Interpreting Exposure Distributions Across Phenotypes

Why Exposure Distributions Need More Than a Mean

A PBPK virtual population can predict a different exposure for every simulated subject. When subjects are grouped by pharmacogenetic phenotype, the resulting distributions can show whether exposure tends to shift across groups. But a group mean alone does not tell us whether the groups are well separated, how much individual exposure overlaps, or how many subjects reach an exposure range of concern. Those are distinct questions, and each calls for a different summary.

The same reasoning applies to AUC and to concentrations at a specified time. AUC summarizes exposure over a defined interval; a concentration distribution describes exposure at a particular sampling time or model-defined point in the dosing interval. Comparisons are interpretable only when dose, regimen, AUC interval or sampling time, and relevant simulation conditions are specified consistently.

Definition: A phenotype-specific exposure distribution is the distribution of a defined exposure measure, such as AUC or concentration at a specified time, among virtual subjects assigned to one phenotype. A population-level exposure distribution is the distribution across all represented phenotypes, combined using their population frequencies. A tail probability at threshold \(T\) is the proportion of subjects whose exposure exceeds \(T\), written \(P(X>T)\), where \(X\) denotes the exposure measure.

The phenotype groups may differ in their mean, variability, shape, or all three. A shift in means is not the same as a change in overlap: two groups can have different means but still contain many subjects with similar exposures. Conversely, groups with similar means can differ in their tails if one distribution is more variable or skewed.

Compare Location, Spread, and Overlap

For each phenotype, report a measure of location, such as the mean and median, together with measures of spread, such as the standard deviation, interquartile range, and selected quantiles. If the distribution is right-skewed, as exposure distributions often can be, the mean can be pulled upward by a small number of high values. The median and upper quantiles help show whether that is happening.

A useful overlap quantity is the probability that a randomly selected subject from one group has greater exposure than a randomly selected subject from another group. This is not the same as the percentage of the two distributions that occupies the same exposure range, nor does it describe an individual’s probability of belonging to a phenotype given an exposure. It is a comparison of two independently selected subjects.

Worked Example: Interpreting a Mean Shift Alongside Overlap

Suppose a hypothetical simulation gives normally distributed AUC values for two phenotype groups under the same regimen. Group A has mean 100 mg·h/L and standard deviation 50 mg·h/L; Group B has mean 160 mg·h/L and standard deviation 60 mg·h/L. Treat the two independent subject AUC values as normally distributed for this calculation.

Let \(X_A\) and \(X_B\) denote AUC for one subject sampled from each group. The difference \(D=X_B-X_A\) is normal, with mean \(160-100=60\) mg·h/L and variance \(60^2+50^2=6100\) \((\mathrm{mg\cdot h/L})^2\). Therefore its standard deviation is \(\sqrt{6100}=78.10\) mg·h/L.

P(X_B>X_A)=P(D>0) =\Phi\left(\frac{60}{78.10}\right) =\Phi(0.7682)\approx 0.779

Under the stated normality and independence assumptions, a randomly selected Group B subject has higher AUC than a randomly selected Group A subject about 77.9% of the time. This indicates a substantial tendency toward higher AUC in Group B, but it is not complete separation: the probability is below 100%, so pairwise reversals are expected. The calculation describes pairwise ordering, not the percentage of subjects above a clinical threshold.

The normal assumption in that example is a calculation aid, not a default property of PBPK outputs. When distributions are skewed, multimodal, or otherwise non-normal, empirical plots and direct comparisons of simulated values are often more informative. If using a log transformation or geometric summaries, state that choice and interpret the results on the corresponding scale.

How Phenotype Groups Form a Population Distribution

A population-level distribution is a frequency-weighted mixture of phenotype-specific distributions. Let phenotype \(g\) have population frequency \(w_g\), where the nonnegative weights sum to 1. For an exposure \(X\), the population probability of exceeding a threshold is the weighted sum of group-specific exceedance probabilities. The population mean is similarly weighted. The population spread, however, includes both within-phenotype variability and differences between phenotype means.

Result: If phenotype \(g\) has mean \(\mu_g\), variance \(\sigma_g^2\), and population frequency \(w_g\), then the population mean \(\mu\) and variance \(\sigma^2\) are

\mu=\sum_g w_g\mu_g,\qquad \sigma^2=\sum_g w_g\left[\sigma_g^2+(\mu_g-\mu)^2\right].

For a threshold \(T\), the population tail probability is \(P(X>T)=\sum_g w_gP(X>T\mid g)\).

The variance result follows by separating a subject’s deviation from the population mean into a within-group deviation and a group-mean deviation. Conditional on group \(g\), write \(X-\mu=(X-\mu_g)+(\mu_g-\mu)\). Squaring and taking the conditional expectation gives \(\sigma_g^2+(\mu_g-\mu)^2\), because the cross-product has expectation zero: \(E(X-\mu_g\mid g)=0\). Averaging over groups with weights \(w_g\) gives the displayed population variance. The first term captures within-group spread; the second captures spread due to differences between group means.

Worked Example: Calculating Population Mean and Variance

Consider a hypothetical population with phenotype A at frequency 0.20 and phenotype B at frequency 0.80. Their AUC means are 180 and 100 mg·h/L, and their standard deviations are 60 and 40 mg·h/L, respectively. Assume these summaries describe the same exposure definition and regimen.

The population mean is the weighted mean:

\mu=(0.20)(180)+(0.80)(100)=36+80=116\ \mathrm{mg\cdot h/L}.

For the variance, first calculate each group’s squared deviation from the population mean. These are \((180-116)^2=4096\) and \((100-116)^2=256\), in squared mg·h/L. Then combine within-group variance and between-group variance:

\begin{aligned} \sigma^2 &=(0.20)\left[60^2+(180-116)^2\right] +(0.80)\left[40^2+(100-116)^2\right]\\ &=(0.20)(7696)+(0.80)(1856)\\ &=1539.2+1484.8=3024\ (\mathrm{mg\cdot h/L})^2. \end{aligned}

Thus the population standard deviation is \(\sqrt{3024}=54.99\) mg·h/L. This is not the weighted average of the two group standard deviations: it also reflects the separation between their means. The population is a mixture of groups, so a single mean and standard deviation do not establish that the overall distribution is normal or even unimodal.

Use Thresholds to Examine Clinically Relevant Tails

A mean can be clinically uninformative when concern centers on unusually high or low exposure. If a relevant threshold \(T\) is available for a particular drug, endpoint, and regimen, estimate the fraction exceeding it in each phenotype and in the combined population. The threshold must be justified for that context; there is no universal AUC cutoff that applies to all drugs or clinical decisions.

Worked Example: Combining Phenotype-Specific Exceedance Rates

Suppose a hypothetical virtual cohort contains 1,000 subjects: 100 in phenotype A and 900 in phenotype B. For an illustrative, drug-specific threshold of 200 mg·h/L, 45 subjects in A and 45 subjects in B exceed the threshold. The observed exceedance rates are \(45/100=0.45\) and \(45/900=0.05\).

The population exceedance fraction is calculated from the group frequencies and conditional exceedance rates:

$$ P(X>200) =(0.10)(0.45)+(0.90)(0.05) =0.045+0.045 =0.090. $$

The estimated population fraction above the threshold is 9.0%, or 90 of 1,000 subjects. Although phenotype A represents only 10% of the cohort, it accounts for 45 of the 90 exceedances, or 50% of the simulated subjects above the threshold. This tail comparison can reveal a concentration of high exposures that the population mean would not communicate. The example treats the threshold and cohort as specified; it does not establish that exceeding this hypothetical threshold causes harm.

With a finite virtual cohort, the estimated tail fraction also varies with Monte Carlo sampling. Here, treating the 1,000 simulated subjects as independent draws, the approximate standard error of the overall estimated fraction is \(\sqrt{0.09(0.91)/1000}=0.00905\), or 0.905 percentage points. This sampling uncertainty is distinct from the spread of AUC values among individuals.

A compact reporting set might include phenotype frequencies; means and medians; standard deviations and interquartile ranges; selected quantiles such as the 5th, 50th, and 95th; and threshold exceedance rates when a defensible threshold exists. Plots of empirical cumulative distributions, density estimates, or violin/box plots can show shape and overlap. A histogram can help diagnose skewness or multiple modes, though its appearance depends on bin choice.

Interpret Comparisons in the Context of the Simulation

Phenotype-specific distributions are conditional summaries: they describe exposure among subjects assigned to each phenotype under the simulated covariates and model assumptions. If groups also differ in body size, organ function, co-medication, or other physiological characteristics, a raw difference between their exposure distributions combines those differences. It should not automatically be attributed to genotype alone. A comparison designed to isolate genotype effects must control or standardize the relevant simulated conditions, while a population prediction should preserve the intended joint distribution of genotype and physiology.

Phenotype assignment can also be uncertain. Functional effects may be uncertain, and genotype-based phenotype can differ from the expressed phenotype under some conditions. When the simulation represents this uncertainty, report which assignment or uncertainty assumptions were used. Do not imply that a sharply separated set of groups is justified if phenotype classification is itself uncertain.

The main alternatives answer different questions. Group means and mean ratios summarize location but can conceal tails. Medians and quantiles are more resistant to extreme values but do not by themselves quantify exceedance. Standard deviations and interquartile ranges summarize spread in different ways, while distribution plots expose shape. Threshold probabilities address a specified decision-relevant region but depend on the suitability of the threshold. No single summary replaces the others.

Common pitfall: “The phenotype with the higher mean has uniformly higher exposure, so its subjects are at higher exposure than the other phenotype.” A higher mean supports a statement about average exposure, not uniform ordering of individuals. The distributions may overlap substantially, and their upper tails may differ in ways not captured by the means. Compare spread, quantiles, overlap, and relevant tail probabilities before making a stronger claim.

Diagnostics and Limitations

Check that each group has enough simulated subjects to estimate its distribution and tails with useful precision. A very small phenotype group can produce unstable upper quantiles and exceedance rates even when the overall virtual cohort is large. Compare empirical distributions and quantiles with the reported summaries: a mean much larger than the median may indicate right-skewness, and a high 95th percentile may reveal a tail that the standard deviation alone does not describe.

Investigate unexpected patterns before interpreting them. Confirm that exposure units, dose, regimen, time interval, and phenotype labels are consistent; check for missing or duplicated subjects and implausible simulated values; and examine whether a small subgroup or influential extreme value drives a summary. A smooth fitted distribution can be useful, but it should not obscure multimodality or a sparse tail. Monte Carlo precision does not validate the PBPK model, the parameter assumptions, or the clinical relevance of a selected threshold.

Finally, a population mixture can change when phenotype frequencies change even if every phenotype-specific distribution stays fixed. Population means, variances, and tail probabilities are therefore population-specific. They should be reported alongside the phenotype weights and the conditions under which the virtual population was constructed.

Check Your Understanding

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

Two phenotype groups have different mean AUC values. Does this alone establish that their AUC distributions do not overlap? Explain.
A population has two phenotypes with frequencies 0.25 and 0.75. Their mean concentrations at a specified time are 12 and 8 mg/L. What is the population mean concentration?
For a threshold \(T\), the phenotype-specific exceedance probabilities are 0.20 and 0.04, and the corresponding frequencies are 0.10 and 0.90. Calculate the population exceedance probability.
Why can the population standard deviation not generally be obtained by taking the frequency-weighted average of phenotype-specific standard deviations?
Name two summaries that could reveal clinically relevant high exposure not apparent from a population mean.

Solutions. (1) No. Different means indicate a location difference, but the distributions may still overlap; compare their spread, shape, or an explicit overlap measure. (2) The weighted mean is \((0.25)(12)+(0.75)(8)=3+6=9\) mg/L. (3) The population exceedance probability is \((0.10)(0.20)+(0.90)(0.04)=0.020+0.036=0.056\), or 5.6%. (4) Population spread includes both within-phenotype variability and between-phenotype differences in means, so the variance decomposition must be used before taking a square root. (5) Examples include the 95th percentile and the fraction above a justified exposure threshold; an empirical distribution plot can also reveal the tail.