Worked Uncertainty Analysis for Hepatic Exposure

Purpose and Scope

Hepatic exposure depends on more than a drug's metabolic capacity. Blood flow determines how quickly drug is delivered to and removed from the liver, while binding affects the unbound concentration available for metabolism. When these inputs are uncertain, their effects can combine in ways that are not obvious from inspecting one parameter at a time. A worked calculation makes the model's assumptions visible and shows which uncertainties matter most for its output.

The calculation here is a simplified, steady-state well-stirred liver model. It estimates the unbound hepatic concentration associated with a specified total drug concentration entering the liver. It is a model-based exposure measure, not a complete PBPK prediction of tissue concentration over time, and it does not by itself establish toxicity or clinical effect.

Definition: Let $Q_h$ be hepatic blood flow (L/h), $f_{u,b}$ the unbound fraction in blood (dimensionless), $CL_{int}$ the intrinsic metabolic clearance (L/h), and $C_{in,b}$ the total drug concentration in blood entering the liver (mg/L). Under the well-stirred, steady-state assumptions below, the unbound hepatic concentration is $$ C_{h,u}=\frac{f_{u,b}Q_h C_{in,b}}{Q_h+f_{u,b}CL_{int}}. $$

Model Assumptions and Derivation

The liver is represented as a single, well-mixed compartment. Blood flow and the drug concentration entering the liver are constant over the calculation; binding is rapid and represented by a constant $f_{u,b}$; and metabolism is linear in the unbound hepatic concentration, with rate $CL_{int}C_{h,u}$. The model ignores spatial concentration gradients, saturable metabolism, time-varying physiology, and other elimination or transport processes. The parameters and inlet concentration must use compatible units and refer to the same drug and physiological context.

Let $C_{h,b}$ denote the total drug concentration in the well-stirred liver blood. The unbound concentration is $C_{h,u}=f_{u,b}C_{h,b}$. At steady state, the rate entering the liver equals the rate leaving in blood plus the metabolic rate: $$ Q_h C_{in,b}=Q_h C_{h,b}+CL_{int}C_{h,u}. $$ Substitute $C_{h,b}=C_{h,u}/f_{u,b}$ into the balance: $$ Q_h C_{in,b}=\frac{Q_h}{f_{u,b}}C_{h,u}+CL_{int}C_{h,u}. $$ Multiplying by $f_{u,b}$ and solving for $C_{h,u}$ gives the definition above. The denominator has units L/h, and the numerator has units $(\mathrm{L/h})(\mathrm{mg/L})=\mathrm{mg/h}$, so the result is in mg/L.

The equation can also be written $C_{h,u}=f_{u,b}F_hC_{in,b}$, where $F_h=Q_h/(Q_h+f_{u,b}CL_{int})$ is the fraction of drug escaping hepatic extraction in this model. This decomposition is useful for interpretation: unbound concentration rises with $f_{u,b}$, but the same increase in unbound fraction also increases the metabolic term in the denominator. The net effect is determined by the full expression, not by binding alone.

Worked Example: Calculate the Nominal Hepatic Concentration

Suppose a hypothetical drug has $Q_h=90.0$ L/h, $f_{u,b}=0.120$, $CL_{int}=400$ L/h, and $C_{in,b}=0.230$ mg/L. These values are illustrative and are not measurements from a real study. Calculate $C_{h,u}$.

First calculate the binding-adjusted intrinsic clearance: $f_{u,b}CL_{int}=0.120(400)=48.0$ L/h. The denominator is $90.0+48.0=138$ L/h, and the numerator is $0.120(90.0)(0.230)=2.484$ mg/h. Therefore,

C_{h,u}=\frac{2.484\ \mathrm{mg/h}}{138\ \mathrm{L/h}} =0.0180\ \mathrm{mg/L}.

This is the deterministic result at the stated input values. It does not express uncertainty, and it is conditional on the model assumptions and the specified inlet concentration.

Local Sensitivity and Uncertainty Propagation

A useful first step is to ask how a small proportional change in each input changes the calculated concentration. Define $D=Q_h+f_{u,b}CL_{int}$. The elasticities, or derivatives of $\log C_{h,u}$ with respect to the log inputs, are

\frac{\partial\log C_{h,u}}{\partial\log Q_h} =\frac{f_{u,b}CL_{int}}{D}=E_h,\qquad \frac{\partial\log C_{h,u}}{\partial\log f_{u,b}} =\frac{Q_h}{D}=F_h,\qquad \frac{\partial\log C_{h,u}}{\partial\log CL_{int}} =-\frac{f_{u,b}CL_{int}}{D}=-E_h.

Here $E_h$ is the model's hepatic extraction fraction and $F_h=1-E_h$ is the hepatic escape fraction. The signs and magnitudes follow directly from differentiating the log of the exposure equation. For example, increasing $CL_{int}$ lowers $C_{h,u}$, while increasing $Q_h$ raises it. The local sensitivity to binding is positive, but its elasticity is $F_h$, not 1, because binding also affects extraction.

For uncertain inputs, a first-order approximation in log space combines these sensitivities with the covariance matrix of the log inputs. If $a=(E_h,F_h,-E_h)$ is the vector of elasticities and $\Sigma$ is the covariance matrix of $(\log Q_h,\log f_{u,b},\log CL_{int})$, then the approximate variance of $\log C_{h,u}$ is $a^\mathsf{T}\Sigma a$. This includes covariance terms when inputs are dependent. With independent inputs, it reduces to the sum of each squared elasticity times the corresponding log-scale variance.

Worked Example: Approximate an Uncertainty Interval

Continue with the nominal inputs in the previous example, treating them as the medians of hypothetical lognormal input distributions. Suppose the standard deviations of $\log Q_h$, $\log f_{u,b}$, and $\log CL_{int}$ are respectively 0.150, 0.300, and 0.400, and suppose the three inputs are independent. For $f_{u,b}$, the lognormal assumption is only a local approximation; its unbounded upper tail is not a fully physically admissible input distribution. These are assumed uncertainty values for illustration.

At the nominal inputs, $D=138$ L/h, so $E_h=48.0/138=0.3478$ and $F_h=90.0/138=0.6522$. Independence gives the following first-order log-scale variance:

\begin{aligned} \operatorname{Var}(\log C_{h,u}) &\approx (0.3478)^2(0.150)^2 +(0.6522)^2(0.300)^2 +(-0.3478)^2(0.400)^2\\ &=0.002722+0.03828+0.01936\\ &=0.06036. \end{aligned}

The approximate standard deviation of $\log C_{h,u}$ is $\sqrt{0.06036}=0.2457$. Using a normal approximation on the log scale, a central 95% uncertainty interval around the nominal value is calculated as $0.0180\exp(\pm1.96(0.2457))$ mg/L. Since $1.96(0.2457)=0.4816$, the lower and upper multipliers are $\exp(-0.4816)\approx0.6178$ and $\exp(0.4816)\approx1.619$. Thus the approximate interval is $0.0111$ to $0.0291$ mg/L.

This is an approximate uncertainty interval conditional on the specified input distributions, independence assumption, model, and fixed inlet concentration. It is not a confidence interval for an estimated population parameter. The contributions to log variance are about 4.51% from flow, 63.42% from binding, and 32.07% from intrinsic clearance. Under these assumptions, uncertainty in binding contributes most to the output variance.

Dependence and the Role of Assumptions

Input dependence can materially change the output range. A covariance term between log binding and log metabolic capacity is $2F_h(-E_h)\operatorname{Cov}(\log f_{u,b},\log CL_{int})$. Because their output sensitivities have opposite signs, negative dependence between these inputs makes this term positive: parameter sets with higher binding fraction tend also to have lower metabolic capacity, reinforcing increases in the calculated exposure. The direction and strength of dependence must be justified; it should not be chosen merely to obtain a wider or narrower range.

Worked Example: Include a Binding–Capacity Correlation

Use the same nominal values and marginal log-scale standard deviations as above. Now assume the hypothetical correlation between $\log f_{u,b}$ and $\log CL_{int}$ is $-0.600$, with flow still independent of both. The covariance is $(-0.600)(0.300)(0.400)=-0.0720$. The added variance contribution is

$$ 2(0.6522)(-0.3478)(-0.0720)=0.0327. $$

Using the unrounded elasticities, the total approximate log variance is $0.0930246$, giving a log-scale standard deviation of $0.3050$. The 95% log-scale half-width is $1.96(0.3050)=0.5978$. Therefore the approximate interval is $0.0180\exp(\pm0.5978)$ mg/L, or $0.00990$ to $0.0327$ mg/L. The interval is wider than under independence because the assumed negative correlation aligns higher $f_{u,b}$ with lower $CL_{int}$, both of which raise the modeled concentration. This calculation is valid only if that dependence assumption is credible.

Diagnostics, Alternatives, and Limitations

The derivative-based approximation is most informative when uncertainty is not so broad that the local sensitivities change substantially across plausible inputs, and when the response is not dominated by sharp nonlinearities or boundaries. Check whether sampled or bounded inputs stay physically admissible: $Q_h>0$, $CL_{int}\geq0$, and $0<f_{u,b}\leq1$. Inspect the output distribution for skewness, extreme tails, or influential combinations. If parameter dependence is used, check that the joint distribution is coherent and that its simulated parameter sets respect the stated bounds.

A simple monotonicity check follows from the model: $C_{h,u}$ increases with $Q_h$ and $f_{u,b}$, and decreases with $CL_{int}$, for positive inputs. If a numerical implementation shows the opposite direction under otherwise fixed conditions, verify the formula, units, and parameter mapping. Also check that $C_{in,b}$ is intentionally held fixed. If inlet concentration is uncertain or depends on the same parameters, its uncertainty and dependence must be propagated too.

For broad or strongly nonlinear uncertainty, sampling parameter sets and evaluating the exposure equation for each set is a more direct alternative to the local approximation. Summaries from sampling can reveal asymmetry and nonlinear interactions, provided the input distributions and dependence structure are specified appropriately. A bounded corner analysis is another useful alternative when the goal is to find the minimum and maximum over stated ranges rather than to assign probabilities. Neither method removes uncertainty about the assumptions or the input distributions.

Worked Example: Compare a Bounded Range With a Probability Interval

Suppose a separate hypothetical plausibility exercise specifies $Q_h$ between 70 and 110 L/h, $f_{u,b}$ between 0.060 and 0.180, and $CL_{int}$ between 250 and 600 L/h, while holding $C_{in,b}=0.230$ mg/L fixed. By the monotonicity results, the minimum occurs at the lowest flow, lowest unbound fraction, and highest intrinsic clearance:

C_{\min}=\frac{(0.060)(70)(0.230)}{70+(0.060)(600)} =\frac{0.966}{106}=0.00911\ \mathrm{mg/L}.

The maximum occurs at the highest flow, highest unbound fraction, and lowest intrinsic clearance:

C_{\max}=\frac{(0.180)(110)(0.230)}{110+(0.180)(250)} =\frac{4.554}{155}=0.0294\ \mathrm{mg/L}.

Thus these chosen bounds imply a bounded range of approximately 0.00911–0.0294 mg/L. It is not a 95% interval: the bounds alone do not specify probabilities within the range. Its comparison with the earlier probability-based approximation is meaningful only with care, because the input ranges and probability distributions were separately assumed.

The model's value is that it exposes the joint roles of delivery, binding, and metabolism in a transparent calculation. Its limitations are equally important. Real hepatic disposition may involve uptake and efflux transport, multiple cell populations, nonuniform perfusion, saturable metabolism, changing concentrations, or differences between blood and plasma measurements. If those processes are material, the well-stirred calculation is a screening approximation rather than an adequate representation of hepatic exposure. A narrow output range is not evidence that the model is correct; it may simply reflect narrow input assumptions or omitted mechanisms.

Common Pitfall

A common mistake is to vary one parameter at a time and treat the resulting changes as the full uncertainty range. That reasoning misses joint changes and dependence among inputs. Another mistake is to interpret a bounded range as though every value in it were equally likely, or to call a model-based uncertainty interval a confidence interval. The correction is to state exactly what was varied, how distributions or bounds were assigned, whether inputs were dependent, and what interpretation the resulting range supports.

Key takeaway: A hepatic exposure estimate is conditional on both a model and its inputs. Derive the concentration from a mass balance, propagate uncertainty using sensitivities or joint sampling, and report the assumptions that determine the resulting range.

Check Your Understanding

Answer each question, then compare with the solutions below.

For fixed positive $Q_h$, $f_{u,b}$, and $CL_{int}$, what is the direction of change in $C_{h,u}$ when $CL_{int}$ increases?
In the nominal calculation, which input contributed the largest share of the approximate log variance?
If $f_{u,b}$ and $CL_{int}$ are negatively correlated on the log scale, does their covariance term increase or decrease the output variance here?
What additional information is needed to interpret a bounded input range as a probability-based uncertainty interval?

Solutions. (1) It decreases: the denominator $Q_h+f_{u,b}CL_{int}$ increases while the numerator is unchanged. (2) The unbound fraction, contributing about 63.42% under the stated independent-input assumptions. (3) It increases the output variance because the sensitivities have opposite signs, so negative covariance yields a positive variance contribution. (4) A probability distribution for inputs, including their dependence structure, is needed; bounds alone do not assign probabilities to values within the range.