Sampling for Michaelis-Menten Elimination

Why Michaelis–Menten Elimination Changes Sampling Priorities

With linear elimination, the elimination rate is proportional to concentration, so the concentration trajectory has a constant fractional decline. Michaelis–Menten elimination is different: the rate approaches a maximum at high concentration and becomes approximately proportional to concentration at low concentration. A schedule that observes only one of these regions may estimate a combination of parameters well while leaving the parameters themselves difficult to distinguish.

The design question here is specific: at which times are concentrations sensitive to elimination capacity and affinity in distinguishable ways? The answer depends on the nominal parameter values, the concentration scale of the assay error, and the feasible sampling window. A high concentration can primarily inform capacity, while observations near the transition between saturated and unsaturated elimination can add affinity information. Neither statement guarantees that a particular time is optimal for every patient or design.

Definition: For a one-compartment model after an IV bolus, with no further input, let $C(t)$ be concentration at time $t$, $C_0$ the known concentration at time zero, $V_{\max}$ the maximum elimination rate expressed in concentration per unit time, and $K_m$ the concentration at which the elimination rate is half its maximum. Michaelis–Menten elimination is described by $$ \frac{dC}{dt}=-\frac{V_{\max}C}{K_m+C}, \qquad C(0)=C_0. $$ Here $V_{\max}>0$, $K_m>0$, and $C_0>0$.

Trajectory and Sensitivity Measures

The model assumes a single, well-mixed compartment; known initial concentration; no absorption, infusion, or other input after time zero; and elimination governed by the stated rate law. It is a structural model, not a guarantee that real concentrations follow that law. Sampling times are treated as fixed, and the local calculations below evaluate sensitivities at specified nominal values of $V_{\max}$ and $K_m$.

Separating variables and integrating from $C_0$ to $C(t)$ gives an implicit concentration-time relationship. This form is convenient because it avoids requiring a closed-form expression for $C(t)$.

\int_{C_0}^{C(t)}\left(1+\frac{K_m}{C}\right)dC =-V_{\max}t \quad\Longrightarrow\quad C-C_0+K_m\ln\!\left(\frac{C}{C_0}\right)+V_{\max}t=0.

For a positive elimination rate, concentration decreases from $C_0$, so $0<C(t)\leq C_0$. To compare information about parameters on a relative scale, define $\eta_V=\ln V_{\max}$ and $\eta_K=\ln K_m$. The sensitivities of log concentration are derivatives with respect to these log parameters, at a fixed time. They quantify the local change in predicted log concentration for a small relative change in a parameter.

g_V(t)=\frac{\partial\ln C(t)}{\partial\eta_V}, \qquad g_K(t)=\frac{\partial\ln C(t)}{\partial\eta_K}.

Differentiate the implicit trajectory equation $F(C,V_{\max},K_m,t)=0$. At fixed $t$ and $C_0$, its concentration derivative is $F_C=1+K_m/C$. Also, $F_{\eta_V}=V_{\max}t$ and $F_{\eta_K}=K_m\ln(C/C_0)$. Implicit differentiation therefore gives the following sensitivity formulas.

\frac{\partial C}{\partial\eta_V} =-\frac{V_{\max}t\,C}{C+K_m}, \qquad \frac{\partial C}{\partial\eta_K} =-\frac{K_m C\ln(C/C_0)}{C+K_m}, $$ $$ g_V(t)=-\frac{V_{\max}t}{C+K_m}, \qquad g_K(t)=-\frac{K_m\ln(C/C_0)}{C+K_m}.

For $t>0$ and $C<C_0$, $g_V$ is negative: increasing capacity lowers predicted concentration. In contrast, $g_K$ is positive: increasing $K_m$ slows elimination at a given concentration and raises the predicted concentration at a fixed time. At $t=0$, both sensitivities are zero under the assumption that $C_0$ is known and fixed. If the response is measured on the concentration scale rather than the log scale, use $\partial C/\partial\eta_V$ and $\partial C/\partial\eta_K$, not the log-concentration sensitivities.

From Sensitivity to Local Information

The error model determines how sensitivities translate into information. For the examples below, suppose the measured log concentration at time $t_i$ is $Z_i=\ln C(t_i)+\epsilon_i$, where the errors are independent and normally distributed with mean zero and known variances $\sigma_i^2$. Under this model, the Fisher information for $(\eta_V,\eta_K)$ is the sum of sensitivity outer products weighted by inverse error variance. This is an exact Fisher information expression for the stated normal model, not a general formula for arbitrary independent errors.

\mathcal I =\sum_i\frac{1}{\sigma_i^2} \begin{pmatrix}g_V(t_i)\\g_K(t_i)\end{pmatrix} \begin{pmatrix}g_V(t_i)&g_K(t_i)\end{pmatrix}.

The formula follows from the normal log likelihood. For one observation with mean $\mu_i=\ln C(t_i)$, the score is $(Z_i-\mu_i)\nabla\mu_i/\sigma_i^2$. Its expected outer product is $\nabla\mu_i\nabla\mu_i^\mathsf{T}/\sigma_i^2$, because $E[(Z_i-\mu_i)^2]=\sigma_i^2$. Independence makes the information contributions add. If the variances depend on parameters, the normal-model Fisher information generally has additional terms. For nonnormal errors, the exact Fisher information also depends on the specified error distribution; a weighted least-squares or Gauss–Newton matrix may be used as an approximation, but must be identified as such.

Worked Examples of Capacity and Affinity Information

Worked Example: A High-Concentration Sample Primarily Informs Capacity

Consider a hypothetical individual with nominal $C_0=20$ mg/L, $V_{\max}=2$ mg/L/h, and $K_m=4$ mg/L. Suppose a candidate sample is taken when the concentration is 16 mg/L. Find its time and log-concentration sensitivities.

From the integrated model, $t=(C_0-C+K_m\ln(C_0/C))/V_{\max}$. Substitution gives $t=(20-16+4\ln(20/16))/2=(4+0.8926)/2=2.446$ h.

The sensitivity to log capacity is $g_V=-V_{\max}t/(C+K_m)=-(2)(2.446)/(16+4)=-0.2446$. The sensitivity to log affinity is $g_K=-K_m\ln(C/C_0)/(C+K_m)=-4\ln(16/20)/20=0.04463$.

The magnitude of the capacity sensitivity is appreciably larger than the affinity sensitivity at this time. Thus, locally, this sample responds more strongly to a relative change in $V_{\max}$ than to the same relative change in $K_m$. It is not wholly uninformative about $K_m$, and this comparison alone does not establish that it is the best capacity sample: assay precision and the other sampling times matter.

Worked Example: A Transition-Region Sample Adds a Different Direction

Use the same hypothetical parameter values, but consider a sample when $C=K_m=4$ mg/L. Calculate its time and sensitivities, then compare its sensitivity direction with the previous sample.

The sampling time is $t=(20-4+4\ln(20/4))/2=(16+6.4378)/2=11.22$ h. At this concentration, $g_V=-(2)(11.22)/(4+4)=-2.805$, while $g_K=-4\ln(4/20)/8=0.8047$.

Both sensitivities are larger in magnitude than at 16 mg/L, and their relative balance differs: the affinity sensitivity is now more substantial. The two samples therefore provide different local sensitivity directions. The $C=4$ mg/L sample alone still supplies only one sensitivity vector for two parameters, so it cannot by itself separate both parameters. Its value is in complementing other times, not in being a complete design on its own.

Worked Example: Two Times Can Separate the Parameters Better Than Repeating One

For the same individual, compare one log-concentration observation at $C=4$ mg/L and one at $C=1$ mg/L with two observations both at $C=4$ mg/L. Suppose both observations have known log-scale standard deviation $\sigma=0.05$. For the $C=1$ mg/L sample, the model gives $t=(20-1+4\ln 20)/2=(19+11.9829)/2=15.49$ h, with $g_V=-(2)(15.49)/(1+4)=-6.197$ and $g_K=-4\ln(1/20)/5=2.397$.

With one observation at each concentration, the sensitivity matrix, with rows ordered by sample time, is approximately $$ S= \begin{pmatrix} -2.805&0.8047\\ -6.197&2.397 \end{pmatrix}. $$ Using the displayed rounded entries, its determinant is $(-2.805)(2.397)-(0.8047)(-6.197)\approx-1.737$; using the unrounded sensitivities gives approximately $-1.73524$, which is nonzero. With equal variance, $\det(\mathcal I)=\det(S)^2/\sigma^4$, so $\det(\mathcal I)\approx(1.73524)^2/(0.05)^4=4.818\times10^5$.

If both observations are instead taken at $C=4$ mg/L, the two rows of $S$ are identical, so $\det(S)=0$ and the two-parameter information matrix is singular. Repeating a time can improve precision in the direction it informs, but identical sensitivity directions do not supply a second independent direction. In this numerical comparison the two distinct times separate the parameters locally; actual design decisions must also account for feasible times, assay properties, and model uncertainty.

Reading the Concentration Regions

The model explains why a spread of concentrations can be useful. When $C$ is much larger than $K_m$, the elimination rate $V_{\max}C/(K_m+C)$ is approximately $V_{\max}$. This near-saturated region tends to emphasize capacity. When $C$ is much smaller than $K_m$, the rate is approximately $(V_{\max}/K_m)C$; the trajectory then depends strongly on the ratio $V_{\max}/K_m$, which can make the two parameters difficult to distinguish. Observations around $C$ comparable to $K_m$ can help reveal the transition between these regimes and add a different sensitivity direction.

These are qualitative guides, not rules that a sample must be taken at exactly $C=K_m$. Sensitivity magnitudes and directions change with time, nominal parameter values, and the response scale. A low-concentration observation can have a large log-concentration sensitivity, as in the example, yet still be weak in practice if its assay error is large or its sensitivity vector is nearly aligned with those from other times. Conversely, a high-concentration sample may have a clear capacity-dominant direction but small absolute sensitivity if taken too close to time zero.

Specify the model and nominal values.
State the bolus or input assumptions, whether $C_0$ is known, the parameterization, and the nominal $V_{\max}$ and $K_m$ used to evaluate the trajectory.

Evaluate candidate concentration-time sensitivities.
Calculate $C(t_i)$ and the sensitivity vector for each feasible time. Match the derivative scale to the modeled observation scale.

Check precision and directional complementarity.
Weight by the specified error model and inspect whether the candidate vectors are large enough and sufficiently distinct to inform both parameters.

Check the assumptions that drive the result.
Re-evaluate the proposed schedule when plausible nominal values, assay precision, initial concentration, or feasible time windows change.

Diagnostics, Alternatives, and Limitations

A useful diagnostic is to inspect both sensitivity magnitude and direction. A near-zero vector contributes little local information. Several large but nearly parallel vectors can still leave one parameter combination weakly informed; a nearly singular information matrix or very high estimated parameter correlation signals this problem. In practice, calculate sensitivities across the feasible time window, not only at a few appealing landmarks, and check whether the preferred schedule changes under plausible parameter values.

Peak, trough, and mid-interval sampling provides a simple time-based description of a concentration profile. For Michaelis–Menten elimination, however, the relationship between those landmarks and the capacity-affinity transition depends on the individual parameters. Sensitivity-based candidate evaluation directly asks which parameter changes a sample can detect under the assumed model and error scale. It does not remove the need for operational constraints, and it does not guarantee robustness to misspecification. An exact individual design evaluates a specified finite set of times; a continuous or numerical search can explore more candidate times, but neither makes the local nominal values certain.

The most common mistake is to conclude that “a sample at high concentration estimates $V_{\max}$, and a low-concentration sample estimates $K_m$.” The first part is a useful tendency, but at low concentration the trajectory often reflects $V_{\max}/K_m$, not $K_m$ alone. The correction is to compare both sensitivity columns together, using the actual error scale, and determine whether the set of observations separates the parameters. A second pitfall is to treat the displayed normal-error information formula as universal: it requires the stated normal, independent, known-variance model. Different error distributions or parameter-dependent variance require their own information calculation.

Finally, these calculations are local. They assume the structural model and nominal values are adequate and, in the derivation above, that $C_0$ is known. Uncertainty in initial concentration, volume, dose delivery, or other model components can alter the sensitivities and create additional confounding. The next practical step is therefore not to regard a calculated schedule as universally optimal, but to check its behavior under scientifically plausible alternatives.

Key takeaway: High-concentration observations often emphasize Michaelis–Menten capacity, while observations around the transition can contribute affinity information. Select times by evaluating the parameter sensitivity vectors together with the assay error model; concentration labels alone do not establish what a schedule identifies.

Check Your Understanding

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

For $C<C_0$, what are the signs of $g_V$ and $g_K$, and what do those signs mean?
Using $C_0=20$ mg/L, $V_{\max}=2$ mg/L/h, and $K_m=4$ mg/L, calculate the time at which $C=4$ mg/L.
Why can two samples at the same time fail to identify both log parameters, even if each assay is precise?
What feature of the observation model justifies the stated Fisher information formula?

Solutions. (1) $g_V<0$, because increasing capacity lowers concentration; $g_K>0$, because increasing affinity parameter $K_m$ slows elimination at fixed concentration and raises concentration at fixed time. (2) $t=(20-4+4\ln 5)/2=(16+6.4378)/2=11.22$ h. (3) Repeated samples at the same time have proportional, indeed identical, sensitivity vectors, so their information matrix has rank at most one for two parameters. Precision can increase along that one direction but cannot create a second direction. (4) The errors are independent normal with known, parameter-independent variances, and the mean is the model-predicted log concentration; under those assumptions the Fisher information is the sum of weighted sensitivity outer products.