Why Sampling Time Determines What Can Be Learned
A study can collect many concentration and response measurements yet still provide little evidence about whether the parent drug or an active metabolite drives the response. The difficulty arises when their concentrations move together: at every sampled time, a high parent concentration accompanies a high metabolite concentration, or both are low. Their separate contributions then make similar predictions. Sampling should therefore target times at which the parent and metabolite exposures have meaningfully different patterns, while still measuring response precisely enough to observe the resulting effect.
The design problem has two linked parts. Concentration samples describe the time courses of the parent and metabolite; response samples connect those exposures to the endpoint. The timing should be informed by formation and elimination kinetics, not chosen only for convenience or by distributing samples uniformly across the dosing interval.
This model is a design device, not a claim that every endpoint follows a linear concentration–response relationship over its full range. It makes the central sampling issue explicit: if \(C_M(t)=aC_P(t)\) at all sampled times for some constant \(a\), then the response data identify only the combined contribution \(\beta_P+a\beta_M\), not the two coefficients separately. The sampling plan should seek time points where this proportionality is broken.
Use Formation and Elimination to Locate Informative Windows
After a dose, parent concentration often declines while an active metabolite forms and is subsequently eliminated. The metabolite may rise after the parent concentration has begun to fall, creating a period when their relative contributions differ. If the metabolite forms and disappears quickly, this period may be brief and require dense sampling. If its elimination is slow, later samples may be particularly useful for observing metabolite exposure after the parent has declined.
A simple one-compartment illustration makes the timing logic concrete. Suppose parent concentration declines as \(C_P(t)=C_0e^{-k_Pt}\), and metabolite concentration is proportional to the difference between parent disappearance and metabolite elimination: $$ C_M(t)=A\left(e^{-k_Pt}-e^{-k_Mt}\right), \qquad k_M>k_P,\quad A>0. $$ This idealized profile starts at zero and rises before declining. Differentiating and setting the derivative to zero gives its peak time: $$ \frac{dC_M(t)}{dt}=A\left(-k_Pe^{-k_Pt}+k_Me^{-k_Mt}\right)=0, \qquad t_{\mathrm{peak}}=\frac{\ln(k_M/k_P)}{k_M-k_P}. $$ The peak is a useful candidate sampling window, not automatically the optimal time. Sampling only at the metabolite peak may not distinguish the two effects if parent concentration is also high there. The useful comparison is the pair of concentrations at that time and at other sampled times.
Worked Example: Locating a Metabolite-Rich Sampling Window
Suppose a hypothetical parent concentration declines with \(k_P=0.20\ \mathrm{h}^{-1}\), while the metabolite is eliminated with \(k_M=0.80\ \mathrm{h}^{-1}\). Under the profile above, when is metabolite concentration expected to peak?
Substitute the rates into the peak-time formula: $$ t_{\mathrm{peak}}=\frac{\ln(0.80/0.20)}{0.80-0.20} =\frac{\ln(4)}{0.60} =\frac{1.3863}{0.60} =2.311\ \mathrm{h}. $$ A concentration sample near 2.31 hours would help characterize the metabolite peak. To assess separation, the schedule should also include a time when the parent-to-metabolite concentration ratio differs—for example, an earlier sample during metabolite formation or a later sample after the parent has declined. The rates justify investigating these windows, but actual concentration profiles should guide final selection.
Quantifying Separation With the Design Matrix
For \(n\) response observations at times \(t_1,\ldots,t_n\), form the design matrix \(X\) with row \(i\) equal to \([1,C_P(t_i),C_M(t_i)]\) when estimating a baseline as well as both concentration coefficients. If baseline is known or removed by design, the two exposure columns alone can be examined. Under independent errors with common variance \(\sigma^2\), the information matrix for the linear model is proportional to \(X^\mathsf{T}X/\sigma^2\). A full-rank exposure design is necessary to estimate both coefficients separately; near dependence produces large uncertainty even when the matrix is technically full rank.
For two response times and a known baseline, let the exposure-only rows be \([C_P(t_1),C_M(t_1)]\) and \([C_P(t_2),C_M(t_2)]\). Their determinant is $$ \Delta=C_P(t_1)C_M(t_2)-C_P(t_2)C_M(t_1). $$ If \(\Delta=0\), the two rows are proportional and cannot distinguish the coefficients. A larger \(|\Delta|\), all else equal, indicates more two-parameter information. With more times, \(\det(X^\mathsf{T}X)\) is a common measure of overall information for the exposure coefficients, subject to how baseline and other model terms are handled. It does not replace checks of parameter-specific precision, model assumptions, or practical feasibility.
Worked Example: Comparing Two Hypothetical Response Schedules
Consider two candidate schedules with two response times each. The entries below are hypothetical model-predicted concentrations in mg/L; values are rounded for display. Assume a known baseline and independent response errors with equal variance at the two times.
| Schedule | Time point | Parent \(C_P\) (mg/L) | Metabolite \(C_M\) (mg/L) |
|---|---|---|---|
| Early | 1 | 6.065 | 4.508 |
| Early | 2 | 3.679 | 2.983 |
| Spread | 1 | 6.065 | 4.310 |
| Spread | 2 | 0.1832 | 2.983 |
For the early schedule, the design-matrix determinant is the cross-product difference: $$ \Delta_{\mathrm{early}} =(6.065)(2.983)-(3.679)(4.508) =18.091895-16.584932 =1.506963. $$ Thus \(\Delta_{\mathrm{early}}^2\approx2.271\). For the spread schedule: $$ \Delta_{\mathrm{spread}} =(6.065)(2.983)-(0.1832)(4.310) =18.091895-0.789592 =17.302303, \qquad \Delta_{\mathrm{spread}}^2\approx299.4. $$ These calculations use the displayed concentrations, so the results are approximate. The spread schedule has substantially greater determinant-based separation in this two-time comparison: one observation has low parent concentration while metabolite concentration remains appreciable. This supports choosing the spread schedule if other design considerations are comparable. It does not guarantee unbiased or precise estimates under model misspecification, unequal errors, or additional nuisance parameters.
A determinant calculation depends on the way the concentration columns are arranged. For example, \(C_P(t_1)C_M(t_1)-C_P(t_2)C_M(t_2)\) is a difference of within-time products, not the two-row design-matrix determinant. Using unrounded model predictions can matter for arithmetic, too. As a numerical check, \(100(e^{-0.7}-e^{-1.1})=16.371422\), whose square is approximately \(268.02\); and \(100(e^{-1.3}-e^{-4.1})\approx25.5959\), whose square is approximately \(655.15\). Those calculations correctly square those particular contrasts, but they are not information determinants unless the contrasts have first been shown to equal the relevant design-matrix determinant. For actual design comparison, use the cross-products in \(\Delta\) or the appropriate information matrix.
Coordinate Concentration and Response Sampling
Concentration and response times need not be identical, but they must be coordinated. Concentration samples near expected changes in profile shape can estimate the exposure trajectories; response samples should cover the times when the parent-to-metabolite balance differs and when the endpoint is expected to reflect those exposures. If response has a delay, a response observed at time \(t\) may relate to earlier concentrations rather than only \(C_P(t)\) and \(C_M(t)\). The sampling plan should reflect the response model being fitted and collect enough concentration data to support its exposure predictions.
Worked Example: Choosing Times Around Different Exposure Patterns
Suppose preliminary PK information for a hypothetical compound suggests that parent concentration is highest soon after dosing, metabolite concentration peaks later, and the metabolite persists after parent concentrations fall. The study can collect four response measurements and has capacity for additional concentration samples. Propose a schedule and explain its purpose.
A defensible initial plan is to measure concentrations at an early post-dose time, near the predicted metabolite peak, and at a later time when the parent is low but the metabolite remains measurable. Place response measurements at or near those exposure windows, with a baseline measurement if baseline is not known. The early sample anchors parent-dominant exposure; the peak-region sample characterizes the transition; and the late sample provides the strongest opportunity to observe metabolite exposure with limited parent contribution.
This plan should be checked against predicted concentrations: if both compounds are near zero at the late time, that observation adds little separation; if the predicted response is delayed, response sampling may need to follow the concentration changes. Replicate measurements can improve precision but do not substitute for variation in the parent-to-metabolite exposure ratio. Before fixing the schedule, compare the proposed design matrix with plausible alternatives and examine the resulting uncertainty under the expected response error.
Assumptions, Diagnostics, and Alternatives
The simple information comparison assumes the candidate concentrations are known or predicted accurately, the response model is an adequate local description, and the error structure used in the calculation is appropriate. The equal-variance linear model also assumes independent errors with common variance. If response variance differs across time, weighted information should be used; if observations are correlated, the covariance structure must be included. If baseline, time trends, or other effects are estimated, assess separation after including those terms rather than judging the two exposure columns in isolation.
Useful diagnostics include plotting \(C_P(t)\) and \(C_M(t)\) against time, plotting one concentration against the other at the planned times, and inspecting the design matrix for rank or near-collinearity. After data collection, large coefficient standard errors, unstable estimates under small changes to the sampling set, or strong coefficient correlation can signal weak separation. A very large determinant alone is not sufficient if one coefficient remains poorly estimated, if the response model is wrong, or if concentrations are imprecise.
Common alternatives and complements address different priorities:
- Uniform or routine clinical sampling is simple and may support general exposure description, but can miss a short metabolite-dominant window.
- Targeted sampling around kinetic landmarks uses expected formation, peak, and elimination behavior to improve separation, but depends on credible prior PK information.
- Information-based optimization compares candidate schedules using a determinant, parameter variances, or another stated criterion. It can formalize trade-offs but is only as reliable as its assumed model and error structure.
- More sampling times or replicates can improve estimation, yet additional observations at nearly redundant exposure ratios may add less information than a smaller number of well-separated times.
The design is limited by dose timing, assay sensitivity, participant burden, endpoint measurement frequency, and uncertainty in formation or elimination rates. A schedule optimized for one predicted kinetic profile may be less informative if the actual metabolite peak is shifted. Where feasible, allow enough flexibility to refine timing from emerging concentration data, and preserve samples around distinct exposure phases rather than spending the entire sampling budget in one narrow window.
Check Your Understanding
Answer each question using the design principles above.
- If metabolite concentration is proportional to parent concentration at every response time, can the two additive response coefficients be estimated separately? Why?
- For two times, identify the correct determinant for exposure rows \([C_P(t_1),C_M(t_1)]\) and \([C_P(t_2),C_M(t_2)]\).
- Why might a sample at the metabolite peak fail to provide strong parent–metabolite separation?
- A proposed late sample has both concentrations below assay quantification limits. What is the design concern?
Solutions. (1) No. Proportional exposure columns make the design rank deficient, so only a combined contribution is identifiable. (2) \(\Delta=C_P(t_1)C_M(t_2)-C_P(t_2)C_M(t_1)\); the cross-products are essential. (3) Parent concentration may also be high at the metabolite peak, leaving the exposure ratio similar to that at other times. (4) The sample may add little usable information because neither concentration is reliably quantified; a different time with measurable and more distinct exposures is preferable.