From a Fitted Model to a Virtual Individual
A fitted population PK model describes both typical pharmacokinetic behavior and variability between individuals. To justify a regimen, it is often useful to move from that population description to individual concentration-time profiles: each profile represents a virtual individual with a particular set of model parameters, exposed to a specified dose history. Exposure metrics can then be calculated for each profile.
The basic operation is straightforward, but its interpretation depends on what is held fixed and what is sampled. A virtual individual’s parameters are generated from the fitted model’s population parameters and its interindividual random-effects distribution. Those parameters, together with the dose regimen and structural PK equations, determine the model-predicted concentrations. Unless the goal is to simulate noisy observations, residual error is not added to the predicted profile.
For a simple one-compartment model with an intravenous bolus dose \(D\), individual clearance \(CL_i\), and individual volume of distribution \(V_i\), the predicted concentration at time \(t\) after the dose is
Here, \(C_i(t)\) is in mg/L when \(D\) is in mg and \(V_i\) is in L; \(CL_i\) is in L/h and \(t\) is in hours. This equation assumes a single, instantaneous IV bolus, linear elimination, and a one-compartment disposition model. It is an example of a structural model, not a universal formula for every drug or regimen.
Assigning Individual Parameters and Generating Profiles
A common parameterization models positive PK parameters using exponential random effects. For example,
The typical values \(CL_{\mathrm{typ}}\) and \(V_{\mathrm{typ}}\) are the model’s typical parameters at the reference covariate values. The random effects \(\eta_{CL,i}\) and \(\eta_{V,i}\) describe individual deviations on the log scale. Under a normal random-effects model, each effect has mean zero and a fitted variance. Covariate effects, if present, must also be applied using the fitted model and the covariate values assigned to the virtual individual.
The simulation procedure is conditional on the model and the choices made for the virtual cohort. It requires a specified regimen, a time grid, covariate values or a defined way to assign them, and a rule for generating random effects. If the fitted model includes dependence between random effects, the joint distribution must be respected; drawing each effect independently would change the model. This tutorial uses diagonal random-effects covariance in the worked examples.
State the regimen, time interval, covariates, fitted parameter values, and random-effects distribution. Make clear whether parameter estimates are treated as fixed.
Draw random effects according to the fitted model and transform them into positive individual PK parameters. Apply any covariate relationships in the model.
Use each individual’s parameters and the specified dose history to calculate concentrations at the chosen times.
Compute the requested metric over its defined interval, verify units and numerical behavior, and retain the individual-level results for interpretation.
Worked Example: A Single-Dose IV Profile for One Virtual Individual
Suppose a hypothetical fitted one-compartment model has \(CL_{\mathrm{typ}}=5\) L/h and \(V_{\mathrm{typ}}=40\) L. Its random-effects standard deviations are 0.30 for clearance and 0.20 for volume, with zero covariance. Consider one virtual individual assigned \(\eta_{CL,i}=\ln(1.2)\) and \(\eta_{V,i}=\ln(0.9)\), and give that individual a 400 mg IV bolus. These stipulated effects correspond to multiplicative factors of 1.2 and 0.9.
The individual parameters are
The initial concentration is \(400/36.0=11.1111\) mg/L, and the elimination rate constant is \(6.00/36.0=0.166667\) h\(^{-1}\). Thus the profile is \(C_i(t)=11.1111\exp(-0.166667t)\) mg/L, with \(t\) in hours. At 2 h,
At 8 h, using the same displayed parameter values,
For this linear IV bolus model, the total AUC from time zero to infinity is \(D/CL_i\), or \(400/6.00=66.67\) mg·h/L. The profile and AUC are conditional on this individual’s assigned parameters; they are not observations from an actual person or study.
Exposure Metrics and a Short Derivation
An exposure metric must be tied to both a definition and a time interval. For example, \(AUC_{0,T}\) is the integral of concentration from time zero through time \(T\), while \(AUC_{0,\infty}\) extends to infinity after a single dose. \(C_{\max}\) is the maximum concentration over the specified interval, and \(C_{\min}\) is the minimum over a specified dosing interval. A grid-based profile may require numerical integration or a search over the grid; an analytic expression is preferable when available and appropriate.
For the one-compartment IV bolus model, the AUC result follows directly by integrating the profile. Set \(k_i=CL_i/V_i\), which has units of h\(^{-1}\). Then
The integral converges because \(k_i>0\), as expected for positive clearance and volume. This also shows why, under this particular linear model, AUC after an IV bolus depends on dose and clearance but not on volume. Volume still affects the initial concentration and the shape of the profile. The result should not be carried over without checking the assumptions: nonlinear elimination, different routes, or a different structural model can change the relationship.
Worked Example: Comparing Exposure Across Three Virtual Individuals
Use the same hypothetical population values, \(CL_{\mathrm{typ}}=5\) L/h and \(V_{\mathrm{typ}}=40\) L, with a 300 mg IV bolus. Consider three stipulated virtual individuals with \((CL_i,V_i)\) values of (4 L/h, 30 L), (5 L/h, 40 L), and (8 L/h, 50 L). These values can be represented by log-scale random effects under the model; they are supplied here to make the arithmetic transparent.
For each individual, the IV bolus initial concentration is \(D/V_i\), and the total AUC is \(D/CL_i\). For individual 1,
For individual 2, \(C_i(0)=300/40=7.50\) mg/L and \(AUC_{0,\infty,i}=300/5=60.0\) mg·h/L. For individual 3, \(C_i(0)=300/50=6.00\) mg/L and \(AUC_{0,\infty,i}=300/8=37.50\) mg·h/L.
| Virtual individual | \(CL_i\) (L/h) | \(V_i\) (L) | \(C_i(0)\) (mg/L) | \(AUC_{0,\infty,i}\) (mg·h/L) |
|---|---|---|---|---|
| 1 | 4.00 | 30.0 | 10.0 | 75.0 |
| 2 | 5.00 | 40.0 | 7.50 | 60.0 |
| 3 | 8.00 | 50.0 | 6.00 | 37.5 |
In this model, the third individual has the largest clearance and therefore the smallest total AUC at the same dose. The table does not establish a general biological relationship beyond the model assumptions, and it does not show how likely any of these parameter combinations are in a particular population.
Worked Example: A Single-Dose Oral Profile and AUC
Consider a hypothetical one-compartment oral model with first-order absorption, bioavailability \(F=1\), dose \(D=100\) mg, \(CL_i=5\) L/h, \(V_i=20\) L, and absorption rate constant \(k_a=1.00\) h\(^{-1}\). Assume \(k_a\ne k_i\), where \(k_i=CL_i/V_i=0.250\) h\(^{-1}\), and assume linear disposition. The predicted concentration after a single oral dose is
The coefficient is \(100/(20(1.00-0.250))=6.6667\) mg/L. At 1 h, substituting the values gives
At 4 h,
The total AUC can be checked by integrating the two exponential terms:
This example illustrates that a profile equation must match the route and absorption model. Applying the IV bolus equation to an oral dose would omit the absorption process and produce an incorrect profile.
Assumptions, Diagnostics, and Alternatives
The simulated profiles are useful only to the extent that the fitted model and simulation conditions are suitable for the intended question. The principal assumptions are that the structural model adequately represents the relevant concentration-time behavior, the fitted parameterization and random-effects distribution are appropriate for the virtual population, and the regimen and covariates are specified correctly. A simulation that treats fitted population estimates as fixed also leaves parameter-estimation uncertainty out of the result; it describes variability conditional on those estimates.
Several checks can identify errors before exposure results are interpreted:
- Check units and positivity. Clearance, volume, dose, time, and concentration units must be compatible. Exponential parameterizations should give positive individual parameters.
- Check profile shape and limiting behavior. For the IV bolus example, \(C_i(0)=D/V_i\), concentrations decline for \(t>0\), and they approach zero as \(t\) increases. A profile that rises after the bolus or becomes negative signals a coding or model-input problem.
- Check exposure against an analytic result where possible. For a linear IV bolus model, numerical integration over a sufficiently long interval should approach \(D/CL_i\). A large disagreement can indicate an inadequate time grid, a truncated interval, or an implementation error.
- Check that the intended model components are included once. Apply the fitted covariate and random-effects relationships, and do not accidentally apply an effect twice or omit it entirely.
- Separate latent predictions from noisy observations. Adding residual error produces simulated observations around a predicted profile, not a different underlying PK trajectory. It may create negative concentrations under some error models, which is a warning about interpretation rather than a reason to silently alter the model.
A common mistake is to calculate a profile using only the typical parameters and then describe it as a simulation of individual variability. Setting every random effect to zero gives a typical individual at the reference covariates, not a virtual cohort representing interindividual variability. A related error is to add residual error when the question concerns model-predicted exposure: that changes a latent concentration profile into a noisy observation and can alter computed metrics.
The main alternative to individual simulation is to evaluate a typical profile, which is simpler but does not represent the spread of individual exposures. Analytical calculations can be more accurate and efficient for simple models and metrics, such as \(D/CL_i\) in the linear IV bolus example, but are not always available for complex structures or time-varying regimens. Numerical evaluation of the structural model is more general, though it requires checks on time resolution and integration. These approaches answer different questions; none removes the need to state what population, model, regimen, and uncertainty are represented.
Check Your Understanding
Work through each question, then compare your reasoning with the solutions.
- For a 200 mg IV bolus in a one-compartment model, an individual has \(CL_i=4\) L/h and \(V_i=20\) L. What are the initial concentration and total AUC?
- Why does setting all random effects to zero not simulate interindividual variability?
- For the oral model in the worked example, what is \(k_i\), and why would the IV bolus equation be inappropriate?
- What does adding residual error to a model-predicted profile represent?
Solutions. (1) \(C_i(0)=200/20=10.0\) mg/L, and \(AUC_{0,\infty,i}=200/4=50.0\) mg·h/L. (2) Zero random effects assign the typical parameter values at the reference covariates; they produce a typical profile rather than a distribution of individual profiles. (3) \(k_i=CL_i/V_i=5/20=0.250\) h\(^{-1}\). The oral model includes absorption, whereas an IV bolus places the dose into the central compartment immediately. (4) It represents noisy simulated observations around the latent model-predicted concentrations, not additional interindividual PK variability.