Separating Baseline Input From Turnover Rate

Why Baseline Concentration Does Not Determine Baseline Input

A biomarker concentration at baseline is often treated as if it directly determines the rate at which the biomarker is produced. It does not. At steady state, the observed concentration reflects a balance between endogenous input and loss: a high concentration can result from high input, slow loss, or both. If baseline is estimated rather than fixed, the model may fit the observed concentrations well while leaving the input and turnover rate poorly separated.

The practical question is therefore not just whether a fitted curve passes through the data. It is whether the observations contain enough distinct information to estimate the baseline concentration, endogenous input, and turnover rate separately. Sensitivity calculations and profile plots help reveal when estimates depend strongly on assumptions or sparse observations.

The units provide an immediate check: multiplying concentration by an inverse-time rate gives concentration per time, the units of input. The equilibrium identity also makes clear that one baseline concentration supplies one relationship between two rates. Without further information, it cannot uniquely determine both.

What the Data Can Identify

For constant input and first-order loss, the turnover equation is

At equilibrium, the derivative is zero, giving \(R_{\mathrm{base}}=k_{\mathrm{in}}/k_{\mathrm{out}}\). If the only observations are steady-state baseline concentrations, the data may estimate this ratio, but do not separately identify the numerator and denominator. A post-perturbation time course can provide information about \(k_{\mathrm{out}}\) through its rate of change, but that information depends on the perturbation being adequately specified and on the sampling times capturing the response.

Profiling Baseline Instead of Fixing It

Consider a hypothetical washout in which endogenous input is completely suppressed at time zero and loss remains first-order. If the concentration at the start of washout is the unknown \(R_{\mathrm{base}}\), then

This form assumes the suppression is complete and immediate, \(k_{\mathrm{out}}\) is constant during observation, and the measured concentrations correspond to the stated times. For observations \(y_i\) at times \(t_i\), suppose the residuals are analyzed by unweighted least squares, with independent errors of common variance. At any fixed candidate value of \(k_{\mathrm{out}}\), the least-squares estimate of \(R_{\mathrm{base}}\) is available in closed form.

This is called profiling: hold one parameter fixed, optimize the other parameter(s), and record the best achievable fit. A profile plot displays the resulting objective against the fixed parameter. A shallow or extended low region means that a range of turnover rates can be paired with different baseline estimates and still fit nearly as well. A sharp minimum indicates stronger information about the rate, though it does not by itself guarantee that the model assumptions are correct.

Sensitivity Shows the Direction of the Trade-Off

A profile summarizes fit across parameter values. Local sensitivity gives a complementary view: it measures how much a prediction changes when a parameter changes slightly. Suppose a known fraction \(I\) of endogenous production is inhibited from time zero, with the baseline steady state \(R_{\mathrm{base}}\) before inhibition. The equation becomes \(dR/dt=k_{\mathrm{in}}(1-I)-k_{\mathrm{out}}R\). Substituting \(k_{\mathrm{in}}=k_{\mathrm{out}}R_{\mathrm{base}}\) and solving from \(R(0)=R_{\mathrm{base}}\) yields

This expression assumes a constant known inhibition fraction, unchanged first-order loss, and a pre-perturbation steady state. It shows that baseline scales the response, while turnover controls how quickly the response approaches its new level. At limited sampling times, changes in the two parameters can compensate, especially when the data do not clearly resolve the response’s curvature.

Reading Profiles and Diagnosing Weak Separation

A useful diagnostic is to plot the profiled objective against \(k_{\mathrm{out}}\), with the corresponding \(\widehat R_{\mathrm{base}}(k_{\mathrm{out}})\) shown on a second panel or as a companion curve. If the objective remains nearly unchanged while the estimated baseline moves substantially, the data weakly distinguish the parameter combinations. For a likelihood-based fit, the analogous display is the profile likelihood; uncertainty limits are obtained using an appropriate likelihood-based criterion and its assumptions, not by treating every low point as equally plausible.

Check whether the apparent profile shape is stable to reasonable choices of observation error model, starting values, and parameter bounds. A minimum located at a bound may indicate that the data favor values outside the explored range, or that the model is weakly identified. A narrow minimum can still be misleading if a major assumption—such as complete input suppression—is false. Inspect observed and predicted time courses and residuals as well as the profile.

The main alternatives answer somewhat different questions. Fixing baseline from an external, reliable measurement can reduce the trade-off, but treats that value as known and may understate uncertainty if it is noisy. Jointly estimating baseline and turnover uses the study data but must expose their dependence rather than hide it. Adding early and late observations, or measuring baseline repeatedly, can improve information when those observations target response scale and response rate, respectively. No fitting method can recover separate parameters if the experimental design contains no information to distinguish them.

Once \(R_{\mathrm{base}}\) and \(k_{\mathrm{out}}\) are estimated, the derived input \(k_{\mathrm{in}}=k_{\mathrm{out}}R_{\mathrm{base}}\) inherits uncertainty from both and their covariance. Treating the two estimates as independent can misrepresent uncertainty in input. Likewise, a precisely fitted baseline does not imply that the endogenous input is precise. The equilibrium relationship is a model-based conversion, not an additional independent observation.

Common Mistake and Practical Limits

A common mistake is to say, “The baseline concentration is measured, so baseline input is known; the remaining parameter is turnover.” That conclusion is only justified if turnover is known independently, because \(k_{\mathrm{in}}=k_{\mathrm{out}}R_{\mathrm{base}}\). The correction is to state what the data directly identify, profile the uncertain parameters, and report the resulting dependence or uncertainty.

Profiles are diagnostics, not repairs. They depend on the model, the error structure, and the parameters allowed to vary. A one-dimensional profile can miss confounding with other estimated parameters if those parameters were held fixed rather than re-optimized. Sparse sampling, uncertain perturbation magnitude, measurement error near baseline, or departures from steady state can all weaken the interpretation. In particular, a slow baseline change is not addressed by a constant-baseline model; the data and assumptions should be assessed before interpreting a turnover estimate mechanistically.