Why fit both inhibitory mechanisms?
An inhibitory drug effect can reduce the response by decreasing its production, or it can reduce the response’s loss. The same word—inhibition—therefore describes changes to different parts of the response balance. A useful identification exercise fits both mechanisms to the same baseline and post-dose data, checks whether their predictions accord with the observed response direction and time course, and asks which parameters the data actually constrain.
The examples below use a hypothetical biomarker measured in response units. Before dosing, assume it is at baseline steady state. During the post-dose interval, assume the drug effect is constant, summarized by a fractional inhibition \(I\). At washout, assume the effect returns immediately to zero. These assumptions isolate the response-model comparison; a real analysis may need to account for changing drug concentrations, uncertain baseline, delayed washout, or measurement error.
Equations and constant-effect solutions
For production inhibition, the response balance during dosing is \(dR/dt=k_{\mathrm{in}}(1-I)-k_{\mathrm{out}}R\), with initial response \(R(0)=R_0\). Solving this first-order linear equation gives:
For loss inhibition, the balance is \(dR/dt=k_{\mathrm{in}}-k_{\mathrm{out}}(1-I)R\), also with \(R(0)=R_0\). Its solution is:
The production-inhibition solution moves downward from baseline toward \(R_0(1-I)\); the loss-inhibition solution moves upward toward \(R_0/(1-I)\). Both begin at \(R_0\). The directions follow directly from the balances: at \(R_0\), production inhibition gives a negative initial rate, while loss inhibition gives a positive initial rate when \(I>0\). The distinction is useful, but it is not by itself proof of mechanism: noisy measurements, a changing drug effect, or baseline drift can complicate the observed trajectory.
Worked fit to a hypothetical post-dose profile
Suppose a hypothetical biomarker has baseline \(R_0=100\) response units. Assume this baseline is known precisely for the exercise. Three measurements during a constant-effect interval are shown below. The values are hypothetical and rounded to three decimal places.
| Time after dose (h) | Response (units) |
|---|---|
| 2 | 114.225 |
| 6 | 134.217 |
| 12 | 150.871 |
The response rises above baseline. The loss-inhibition solution can describe that direction with \(I>0\); the production-inhibition solution cannot. The following fits make the comparison explicit.
Worked Example: Fitting the Loss-Inhibition Model
The observations are consistent, to rounding, with \(I=0.40\) and \(k_{\mathrm{out}}=0.20\ \mathrm{h}^{-1}\). Substituting these values into the loss-inhibition solution gives, at 2 h:
At 6 h, the exponent is \(-(0.20)(0.60)(6)=-0.72\), so \(R_{\mathrm{loss}}(6)=100[1+(0.40/0.60)(1-e^{-0.72})]=134.217\) units. At 12 h, \(R_{\mathrm{loss}}(12)=100[1+(0.40/0.60)(1-e^{-1.44})]=150.871\) units. These predictions match the rounded observations. Thus this hypothetical fit estimates a 40% fractional inhibition of loss and a baseline loss-rate constant of \(0.20\ \mathrm{h}^{-1}\).
Worked Example: Comparing the Production-Inhibition Fit
For production inhibition with \(I\geq 0\), the predicted response satisfies \(R_{\mathrm{prod}}(t)\leq R_0\) for every \(t\geq0\), because \(1-e^{-k_{\mathrm{out}}t}\geq0\). All three observations are above 100 units. Under least squares with these observations, the constrained optimum is therefore \(I=0\), which predicts 100 units at every time; changing \(k_{\mathrm{out}}\) cannot improve that fit when \(I=0\).
The resulting sum of squared residuals is approximately \( (114.225-100)^2+(134.217-100)^2+(150.871-100)^2 =202.35+1170.8+2588.0=3961\ \text{units}^2\). The constrained production-inhibition model thus fails to reproduce the upward profile, whereas the loss-inhibition fit reproduces it closely. This is strong evidence against production inhibition under the stated assumptions and for these hypothetical data, not universal proof about a drug’s biological mechanism.
Local uncertainty and what the fit identifies
A close fit does not by itself show that the parameter estimates are precise. For a local approximation, let \(J\) be the sensitivity matrix whose rows contain the derivatives of the model predictions with respect to \(I\) and \(k_{\mathrm{out}}\) at each sampling time. Suppose, specifically for this uncertainty calculation, that the response errors are independent, have mean zero, and have known standard deviation \(\sigma=2\) response units. With the baseline fixed and the model locally approximated by a linear function of the parameters, the approximate covariance matrix is \(\sigma^2(J^\mathsf{T}J)^{-1}\).
For the loss-inhibition model at \(I=0.40\), \(k_{\mathrm{out}}=0.20\ \mathrm{h}^{-1}\), and \(R_0=100\), differentiating the stated solution gives the following sensitivities at 2, 6, and 12 h:
Using the exact derivatives before rounding gives approximately:
The matrix is invertible, so the two parameters are locally distinguishable in this design, but its large off-diagonal term indicates substantial local correlation. The approximate standard errors from \(\sigma^2(J^\mathsf{T}J)^{-1}\) are \(0.0318\) for \(I\) and \(0.0375\ \mathrm{h}^{-1}\) for \(k_{\mathrm{out}}\). Using a normal Wald approximation, the corresponding 95% intervals are \(0.40\pm1.96(0.0318)\), or \(0.338\) to \(0.462\), for \(I\), and \(0.20\pm1.96(0.0375)\), or \(0.126\) to \(0.274\ \mathrm{h}^{-1}\), for \(k_{\mathrm{out}}\). These are approximate intervals conditional on the model, fixed baseline, known error standard deviation, and local linearization; they are not guaranteed to have exact 95% coverage for a nonlinear model.
Worked Example: Checking Local Parameter Uncertainty
The determinant of the displayed sensitivity cross-product matrix is approximately \(42501(30541)-34310^2=1.208\times10^8\). The diagonal elements of its inverse are therefore approximately \(30541/(1.208\times10^8)\) and \(42501/(1.208\times10^8)\). Multiplying each by \(\sigma^2=4\) and taking square roots gives \(SE(I)\approx0.0318\) and \(SE(k_{\mathrm{out}})\approx0.0375\ \mathrm{h}^{-1}\).
The off-diagonal element of the inverse is negative, so the local covariance between the estimates is negative: within this profile, a somewhat larger \(I\) can be partly offset by a smaller \(k_{\mathrm{out}}\). This trade-off is why a numerical estimate should be accompanied by uncertainty and a check that the sampling times provide meaningful information about both parameters.
Recovery measurements as an additional check
Once the drug effect has been removed, both models return to the drug-free equation \(dR/dt=k_{\mathrm{in}}-k_{\mathrm{out}}R\). If washout occurs at time \(T\), the recovery profile at \(s\) hours after washout is \(R(T+s)=R_0+[R(T)-R_0]e^{-k_{\mathrm{out}}s}\). Provided \(R(T)\ne R_0\), \(s>0\), and \([R(T+s)-R_0]/[R(T)-R_0]>0\), an exact pair of recovery values and a known \(R_0\) imply \(k_{\mathrm{out}}=-\ln\{[R(T+s)-R_0]/[R(T)-R_0]\}/s\). In practice, multiple noisy recovery measurements should be fitted rather than treating one ratio as exact.
Worked Example: Estimating Turnover From Hypothetical Recovery Data
Suppose the response is 150.871 units at washout and, 4 h later, is observed to be 123.0 units. With \(R_0=100\), the excess response falls from \(50.871\) to \(23.0\) units. The recovery estimate is:
This is close to the post-dose estimate of \(0.20\ \mathrm{h}^{-1}\), supporting a coherent turnover interpretation in this hypothetical example. Recovery can help constrain the drug-free loss rate because the drug effect is absent during that phase. It does not automatically establish which process the drug inhibited; that comparison still depends on the full time course and the assumptions about washout and baseline.
Diagnostics, alternatives, and limitations
Compare residuals over time, not only a single fit statistic. Systematic residual curvature can indicate that the constant-effect assumption is wrong, the chosen mechanism is inadequate, or an omitted process such as baseline drift matters. Check that estimates are within their admissible ranges, that optimization reaches the same solution from reasonable starting values, and that predictions remain stable when sampling times or individual observations are examined. A flat likelihood or a strong parameter trade-off warns that the data do not sharply identify the parameters, even if an optimizer reports a precise-looking point estimate.
The main alternative to selecting one mechanism is to fit both and compare their likelihoods or another appropriate measure of fit while accounting for model complexity and the observation-error assumptions. Residual patterns and biological plausibility also matter. If the models have similar predictions over the observed sampling window, a small difference in fit does not establish the mechanism. Conversely, a model that cannot produce the observed direction under its parameter constraints is plainly inadequate for that dataset, as production inhibition is in the worked profile.
The uncertainty calculation above is local and assumes independent errors with known standard deviation, a fixed baseline, and an adequate model. If the error scale is estimated from only three post-dose points and two fitted parameters, the residual degrees of freedom are \(3-2=1\), so that estimate would be very unstable. Uncertainty in \(R_0\), correlated errors, concentration changing over time, incomplete washout, or inhibition near a parameter boundary can further change the uncertainty assessment. A local Wald interval can be misleading when the likelihood is asymmetric or the parameters are strongly correlated; profile likelihood is a useful alternative for investigating such cases.
Check whether \(R_0\) is stable and whether treating it as fixed is justified.
Use the same observations and error assumptions, and check predictions against the observed direction and shape.
Inspect parameter uncertainty and correlations, then test whether the drug-free recovery profile supports the estimated turnover rate.
Check Your Understanding
Answer each question, then compare with the solutions below.
- At baseline, what relationship links \(k_{\mathrm{in}}\), \(k_{\mathrm{out}}\), and \(R_0\)?
- For \(R_0=80\), \(I=0.25\), and \(k_{\mathrm{out}}=0.10\ \mathrm{h}^{-1}\), calculate the production-inhibition prediction at 5 h.
- Why can the constrained production-inhibition model not fit the rising hypothetical profile in the tutorial?
- What additional information can a post-washout recovery profile provide?
- What does a strong local correlation between \(I\) and \(k_{\mathrm{out}}\) imply for interpretation?
Solutions: (1) Baseline steady state requires \(k_{\mathrm{in}}=k_{\mathrm{out}}R_0\). (2) \(R=80[1-0.25(1-e^{-0.10(5)})]=80[1-0.25(1-e^{-0.5})]\approx72.13\) response units. (3) With \(I\geq0\), \(R_{\mathrm{prod}}(t)\leq R_0\), while the observed responses exceed baseline. (4) It can help estimate or check the drug-free turnover rate because the drug effect is absent during recovery. (5) It indicates that different parameter combinations can produce similar local predictions, so estimates should be interpreted jointly and their uncertainty examined.