specified human liver microsome system. The modeling team intends to use this quantity as a CYP3A-related victim input in a DDI model.

Parameter identity. Record “victim drug A, CYP3A fraction metabolized, \(f_m=0.65\)” and identify drug A as the victim. This prevents the value from being mistaken for a perpetrator property or for a fraction associated with another enzyme.

Source and context. Record the hypothetical study identifier, assay system, relevant experimental conditions, and the reported result. The range is copied as reported; it is not described as a confidence interval unless the study defines it that way.

Uncertainty and status. Record 0.45–0.80 as the study-reported range, explain its stated basis, and mark the parameter provisional if the assay’s relevance to the intended model is not fully established. Do not replace the range with a narrower one merely because the central value is convenient.

Model use. State that the value is intended as a victim CYP3A input and specify the model component that will use it. The map does not establish that the reported value is the uniquely correct input for every model; it makes the evidence and intended application inspectable.

Keep In Vitro, Clinical, and Label Evidence Distinct

Different source types answer different questions. In vitro experiments may provide assay-specific observations about enzyme-mediated activity. Clinical DDI studies describe exposure changes under particular doses, schedules, populations, and conditions. Product information may summarize clinical findings, provide regulatory interpretation, or state a precaution without reporting a parameter estimate. These sources can complement one another, but they should not be collapsed into a single number unless a justified method connects them.

For each source, preserve enough context to interpret its scope. For an in vitro result, that commonly includes the biological system, substrate or probe, concentration range, incubation conditions, and reported endpoint. For a clinical result, record the victim and perpetrator regimens, study population, design, endpoint, and reported estimate and uncertainty. For product information, preserve the document version and exact relevant statement or reported result. These fields help distinguish evidence about enzyme activity from evidence about a clinical outcome.

Worked Example: Mapping a Perpetrator Assay Result

In a hypothetical enzyme assay, perpetrator drug B has a reported CYP3A IC₅₀ of 2.4 µmol/L. The assay report describes the system and conditions but does not provide a conversion to \(K_i\). The model team wants to use the evidence when developing a reversible-inhibition component.

What to enter. Record the reported quantity as “CYP3A IC₅₀, 2.4 µmol/L,” with the assay context and source. Do not enter 2.4 µmol/L in a field labelled \(K_i\): the reported endpoint and the intended model parameter are not automatically the same quantity.

What remains uncertain. Note that the report provides no \(K_i\) conversion and identify any assay limitations or missing conditions. If a later method is used to derive a model input, document that method, its assumptions, and the derived value in a separate field or linked record.

Intended use and status. Mark the result as evidence relevant to the perpetrator’s reversible CYP3A effect, rather than as a finalized model parameter. Its status can be “contextual” or “provisional” until the model-specific parameterization is justified. The reported value remains traceable even if another source later supports a different input.

Worked Example: Reconciling Clinical and Product-Information Evidence

Suppose a hypothetical clinical study reports a victim-drug AUC ratio of 3.1 during coadministration with perpetrator drug C, with a 90% confidence interval of 2.6 to 3.7. A hypothetical product-information document describes drug C as a strong CYP3A inhibitor but gives no assay potency value.

Separate the records. The clinical record contains the AUC ratio, its interval, the tested regimens, and the study context. The product-information record contains the qualitative classification, the document version, and the statement’s context. These are distinct evidence items; the qualitative classification is not another numerical estimate of 3.1.

Assign suitable uses. The clinical result can be recorded as a benchmark for checking a model prediction under comparable conditions. The product statement can be recorded as supporting context for the perpetrator’s classification. Neither source alone supplies a numerical in vitro potency parameter.

Interpret the interval correctly. Under the study’s statistical procedure, the 90% confidence interval describes uncertainty in the estimated clinical AUC ratio; it is not a 90% probability statement about the fixed true ratio. The map should retain the interval and its confidence level rather than relabeling it as a plausible parameter range for a different model input.

How to Resolve Disagreement Without Hiding It

Sources may differ because they measure different endpoints, use different experimental systems, or describe different clinical conditions. A numerical value from one source should not be averaged with a qualitative classification from another. Even two numerical values should not be pooled or averaged simply because they refer to the same drug or enzyme. First establish whether they estimate the same quantity under sufficiently comparable conditions, and whether a defensible method supports combining them.

If the evidence does not support a single selected value, keep the competing records and explain the current decision. A map can identify one input as selected for a stated model purpose, another as a sensitivity-analysis alternative, and a third as contextual evidence. If a value is not used, record why: for example, the endpoint is not the required parameter, the source context is insufficiently comparable, or the reported result is qualitative. This is more informative than silently deleting inconvenient evidence.

Identify the quantity.
Confirm the parameter, drug, enzyme, role, units, and endpoint before comparing sources.

Preserve each source as reported.
Record source context, reported value or statement, and uncertainty without prematurely converting or combining them.

State the connection to the model.
Document any transformation, assumption, or reason a source is being used as a benchmark rather than as a parameter estimate.

Assign a reviewable status.
Mark the input as selected, provisional, contextual, or not used, and explain the decision and relevant alternatives.

Diagnostics, Pitfalls, and Limitations

A map is failing as a communication tool if a reviewer cannot trace a model input back to a particular source, reconstruct how the input was obtained, or determine what the input is intended to represent. Check for missing units, ambiguous parameter labels, an absent source version, an unrecorded conversion, and uncertainty values whose meaning is unclear. Also check that the victim or perpetrator role and enzyme are stated consistently across the record and the model.

Wrong reasoning: “The clinical study shows a 3.1-fold AUC ratio, so 3.1 is the inhibitor potency parameter.” Correction: An AUC ratio is a clinical exposure outcome under a specified regimen, not automatically an in vitro potency value. Record it as a clinical result or benchmark unless a justified method maps it to a model parameter.
Wrong reasoning: “The product information calls the drug a strong inhibitor, so its numerical potency is known.” Correction: A qualitative classification is useful evidence, but it does not supply a specific numerical assay parameter.
Wrong reasoning: “Two sources disagree, so their values should be averaged.” Correction: First determine whether they concern the same quantity and comparable conditions. Preserve distinct evidence when they do not.
Wrong reasoning: “No reported uncertainty means the parameter is certain.” Correction: It means that the source did not provide a particular uncertainty description. Record the gap and any relevant limitations explicitly.

The map cannot eliminate uncertainty, establish that a source is applicable to every population or regimen, or prove that the model structure is adequate. It also cannot replace the method used to derive a parameter from experimental observations. Its value is traceability: it makes evidence selection and model use visible so that assumptions can be reviewed, tested, or updated.

Key takeaway: Keep the reported evidence, the model parameter, any transformation between them, the uncertainty, and the intended model use linked but distinct. A clear map does not resolve every scientific judgment; it makes those judgments explicit and auditable.

Check Your Understanding

Answer each question, then compare your reasoning with the solutions.

A study reports an assay-specific IC₅₀. What should the map preserve if the model requires a different inhibition parameter?
A product document calls a perpetrator a strong inhibitor but gives no numeric potency. What can and cannot be entered from this evidence?
A clinical DDI study reports an AUC ratio with a confidence interval. Why should this not automatically be recorded as uncertainty in an in vitro parameter?
Two sources give different values for a parameter. What should be checked before selecting or combining them?
Name two diagnostics that suggest a DDI evidence and parameter map is incomplete.

Solutions. (1) Preserve the reported IC₅₀, assay context, and source; separately document any justified conversion to the model-required parameter and its assumptions. (2) Record the qualitative classification and document context; do not invent a numeric potency value. (3) The interval describes uncertainty in a clinical exposure estimate under the study conditions, not automatically uncertainty in a different assay parameter. (4) Check whether both sources estimate the same quantity under comparable conditions and whether a defensible method supports combining them; otherwise retain them separately and explain the choice. (5) Examples include missing units, no traceable source, an unclear uncertainty type, an undocumented transformation, or an unstated intended model use.