Common Errors in Pharmacogenetic Virtual Populations

Why Small Genetic Assumptions Can Have Large PBPK Consequences

A pharmacogenetic virtual population links genetic variation to model parameters, such as enzyme abundance or intrinsic clearance, and then propagates those differences through a PBPK model. Errors at the genetic-input stage can therefore produce plausible-looking but misleading exposure distributions. The issue is not simply whether a variant is present: the model must also represent its function, the number of relevant gene copies, and how common the genotype is in the population being simulated.

The earlier material on pharmacogenetic virtual populations establishes how genotypes can be sampled and connected to PBPK parameters. This tutorial focuses on errors in those steps and on practical checks that help distinguish a defensible assumption from an unsupported shortcut. The examples use hypothetical data and simplified parameter relationships for teaching; they are not clinical recommendations.

Definition: A pharmacogenetic virtual population error is an input or interpretation mistake that misrepresents genetic variation, its functional consequences, or its frequency in the target cohort, and thereby distorts simulated parameters or outcomes. The error can occur even when the PBPK equations are implemented correctly.

Error 1: Treating Ancestry as an Individual Genotype

Ancestry, ethnicity, and geographic origin are not genotypes. A group label may be associated with different allele frequencies in a particular dataset, but it does not establish which alleles an individual carries. People within any broadly defined group vary genetically, and individuals with mixed ancestry may not be represented well by a single label. A label can also be a poor match to the population from which a frequency estimate was obtained.

For a virtual population, the distinction is between assigning every subject a genotype based on a group label and using a justified frequency distribution to generate variation across subjects. The latter still does not predict any one person’s genotype. It is a population-level modelling assumption, and it should reflect the target cohort and the available evidence. If individual genotypes are available, those data are different from assigning genotypes by ancestry category.

Worked Example: Replacing an Ancestry-Based Assignment

Suppose a hypothetical target cohort is being simulated for a biallelic locus. A suitable data source suggests an allele frequency of $p=0.20$ for the allele of interest. For illustration, assume random mating and Hardy–Weinberg equilibrium (HWE). The corresponding genotype proportions are $p^2$, $2p(1-p)$, and $(1-p)^2$.

The expected proportions are:

\begin{aligned} P(AA) &= 0.20^2 = 0.0400,\\ P(Aa) &= 2(0.20)(0.80) = 0.320,\\ P(aa) &= 0.80^2 = 0.640. \end{aligned}

In a hypothetical population of 1,000 subjects, this corresponds to expected counts of 40, 320, and 640, respectively. These are expectations, not guaranteed counts. If instead the modeller labels all subjects as belonging to a group and assigns everyone a single “typical” genotype, the resulting population would have no genotype variation at this locus and would not represent the stated frequency distribution.

The HWE calculation is only an illustration: its assumptions may not hold in the target cohort, and a suitable observed genotype distribution may be preferable. The correction is not to replace one ancestry label with another, but to use relevant genetic evidence and state the assumptions behind the resulting distribution.

Error 2: Giving Every Variant a Known Functional Effect

A detected sequence variant does not automatically have a known effect on protein expression, substrate-specific activity, or clinical phenotype. Evidence may be absent, conflicting, or limited to a different substrate or experimental system. A variant with uncertain function should not be silently converted into a reduced-function or no-function allele simply because that assignment makes the model easier to run.

The model should distinguish what is observed from what is inferred. Genotype data identify a genetic state; functional annotation supports (or fails to support) a relationship between that state and a model parameter. If the effect is uncertain, the uncertainty can be represented through explicitly defined alternatives, such as a reference-function case and one or more plausible effect scenarios. These scenarios are assumptions for sensitivity analysis, not established clinical classifications.

Worked Example: Handling an Uncertain Variant

In a hypothetical set of 100 virtual subjects, 20 carry a variant whose effect on a modelled enzyme has not been established. The baseline enzyme-related intrinsic clearance parameter is $10$ clearance units. Suppose a modeller assigns carriers half the baseline value without supporting evidence.

The resulting arithmetic mean of the assigned parameter across the 100 subjects would be:

\frac{80(10)+20(5)}{100} =\frac{900}{100} =9.00\ \text{clearance units}.

This value follows from the assumed assignment, not from evidence that the variant reduces clearance by 50%. It should not be reported as an evidence-based population estimate. A more transparent analysis would retain the reference-function assumption as one scenario and separately test specified alternatives, such as a carrier parameter of 8 clearance units or 5 clearance units, clearly labelling each as hypothetical. The PBPK model can then show how much the predicted outputs depend on the uncertain functional assumption.

The arithmetic mean of a parameter in this example is not itself a prediction of mean exposure. PBPK outcomes can depend nonlinearly on several pathways and physiological inputs, so the model must be run under each scenario before interpreting exposure changes.

Error 3: Ignoring Copy-Number Variation

A genotype representation that records only allele names can be incomplete when gene copies vary. Copy-number variation (CNV) includes deletions and duplications; some pharmacogenes can also have complex structures, including hybrid genes. In such cases, a label that appears to contain two alleles may fail to capture how many copies of a functionally relevant allele are present. The biological consequence depends on the gene, the structure of the copies, and the evidence about their function.

This is not fixed by assuming that every extra copy has the same effect. A duplicated copy may be functional, nonfunctional, or structurally complex. A model must represent the relevant copy-number information and specify how each represented copy maps to the PBPK parameter. When copy-number status is missing or uncertain, that limitation should be stated rather than hidden by coding the subject as an ordinary two-allele genotype.

Worked Example: A Simplified Copy-Number Calculation

Consider a hypothetical enzyme model in which one functional copy contributes 6 L/h to an enzyme-related intrinsic-clearance component. Other pathways contribute 4 L/h. Assume, only for this simplified example, that each functional copy contributes equally and additively, and that total clearance is linear under the conditions considered.

For genotype $*1/*4$, suppose $*1$ is treated as one functional copy and $*4$ as contributing no function. The modelled total is $4+6=10$ L/h. For $*1x2/*4$, suppose the duplication notation indicates two functional $*1$ copies and one nonfunctional $*4$ copy. The modelled total is $4+2(6)=16$ L/h.

Under the additional simplified assumption that exposure is inversely proportional to clearance at the same dose, the ratio of exposure for the duplicated genotype to that for $*1/*4$ is:

\frac{\mathrm{Exposure}_{*1x2/*4}} {\mathrm{Exposure}_{*1/*4}} \approx \frac{10\ \mathrm{L/h}}{16\ \mathrm{L/h}} =0.625.

This calculation illustrates why copy number can matter; it does not establish a clinical exposure ratio. It assumes equal expression per functional copy, additive contributions, no change to other pathways, and linear exposure-clearance behaviour. A real PBPK implementation must use gene- and substrate-appropriate evidence and test whether those assumptions are defensible.

Error 4: Reusing One Frequency Table for Every Target Cohort

An allele-frequency table is an estimate from a particular source population, with its own sampling, ancestry composition, and inclusion criteria. It is not a universal frequency table. Transferring it to a different target cohort without justification can change the simulated genotype proportions and, through genotype-dependent parameters, the predicted population exposure distribution.

Frequency estimates can also differ in how variants and alleles are defined, whether rare variants were measured, and how missing or ambiguous genotypes were handled. A table should therefore be checked for population relevance, variant definition, sample size and uncertainty, and compatibility with the intended genotype representation. If no suitable estimate exists, alternative plausible frequencies can be examined as scenarios rather than presenting one borrowed value as exact.

Worked Example: Frequency Transfer Changes Expected Counts

Suppose a hypothetical reduced-function allele has frequency $p=0.10$ in one reference dataset and $p=0.30$ in a dataset considered more relevant to a target cohort. To isolate the consequence of transferring the frequency, assume HWE in both cases. The expected frequency of individuals with two copies of this allele is $p^2$.

Using $p=0.10$, the expected proportion is $0.10^2=0.0100$, or 1.00%. In a simulated cohort of 200, the expected count is $200(0.0100)=2$. Using $p=0.30$, the expected proportion is $0.30^2=0.0900$, or 9.00%, giving $200(0.0900)=18$ expected subjects.

The difference is 16 expected subjects, and the second expected count is nine times the first. This result does not prove either frequency is correct for a particular cohort; it shows why an apparently small choice of frequency source can materially alter a virtual population. HWE is an explicit simplifying assumption here, not a substitute for checking the target cohort’s genotype distribution.

A Practical Check: Trace the Assumption to the PBPK Output

A useful way to diagnose these errors is to make each link in the modelling chain explicit: genetic observation, allele or copy interpretation, functional assumption, parameter mapping, and simulated outcome. Then ask whether the output changes when an uncertain link is varied. The calculation below shows why frequency errors affect population predictions.

Derivation: Let $g$ index possible genotypes, let $P(g)$ be the proportion of genotype $g$ in the target population, and let $Y(g)$ be the PBPK outcome for that genotype under the model assumptions. The population mean outcome is $$ E[Y]=\sum_g P(g)Y(g). $$ If the model uses incorrect proportions $\widetilde{P}(g)$, its mean is $\sum_g \widetilde{P}(g)Y(g)$. Subtracting the target mean gives $$ \widetilde{E}[Y]-E[Y] =\sum_g\left[\widetilde{P}(g)-P(g)\right]Y(g). $$ Because both sets of proportions sum to one, the differences in proportions sum to zero. But the outcome error need not be zero: it depends on how those differences align with genotype-specific outcomes. If all genotypes have the same $Y(g)$, changing their proportions does not change the mean; if outcomes differ, frequency misspecification can shift it.

This result is a bookkeeping identity, not a guarantee that a mean fully describes exposure. The entire distribution may matter, particularly when genotype groups have distinct tails or when the outcome is nonlinear. Inspect genotype counts and group-specific PBPK outputs as well as an overall mean.

Practical diagnostics include checking that simulated genotype frequencies match the specified input distribution within expected Monte Carlo variation; confirming that genotype-to-parameter mappings include the intended functional and copy-number states; and comparing outputs under alternative plausible functional assignments or frequency sources. Unexpectedly identical genotype-specific parameters may indicate that genetic variation was not propagated. Conversely, a large output difference should prompt review of the parameter mapping and model assumptions, not immediate attribution to a biological effect.

Choosing a More Defensible Approach

The main alternative to a single, fixed genetic assignment is not an assumption-free model: every approach needs evidence and modelling choices. The useful comparison is how explicitly each approach represents individual data, uncertainty, and population variation.

Approach	Useful when	Main limitation to check
Use observed individual genotypes and copy-number data	Relevant, sufficiently complete data are available for the intended subjects	Missing or ambiguous calls and nonrepresentative sampling can remain
Sample genotypes from a relevant frequency distribution	The aim is a virtual population and individual genotype data are unavailable	Results depend on the frequency source and any assumptions used to construct genotype proportions
Run alternative frequency or function scenarios	Evidence is uncertain or more than one target population is plausible	Scenarios describe conditional model outcomes; they do not resolve which assumption is true

These approaches can be combined where appropriate, for example by using observed genotype data for part of a cohort and explicitly handling missing information. However, the method chosen should match the question. A population-level simulation should not be presented as if it assigns reliable genotypes to named individuals.

Common Pitfalls and Limitations

“The subject has this genotype because of their ancestry label.” A group label does not determine an individual genotype. Use individual genetic data where available, or a justified population distribution for population-level simulation.
“A variant is present, so its functional effect is known.” Presence and function are separate claims. Preserve uncertainty and test explicitly stated alternatives when evidence is limited.
“Two allele labels fully describe the genotype.” Deletions, duplications, and complex structures can make that representation incomplete. Check whether copy number and gene structure are relevant and available.
“This frequency table is a standard value for everyone.” Frequency estimates depend on their source and variant definitions. Check whether the data match the target cohort and state any transfer assumption.
“A changed mean exposure proves the genetic assumption is correct.” A simulation result is conditional on its inputs. Compare assumptions and inspect intermediate parameters before drawing biological conclusions.

Even careful modelling is limited by available genetic data, functional evidence, and the validity of the genotype-to-parameter relationship for the drug and model context. Sensitivity analysis can reveal dependence on an uncertain assumption, but it does not replace evidence. Keep the distinction clear between measured genotype information, functional interpretation, and the consequences predicted by the PBPK model.

Check Your Understanding

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

Why is it incorrect to assign every subject in an ancestry category the same pharmacogenetic genotype?
A variant is observed in a virtual subject, but its effect on the modelled enzyme is uncertain. What should the modeller avoid, and what is one defensible way to examine the uncertainty?
Under the simplified assumptions in the copy-number example, what total clearance is assigned to a genotype with three functional copies if other pathways contribute 4 L/h and each functional copy contributes 6 L/h?
Under HWE, a biallelic allele has frequency $p=0.20$. What is the expected proportion with two copies of that allele?
Why can a frequency-table mismatch change the simulated mean PBPK outcome even though genotype-specific outcome calculations are unchanged?

Solutions. (1) An ancestry category is not an individual genotype; individuals within a category vary, so use observed individual data or a justified population distribution. (2) Avoid assigning an unsupported functional effect as though it were known. Compare explicitly labelled, plausible functional scenarios and report that the resulting outputs are conditional on them. (3) The total is $4+3(6)=22$ L/h, given the stated additive assumptions. (4) The proportion is $p^2=0.20^2=0.0400$, or 4.00%. (5) The population mean weights each genotype-specific outcome by its genotype proportion, $E[Y]=\sum_g P(g)Y(g)$. Changing $P(g)$ can therefore change the mean when genotype-specific outcomes differ.