Worked Analysis of a Genotype-by-Inhibitor Effect on Clearance

Why Analyze the Joint Effect?

A genotype may affect clearance differently during inhibitor treatment than it does without treatment. In that case, a single genotype effect and a single inhibitor effect do not fully describe the four subgroup predictions. A genotype-by-inhibitor interaction represents this departure from a model in which the two effects combine consistently on a specified scale.

The scale matters. In this tutorial, clearance is modeled on the logarithmic scale, so “additive effects” means that the genotype and inhibitor coefficients add in log clearance. On the original clearance scale, those same effects combine multiplicatively. The interaction coefficient describes whether that multiplicative combination changes across subgroups.

Definition: Let $G_i=0$ for the reference genotype and $G_i=1$ for the chosen genotype group, and let $I_i=0$ without the inhibitor and $I_i=1$ during inhibitor exposure. A population model with a genotype-by-inhibitor interaction is $$ \log(CL_i)=\log(CL_{\mathrm{ref}})+\beta_GG_i+\beta_II_i+\beta_{GI}G_iI_i+\eta_i. $$ Here $CL_i$ is individual clearance, $CL_{\mathrm{ref}}$ is the typical clearance for the reference-genotype, no-inhibitor subgroup, and $\eta_i$ is an individual random effect, commonly modeled as $N(0,\omega^2)$. The additive model is the special case $\beta_{GI}=0$.

The fixed-effect predictions below refer to typical clearance, obtained by setting $\eta_i=0$. Under the stated normal random-effect model, this is the median individual clearance conditional on the covariates; it is not the arithmetic mean clearance across individuals. The coefficient $\beta_G$ describes the genotype contrast when $I=0$, and $\beta_I$ describes the inhibitor contrast when $G=0$. With an interaction, neither is a universal effect across the other covariate’s levels.

What the Four Subgroup Predictions Mean

The four combinations of genotype and inhibitor status make the model’s interpretation explicit. Setting $G$ and $I$ to zero or one gives:

\begin{aligned} CL_{00}&=CL_{\mathrm{ref}},\\ CL_{10}&=CL_{\mathrm{ref}}\exp(\beta_G),\\ CL_{01}&=CL_{\mathrm{ref}}\exp(\beta_I),\\ CL_{11}&=CL_{\mathrm{ref}}\exp(\beta_G+\beta_I+\beta_{GI}). \end{aligned}

For example, the genotype-to-reference clearance ratio without inhibitor is $\exp(\beta_G)$. With inhibitor, it is $\exp(\beta_G+\beta_{GI})$. The interaction is therefore a ratio of ratios:

\frac{CL_{11}/CL_{01}}{CL_{10}/CL_{00}}=\exp(\beta_{GI}).

A value of $\beta_{GI}=0$ means the genotype clearance ratio is the same at both inhibitor levels, equivalently that the inhibitor clearance ratio is the same for both genotype groups. A negative interaction coefficient means the combined group’s clearance is lower than predicted by the additive model on the log scale. It does not, on its own, establish a biological mechanism or a causal effect.

Worked Example: Calculating Clearance in All Four Subgroups

Suppose a hypothetical population PK analysis estimates $CL_{\mathrm{ref}}=10.0$ L/h, $\beta_G=-0.223$, $\beta_I=-0.357$, and $\beta_{GI}=-0.470$. These are illustrative fitted values, not results from a real study. Calculate the typical clearance for each subgroup.

For the reference genotype without inhibitor, $G=0$ and $I=0$, so $CL_{00}=10.0$ L/h.

For the chosen genotype without inhibitor, $CL_{10}=10.0\exp(-0.223)=10.0(0.800)=8.00$ L/h.

For the reference genotype with inhibitor, $CL_{01}=10.0\exp(-0.357)=10.0(0.700)=7.00$ L/h.

For the chosen genotype with inhibitor, include all three coefficients: $CL_{11}=10.0\exp(-0.223-0.357-0.470)=10.0\exp(-1.050)=3.50$ L/h.

Thus the fitted typical clearances are 10.0, 8.00, 7.00, and 3.50 L/h for reference/no inhibitor, chosen genotype/no inhibitor, reference/inhibitor, and chosen genotype/inhibitor, respectively. The interaction model predicts a lower clearance for the joint subgroup than the additive model would predict.

From an Interaction Coefficient to a Contrast

A coefficient is most useful when translated into a comparison on the original clearance scale. The genotype ratio without inhibitor in the example is $\exp(-0.223)=0.800$, or 80.0% of reference clearance. With inhibitor, the ratio is $\exp(-0.223-0.470)=0.500$, or 50.0%. The ratio of those two genotype ratios is $0.500/0.800=0.625$, matching $\exp(-0.470)=0.625$. Under the fitted model, the genotype-associated clearance ratio is 37.5% lower with inhibitor than without it, on this ratio-of-ratios comparison.

Worked Example: Quantifying Uncertainty in the Interaction

Suppose the hypothetical fitted interaction estimate is $\hat{\beta}_{GI}=-0.470$, with standard error $SE=0.200$. Using a normal-approximation 95% confidence interval, calculate an interval for the interaction and translate it to a ratio of ratios.

On the log-clearance scale, the interval is $-0.470\mathbin{\pm}1.96(0.200)=-0.470\mathbin{\pm}0.392$, giving $(-0.862,-0.078)$. Exponentiating both endpoints gives $(\exp(-0.862),\exp(-0.078))=(0.422,0.925)$.

The fitted ratio of ratios is 0.625, and this approximate interval ranges from 0.422 to 0.925. In context, the interval describes values compatible with the fitted model and its standard-error approximation for how the genotype clearance ratio changes with inhibitor status. It is not a 95% probability statement about the fixed parameter. The interval’s construction also depends on the model and approximation being adequate.

How the Interaction Changes a PK Prediction

A clearance contrast can matter for predicted concentration-time profiles, not just for a parameter table. The next calculation illustrates one implication under a deliberately simple structural assumption. It does not replace checking the fitted concentration model.

Worked Example: Implication for Half-Life at Fixed Volume

Assume, hypothetically, a one-compartment model with first-order elimination and the same typical volume of distribution $V=40.0$ L in every subgroup. Under these assumptions, $t_{1/2}=0.693V/CL$. Compare the reference/no-inhibitor half-life with that for the joint subgroup, both with and without the interaction.

For the reference/no-inhibitor subgroup, $t_{1/2}=0.693(40.0\text{ L})/(10.0\text{ L/h})=2.77$ h.

With the interaction, the joint subgroup has $CL=3.50$ L/h, so $t_{1/2}=0.693(40.0)/3.50=7.92$ h.

Without the interaction, the additive model predicts $CL_{11}=10.0\exp(-0.223-0.357)=10.0\exp(-0.580)=5.60$ L/h. Its corresponding half-life is $0.693(40.0)/5.60=4.95$ h.

In this simplified example, the interaction changes the predicted joint-subgroup half-life from 4.95 to 7.92 h. This comparison assumes equal volume and the stated one-compartment elimination model; if volume also varies by subgroup or the structural model is inadequate, these half-life calculations do not capture the full PK implications.

Fitting and Checking the Two Models

In a population PK analysis, the fixed-effect coefficients are estimated from concentration-time data through the selected structural and statistical model. The covariate terms do not fit observed subgroup averages directly: individual concentrations also depend on dose, sampling times, other PK parameters, individual variability, and residual error. A valid comparison therefore requires the same appropriate structural model, data handling, and other covariate terms in both candidate models, with the interaction term added in the interaction model.

The following short derivation shows why the no-interaction model makes a specific prediction. If $\beta_{GI}=0$, then $CL_{11}=CL_{\mathrm{ref}}\exp(\beta_G+\beta_I)$, while $CL_{10}CL_{01}/CL_{00}=CL_{\mathrm{ref}}\exp(\beta_G)\exp(\beta_I)$. Therefore:

\beta_{GI}=0 \quad\Longrightarrow\quad CL_{11}=\frac{CL_{10}CL_{01}}{CL_{00}}.

The equality is on the modeled typical-clearance scale. The interaction model allows the joint subgroup prediction to depart from that multiplicative combination. This algebra does not show that an interaction is supported by the data; support depends on the fitted model, the uncertainty, and whether the data can distinguish the alternatives.

The main alternative is the additive log-clearance model, which uses three fixed-effect terms rather than four and imposes the equality above. A separate estimate for each of the four subgroup clearances is another way to describe the same cells, but it gives up the structured parameterization and may be unstable when some cells contain little information. The interaction model can represent effect modification, but it uses an additional parameter and can be poorly determined when subgroup information is sparse.

Approach	What it assumes or estimates	Useful check
Additive log-clearance model	Genotype and inhibitor coefficients combine with $\beta_{GI}=0$	Compare its joint-subgroup prediction with the observed concentration patterns and the interaction model’s prediction
Interaction model	The genotype contrast may differ by inhibitor status	Check whether the interaction is estimated with useful precision and whether subgroup predictions are credible
Four separate subgroup clearances	Each genotype-by-inhibitor cell has its own typical clearance	Check whether each cell has enough data to support its estimate; do not assume sparse cells are well estimated

Before interpreting an interaction, inspect the data and model diagnostics across all four cells. Check subgroup counts, dose and sampling-time coverage, inhibitor exposure records, genotype coding, and whether other covariates are unevenly distributed. Examine observed and model-predicted concentration patterns by subgroup, as well as residual patterns and parameter uncertainty. A model may return a numerical estimate even when the data provide little information to distinguish an interaction from random variation or other model components.

Assumptions, Failure Modes, and Interpretation

Covariate definition and timing: The genotype groups and inhibitor status must be defined consistently with the intended comparison. If inhibitor status changes over the modeled period, a single status for an individual may misrepresent exposure; the covariate representation must match the time course being analyzed.
Cell information: The joint genotype-and-inhibitor subgroup needs informative observations. Very few individuals, narrow sampling, or limited dose information in a cell can make its predicted clearance and the interaction uncertain.
Structural and statistical adequacy: The interaction term cannot repair a misspecified structural PK model, inappropriate residual model, or poorly represented variability. Apparent subgroup patterns may reflect those problems rather than a covariate effect.
Scale-specific interpretation: No interaction on log clearance means equal ratios, not equal absolute changes in L/h. On another scale, the same data can yield a different definition of interaction.
Association, not mechanism: A fitted genotype-by-inhibitor contrast is conditional on the model and recorded covariates. It is not by itself proof that the inhibitor causes the genotype-specific difference or that a particular biological pathway explains it.

A common mistake is to read $\beta_G$ as the genotype effect in every inhibitor group. In the interaction model, $\beta_G$ is the genotype contrast only when $I=0$; when $I=1$, the contrast is $\beta_G+\beta_{GI}$. Likewise, $\beta_I$ is the inhibitor contrast only when $G=0$; when $G=1$, it is $\beta_I+\beta_{GI}$. Another mistake is to call the effects additive on the original clearance scale. The coefficients add on log clearance, while the corresponding clearance ratios multiply.

Key takeaway: Interpret a genotype-by-inhibitor effect by stating the reference subgroup, calculating all four typical clearances, and translating the interaction into a ratio of ratios. Then assess uncertainty, subgroup information, and model diagnostics before drawing a scientific conclusion.

Check Your Understanding

Use the model and hypothetical fitted values in this tutorial. Answers are included so that each calculation can be checked.

If $\beta_{GI}=0$, what relationship must hold among the four typical clearances?
Using the example coefficients, what is the inhibitor-to-no-inhibitor clearance ratio for the chosen genotype?
What does $\exp(\beta_{GI})=0.625$ mean in this analysis?
Why can an interaction estimate be unreliable when one genotype-by-inhibitor cell contains very little information?
Does $\beta_G$ describe the genotype contrast among inhibitor-exposed individuals? Explain.

Solutions. (1) $CL_{11}=CL_{10}CL_{01}/CL_{00}$. (2) The ratio is $\exp(\beta_I+\beta_{GI})=\exp(-0.357-0.470)=\exp(-0.827)=0.437$; typical clearance during inhibitor exposure is 43.7% of that without inhibitor for the chosen genotype. (3) The genotype clearance ratio with inhibitor is 0.625 times its ratio without inhibitor. (4) The joint subgroup prediction depends on information from that cell; sparse or uninformative data can produce high uncertainty and weak separation of the interaction from other model components. (5) No. Among inhibitor-exposed individuals, the genotype contrast on log clearance is $\beta_G+\beta_{GI}$, corresponding to a clearance ratio of $\exp(\beta_G+\beta_{GI})$.