Genotype-by-Co-medication Interactions

Why Model a Genotype-by-Co-medication Interaction?

A co-medication may alter a pharmacokinetic parameter, but its effect need not be identical for every genetic group. For example, a drug that inhibits an enzyme might have a different association with clearance among people with different enzyme genotypes. A model with separate genotype and medication main effects assumes a particular pattern: the medication effect is the same on the model’s chosen scale across genetic groups. An interaction term allows that effect to vary.

The aim is not to label a combination of genotype and medication as inherently important. It is to specify which parameter may differ, define the reference groups, and state exactly what comparison the interaction represents. The interpretation depends on the parameter scale and on how genotype and medication status are coded.

A Binary Interaction Model

For illustration, let \(G=0\) denote a reference genotype group and \(G=1\) a comparison group, such as carriers of a specified allele. Let \(C=0\) indicate that the co-medication is absent and \(C=1\) that it is present. These are illustrative indicator codings; genotype can instead be represented by allele dosage or by multiple categories. Suppose the typical value of clearance, \(CL\), is modeled as

Here \(\theta_0\) is the log typical clearance for the reference genotype group without the co-medication. The coefficient \(\beta_G\) is the genotype contrast when \(C=0\); \(\beta_C\) is the medication contrast when \(G=0\); and \(\beta_{GC}\) is the interaction coefficient. If the population PK model includes individual random effects, such as \(\eta_{CL}\), they can be included in the individual-level parameter model as appropriate. The covariate coefficients above describe the typical value, not an individual’s observed concentration.

Exponentiating the model makes the four group-specific typical clearances explicit:

The first index is \(G\), and the second is \(C\). Within the reference genotype group, the co-medication’s clearance ratio is \(CL_{01}/CL_{00}=\exp(\beta_C)\). Within the comparison genotype group, that ratio is \(CL_{11}/CL_{10}=\exp(\beta_C+\beta_{GC})\). Thus, \(\beta_C\) alone is not the medication effect for every participant: it is the effect when \(G=0\).

Deriving the Interaction Contrast

The interaction can be derived directly by comparing the two medication effects. On the log scale, the medication contrast for \(G=0\) is \(\log(CL_{01})-\log(CL_{00})=\beta_C\). For \(G=1\), it is \(\log(CL_{11})-\log(CL_{10})=\beta_C+\beta_{GC}\). Subtracting the first contrast from the second gives the interaction coefficient:

Exponentiating this difference gives a ratio of ratios:

This derivation also clarifies the meaning of “no interaction.” The null restriction is \(H_0:\beta_{GC}=0\), equivalent to \(\exp(\beta_{GC})=1\). It means the co-medication ratios are equal across the two genotype groups under this model. It does not mean that neither genotype nor co-medication is associated with clearance; either or both main effects can be nonzero.

Estimating an Interaction from Group Contrasts

In practice, model estimates come from the PK data and the specified structural and variability models, rather than from four raw group means alone. Still, a four-cell example makes the contrast intuitive. Suppose a hypothetical analysis gives typical clearance estimates of 6.0 L/h and 4.8 L/h for the reference genotype without and with medication, and 7.2 L/h and 3.6 L/h for the comparison genotype without and with medication.

For a statistical test in this binary case, the interaction-specific null is \(H_0:\beta_{GC}=0\), against an alternative allowing \(\beta_{GC}\ne0\). The coefficient’s standard error and an appropriate inferential procedure are needed to assess uncertainty. The group-contrast arithmetic alone is not a test and does not account for the population model, residual error, or uncertainty in the estimates.

Other Genotype Codings

The interaction structure must match the scientific question and the chosen genotype representation. With allele dosage \(G=0,1,2\), a linear dosage interaction can be written using the same product term \(GC\). Under this coding, each additional allele changes the medication contrast by the same amount on the log scale. That is a modeling assumption, not a mathematical consequence of having three genotype values.

If the scientific question does not justify a linear dosage effect, represent genotype with indicator variables for its categories and include a medication interaction for each non-reference category. For three genotype groups, this generally gives two interaction coefficients. The no-interaction null then requires both coefficients to equal zero. The genotype main-effect coefficients also have reference-group interpretations, and changing the reference category changes coefficient values but not the model’s fitted group contrasts when the model is reparameterized correctly.

Model Assumptions, Checks, and Alternatives

An interaction coefficient is interpretable only relative to the model that defines it. Check that genotype and medication coding are explicit, that the co-medication status corresponds to the relevant PK observation or interval, and that the interaction is applied to the intended PK parameter. If medication exposure changes over time, using a single fixed status when participants switch treatment can misrepresent the covariate pattern. The covariate history should align with the model’s observation times and the intended exposure definition.

The log-linear form also assumes multiplicative covariate effects on the parameter. A model can show no interaction on the log-clearance scale while showing a difference in absolute clearance changes, because equal ratios do not imply equal differences in L/h. Conversely, an additive-scale interaction need not be absent just because the multiplicative interaction is absent. State the scale whenever reporting effect modification.

Useful checks include confirming the four genotype-medication combinations are represented in the data, verifying the product term was coded as intended, and checking whether fitted group-specific predictions align with the reported coefficients. Estimates that change markedly under reasonable coding or model choices, or that have very large uncertainty, signal that the interaction may be weakly supported. Sparse subgroup representation can make an interaction hard to estimate; detailed identifiability questions require additional consideration beyond the contrast definition here.

The main alternative is a main-effects-only model, which sets \(\beta_{GC}=0\) and assumes a common medication ratio across genotype groups. A stratified analysis estimates separate medication effects within genotype groups, which can be intuitive but may not share information across groups and can be inefficient when data are limited. A regression interaction estimates the difference between those group effects within one model. For categorical genotypes, a dosage interaction is a more restrictive alternative than category-specific interactions because it imposes a linear per-allele pattern.

A co-medication interaction is an association in the fitted model; it does not by itself show that the medication caused the parameter difference. Other features of treatment and patient characteristics may also differ between groups. Nor does a clearance interaction automatically imply the same interaction in exposure, efficacy, or safety: those outcomes depend on the full PK and, where relevant, pharmacodynamic model.

Common Mistakes

• Reading \(\beta_C\) as the medication effect for everyone. In the binary model it is the medication log-ratio when \(G=0\); for \(G=1\), the log-ratio is \(\beta_C+\beta_{GC}\).
• Calling the interaction coefficient a percentage-point change. On the log-parameter scale, exponentiating \(\beta_{GC}\) gives a ratio of medication ratios, not an absolute difference in clearance.
• Assuming no interaction means no covariate effects. The restriction \(\beta_{GC}=0\) permits nonzero genotype and medication main effects.
• Reporting a dosage interaction as unrestricted genotype-specific effects. A linear dosage product imposes a common per-allele pattern; category-specific terms are needed for unrestricted contrasts.
• Interpreting the fitted association as causal. The interaction model describes conditional parameter contrasts under its assumptions; it does not alone establish a causal medication effect.