A CYP2D6 Virtual-Population Construction Workshop

From CYP2D6 Alleles to a Virtual Cohort

A pharmacogenetic PBPK cohort is not created by assigning one average genotype to a population. It is constructed by representing the genetic variation that changes drug disposition, assigning that variation to virtual subjects, and then carrying the resulting parameters through the exposure model. CYP2D6 is a useful workshop example because different diplotypes can imply markedly different enzyme activity and therefore different clearance.

The calculations below use explicitly hypothetical allele frequencies and a simplified clearance model. They illustrate a reproducible construction workflow; they are not clinical frequency estimates, clinical dosing guidance, or a substitute for a drug-specific PBPK model. The earlier material on pharmacogenetic phenotype frequencies and genotype-dependent parameters supplies the general concepts used here. The focus is on connecting those concepts in one CYP2D6 cohort and checking each calculation.

Definition: A diplotype is the pair of alleles carried at a gene locus. An activity score (AS) is a numerical summary of the functional activity assigned to the alleles in that diplotype. A virtual cohort is a set of simulated subjects whose assigned genotypes and other covariates represent a defined target population under stated assumptions.

For this workshop, use three hypothetical CYP2D6 alleles: *1 with allele activity value 1, *4 with value 0, and *10 with value 0.25. The diplotype AS is the sum of the two allele values. For clarity, assign phenotype labels using these operational bins: poor metabolizer (PM), AS = 0; intermediate metabolizer (IM), 0 < AS < 1.25; normal metabolizer (NM), 1.25 ≤ AS ≤ 2.25; and ultrarapid metabolizer (UM), AS > 2.25. These bins are stipulated for this worked example; phenotype translation can depend on the allele-function assignments and classification standard being applied. No copy-number variants are included here.

Constructing Diplotype Probabilities

Suppose the target population has allele frequencies $p_{*1}=0.70$, $p_{*4}=0.20$, and $p_{*10}=0.10$, which sum to 1. Under random mating, Hardy–Weinberg proportions, and no allele-specific sampling or selection in this simplified construction, the probability of a homozygous diplotype $i/i$ is $p_i^2$, and the probability of a heterozygous diplotype $i/j$ is $2p_ip_j$. These are population probabilities, not guarantees that a finite simulated cohort will have exactly those proportions.

Worked Example: Enumerating the CYP2D6 Diplotypes

Calculate each diplotype probability, its activity score, and its phenotype label. The six unordered diplotypes are the complete set for these three alleles.

Diplotype	Probability calculation	Probability	AS	Phenotype
1/1	0.70²	0.49	2.00	NM
1/4	2(0.70)(0.20)	0.28	1.00	IM
1/10	2(0.70)(0.10)	0.14	1.25	NM
4/4	0.20²	0.04	0	PM
4/10	2(0.20)(0.10)	0.04	0.25	IM
10/10	0.10²	0.01	0.50	IM

For example, the heterozygous *1/*4 probability is $2(0.70)(0.20)=0.28$, and its AS is $1+0=1$. The diplotype probabilities sum to $0.49+0.28+0.14+0.04+0.04+0.01=1.00$, a necessary check that the enumeration is complete and normalized.

The phenotype proportions follow by adding probabilities within each operational category: PM = 0.04, IM = $0.28+0.04+0.01=0.33$, NM = $0.49+0.14=0.63$, and UM = 0. Their sum is 1.00. A zero UM proportion here is a consequence of the restricted allele set and the absence of duplications, not a general claim about CYP2D6 populations.

Scaling Clearance and Interpreting Exposure

Activity score can be used to scale the CYP2D6-mediated component of intrinsic clearance in a simplified model. Let the reference AS be 1, and suppose CYP2D6 contributes a fraction $f=0.60$ of total clearance at that reference activity. The remaining fraction, $1-f=0.40$, is assumed not to vary with CYP2D6 AS. Under linear, time-invariant pharmacokinetics at the same dose and bioavailability, exposure is inversely proportional to total clearance. These assumptions are deliberately restrictive and must be checked before using the result as a PBPK prediction.

Model: With reference total clearance normalized to 1, the relative total clearance at activity score $AS$ is $0.40+0.60(AS/1)$. The resulting relative exposure ratio is the reference clearance divided by this clearance: $$ R(AS)=\frac{1}{0.40+0.60AS}. $$ Here $R=1$ denotes exposure at the reference AS of 1. The ratio is dimensionless; actual exposure units depend on the outcome being modelled, such as AUC in concentration·time units.

This expression follows from adding the clearance components. The invariant pathway contributes $0.40$, while the CYP2D6 component contributes $0.60AS$ because the reference AS is 1. Total relative clearance is therefore $CL_{\mathrm{rel}}=0.40+0.60AS$. If dose and bioavailability are held fixed and exposure is proportional to $1/CL$, then $R=1/CL_{\mathrm{rel}}$. This derivation also shows why a zero CYP2D6 AS does not imply zero total clearance: the non-CYP2D6 component remains.

Worked Example: Clearance Scaling by Diplotype

Calculate relative clearance and relative exposure for *4/*4 and *4/*10. The AS values are 0 and 0.25, respectively, from the allele scores stated above.

For *4/*4, $AS=0$. Thus $CL_{\mathrm{rel}}=0.40+0.60(0)=0.40$, and $R=1/0.40=2.500$. Under this model, the predicted exposure is 2.500 times the reference exposure, not infinite, because other clearance pathways retain 40% of reference total clearance.

For *4/*10, $AS=0.25$. Then $CL_{\mathrm{rel}}=0.40+0.60(0.25)=0.55$, and $R=1/0.55=1.818$. The model therefore predicts 1.818 times reference exposure. This calculation assumes the same dose, bioavailability, non-CYP2D6 clearance, and physiological covariates in the two comparisons.

For the six diplotypes, the same formula gives relative exposure ratios of 0.625 for *1/*1, 1.000 for *1/*4, 0.870 for *1/*10, 2.500 for *4/*4, 1.818 for *4/*10, and 1.429 for *10/*10, to three significant figures. A phenotype-level mean should be weighted by the diplotype probabilities conditional on that phenotype. It is not generally the exposure ratio at the phenotype’s average AS, because $R(AS)$ is nonlinear and the phenotype contains different diplotypes.

Worked Example: Mean Exposure Ratio Within the IM Group

The IM phenotype has total probability 0.33 and contains *1/*4, *4/*10, and *10/*10. Their probabilities in the full population are 0.28, 0.04, and 0.01. Using the relative exposure ratios calculated from the clearance model, the conditional mean is:

\bar{R}_{\mathrm{IM}}= \frac{0.28(1.000)+0.04(1.81818)+0.01(1.42857)}{0.33} = \frac{0.367013}{0.33} \approx 1.112.

Thus the model’s mean relative exposure in this IM group is approximately 1.112 times the reference exposure. The weights are conditionalized by dividing by 0.33; using 0.28, 0.04, and 0.01 without that denominator would instead give the contribution of the IM group to the population-wide mean, not the mean within IM.

As a separate arithmetic check, suppose a different hypothetical PBPK run reports ratios of 1.429, 2.105, and 1.818 for three cells with full-population weights 0.28, 0.04, and 0.01, respectively. The weighted conditional mean from those reported outputs is:

\frac{0.28(1.429)+0.04(2.105)+0.01(1.818)}{0.33} = \frac{0.40012+0.08420+0.01818}{0.33} = \frac{0.50250}{0.33} \approx 1.523.

This second value is conditional on those separate PBPK outputs; it is not the mean produced by the clearance formula above. Keeping the source of each ratio explicit prevents a weighted calculation from silently mixing incompatible model outputs.

Turning the Enumeration Into a Virtual Population

For a small genotype set, exact enumeration is often the clearest construction: retain each diplotype probability and calculate its consequences directly. For a finite virtual cohort, a subject-level implementation can sample diplotypes according to the probabilities, assign AS and phenotype, and then sample or assign physiological covariates. The earlier material on sampling discrete genotypes, continuous variability, and dependence among covariates applies when expanding this workshop into a larger cohort.

If the cohort size is $N$, a Monte Carlo cohort will have observed phenotype proportions that fluctuate around the probabilities 0.04, 0.33, 0.63, and 0 under this construction. In particular, a cohort with no PM subjects is possible when $N$ is small, even though the underlying PM probability is 0.04. Exact enumeration avoids this sampling variation for genotype-cell summaries, while sampling is useful when individual-level covariates and PBPK simulations are needed.

Specify the target and allele inputs.
State the target population, allele frequencies, functional assignments, and whether copy-number variants or uncertain alleles are represented.

Construct diplotypes and verify probabilities.
Use an explicit sampling assumption such as random mating, calculate all diplotype probabilities, and check that they sum to 1.

Map diplotypes to parameters and phenotypes.
Calculate AS from allele values, apply stated phenotype bins, and preserve the link between each subject’s genotype and CYP2D6-dependent parameter.

Run the exposure model and summarize by group.
Keep dose and comparison conditions explicit; calculate group summaries using the relevant subject or genotype-cell weights.

Diagnostics, Alternatives, and Limitations

Several checks can reveal a construction error before interpreting exposure differences. First, allele frequencies and diplotype probabilities must each sum to 1. Second, every diplotype must receive one defined AS and one phenotype under the stated rules. Third, the reference genotype must give $R=1$; in this example, AS = 1 yields $1/(0.40+0.60)=1$. Fourth, increasing AS must increase the CYP2D6 clearance component and, under the model, decrease relative exposure. A contrary trend suggests a coding or parameter-mapping error.

The analytic model and the sampled PBPK cohort answer related but different questions. Exact enumeration provides stable, transparent weighted results when the genotype set is small and the model is deterministic. Monte Carlo sampling represents individual variability and can support subject-level PBPK runs, but its group estimates have finite-sample variation. Stratified sampling can ensure that rare genotypes are represented for within-group comparisons; if the goal is a population-wide mean, those deliberately altered sampling fractions must be corrected with population weights.

The clearance equation is not a universal CYP2D6 exposure law. It assumes a fixed CYP2D6 fraction of total clearance, linear scaling of the CYP2D6 component with AS, and otherwise identical subjects. In a full PBPK analysis, enzyme abundance, hepatic blood flow, substrate concentration, other enzymes, absorption, distribution, and covariate relationships can alter the prediction. Different genotypes may also occur alongside different physiological covariates in a target population. Therefore, a group exposure difference in a simulation is a model-based comparison under its assumptions, not evidence that genotype alone caused a clinical difference.

Common pitfall: Averaging exposure ratios using the unconditional genotype probabilities and then calling the result a phenotype-specific mean gives the wrong estimand. For a phenotype with population probability $P(G)$, divide the sum of each included cell’s population weight times its exposure by $P(G)$. Also do not assume $R(\text{mean AS})$ equals the weighted mean of $R(AS)$; nonlinearity generally makes these different.

Practice and Solutions

Check Your Understanding

Use the allele values, frequencies, phenotype bins, and clearance model defined above. Solutions are included so each calculation can be checked.

What is the probability, AS, and phenotype of diplotype *1/*10?
Calculate the total NM proportion.
Calculate the relative exposure ratio for AS = 0.50.
Why should a phenotype-specific mean exposure ratio be divided by the phenotype probability?
Name one reason the analytic clearance calculation may differ from a full PBPK simulation.

Solutions. (1) The probability is $2(0.70)(0.10)=0.14$; AS is $1+0.25=1.25$, so the stipulated bin assigns NM. (2) NM includes *1/*1 and *1/*10, giving $0.49+0.14=0.63$. (3) $R=1/[0.40+0.60(0.50)]=1/0.70=1.429$. (4) The numerator weighted by full-population probabilities is the group’s contribution to the population mean; dividing by the group probability conditions the weights on membership in that phenotype and yields its mean. (5) For example, the analytic model fixes the non-CYP2D6 clearance fraction and assumes CYP2D6 clearance scales linearly with AS, whereas a full PBPK model can represent different physiology or nonlinear processes.