From Translational Evidence to Human Prediction
A human concentration–time profile is uncertain because the parameters used to predict it are uncertain. Translational data may support a range of plausible human clearances, volumes, or potency values rather than a single value. A Bayesian analysis represents that uncertainty with a probability distribution conditional on the model, the observed evidence, and the prior distribution. Predictions are then made across that distribution, rather than by inserting only one parameter estimate into the model.
The target must be stated precisely: for example, uncertainty in the expected concentration at a specified time, or uncertainty in a future measured concentration that also includes residual error. The distinction matters. The first is uncertainty in a model-based quantity; the second additionally allows for deviations between that quantity and an observation. As in the earlier material on defining translational prediction targets, the prediction target determines which sources of uncertainty belong in the calculation.
This expression requires a specified model and likelihood, a prior distribution, and a well-defined data set. The posterior describes uncertainty conditional on these choices; it does not automatically account for structural model error or uncertainty in the model itself. A posterior predictive distribution for a future outcome \(Y_{\mathrm{new}}\) is obtained by averaging its conditional distribution over the parameter posterior:
If the target is the model-predicted mean profile \(g(t,\theta)\), the corresponding posterior distribution is instead the distribution of \(g(t,\theta)\) as \(\theta\) varies according to \(p(\theta\mid D,M)\). This distinction separates parameter uncertainty in the mean profile from residual or observation variability included in \(p(Y_{\mathrm{new}}\mid\theta,M)\).
A Conjugate Update on the Log Scale
A useful transparent case occurs when a positive parameter is modeled on the log scale. Write \(\eta=\log(CL)\), where \(CL\) is human clearance. Suppose the prior for \(\eta\) is Normal, and a translational estimate \(z\) is Normally distributed around \(\eta\) with known standard error \(s_z\). The assumptions are that this likelihood adequately represents the evidence, the estimate is unbiased on the log scale, and its error variance is known. With a Normal prior, the posterior is also Normal.
The posterior variance is smaller than either input variance when both are positive, because the posterior precision is the sum of prior and data precisions. This result follows by multiplying the two Normal densities: the terms involving \(\eta\) in the exponent are quadratic, with coefficient \(-\tfrac12(1/v_0+1/s_z^2)\), and the linear term is \(\eta(m_0/v_0+z/s_z^2)\). Completing the square gives variance \(v_1\) and mean \(m_1\) above. This update is not a universal recipe: a non-Normal likelihood or a nonlinear translational model generally requires a different posterior calculation.
Worked Example: Updating Human Clearance From Translational Evidence
All values are hypothetical. Suppose the prior for log human clearance, with clearance measured in L/h, is \(N(\log 5,0.40^2)\). A translational analysis provides an estimate \(z=\log 6\) with standard error \(0.20\) on the log scale. Treat this estimate as Normal with mean \(\log(CL)\), and assume the standard error is known.
The prior variance is \(v_0=0.16\), and the data variance is \(s_z^2=0.04\). Therefore,
The posterior mean on the log scale is
Thus the posterior median clearance is \(\exp(1.7553)=5.785\) L/h. A central 90% posterior interval for log clearance is \(m_1\pm1.645\sqrt{v_1}\), or \(1.7553\pm0.2943\). Exponentiating the endpoints gives approximately 4.31 to 7.76 L/h. Conditional on this prior, likelihood, data, and model, 90% of the posterior probability for clearance lies in that interval. It is not a statement that the fixed clearance has a 90% repeated-sampling probability of lying there.
Predicting a Human Concentration Profile
For each draw from the posterior, evaluate the human PK model using that draw and the specified dose and covariates. The resulting collection of profiles represents uncertainty in the model-predicted concentration curve. With a one-compartment intravenous bolus model and fixed volume \(V\), for example, concentration at time \(t\) is \(C(t)=D/V\exp(-CLt/V)\). This model assumes instantaneous input, linear elimination, and fixed \(V\); it is not appropriate without qualification for other routes or more complex disposition.
Worked Example: Posterior Uncertainty in a One-Compartment Profile
Use the clearance posterior from the previous example, and suppose the hypothetical human dose is 500 mg intravenously with fixed \(V=50\) L. The model is \(C(t)=D/V\exp(-CLt/V)\), so \(D/V=10\) mg/L. Assume no additional uncertainty in dose or volume, and no residual error because the target is the model-predicted concentration, not a future assay result.
Because concentration decreases monotonically as clearance increases, the 90% posterior interval for concentration is obtained by evaluating the concentration model at the upper and lower clearance limits in reverse order. The median profile uses the posterior median clearance, 5.785 L/h. For example, at 8 h:
The following values show the median profile and the pointwise 90% posterior intervals obtained by transforming the clearance quantiles 4.31 and 7.76 L/h. The lower concentration endpoint uses the higher clearance.
| Time (h) | Lower 90% endpoint (mg/L) | Median (mg/L) | Upper 90% endpoint (mg/L) |
|---|---|---|---|
| 2 | 7.33 | 7.93 | 8.42 |
| 8 | 2.89 | 3.96 | 5.02 |
| 24 | 0.241 | 0.622 | 1.26 |
These are pointwise intervals: at each listed time, they summarize posterior uncertainty at that time. They are not a simultaneous 90% band guaranteeing that an entire future curve lies within the band. The increasing separation at later times reflects the influence of clearance uncertainty on elimination over time.
In a practical simulation, generate posterior draws \(\theta^{(1)},\ldots,\theta^{(S)}\), evaluate \(g(t,\theta^{(s)})\) at each desired time, and summarize the resulting values by posterior medians and quantiles. If the target is a future measured concentration, also draw an observation from the conditional observation model for each parameter draw. This adds only the variability explicitly represented by that model. For instance, a log-normal assay error model adds assay variability; it does not automatically add between-person variability unless individual parameters are also sampled.
Name the human quantity, time points, dose, and whether the target is a mean profile or a future observation.
Use draws that represent the joint posterior distribution, preserving parameter dependence when it is present.
Calculate the human profile for every draw; for future observations, additionally draw from the specified conditional observation model.
Report posterior medians or means and quantiles with units, and check that the model and sampled parameter values yield biologically and numerically plausible profiles.
Propagating Potency Uncertainty to a PD Response
The same approach applies when a response depends nonlinearly on uncertain potency. Nonlinearity means that evaluating the response at a posterior mean parameter does not generally give the posterior mean response: \(E[g(\theta)\mid D]\) need not equal \(g(E[\theta\mid D])\). Predictions should therefore be computed for parameter draws before summarizing the responses.
Worked Example: Response Uncertainty From an EC50 Posterior
Suppose, hypothetically, that log \(EC_{50}\) has prior \(N(\log 2,0.50^2)\), and a translational potency estimate is \(z=\log 3\) with known standard error 0.25 on the log scale. Assume the estimate is unbiased and Normal conditional on log \(EC_{50}\). The Normal update gives posterior variance \((4+16)^{-1}=0.0500\), posterior standard deviation 0.2236, and posterior mean
The posterior median \(EC_{50}\) is \(\exp(1.0175)=2.767\) concentration units. A 90% posterior interval is \(\exp(1.0175\pm1.645(0.2236))\), approximately 1.916 to 3.997 concentration units. At a fixed concentration \(C=3\) in the same units, suppose the model is \(E(C)=100C/(EC_{50}+C)\), with fixed \(E_{\max}=100\) response units and zero baseline. The median-parameter response is
Because response decreases as \(EC_{50}\) increases at fixed concentration, transforming the posterior endpoints in reverse order gives a 90% interval of approximately \(100(3)/(3.997+3)=42.88\) to \(100(3)/(1.916+3)=61.02\) response units. This is an interval for the model-predicted response arising from potency uncertainty. If individual response observations have additional residual variability, that variability must be represented and simulated separately to predict future observations.
Assumptions, Diagnostics, and Alternatives
A posterior prediction is conditional on both the statistical model and the translational assumptions used to construct it. The update in the examples assumed a Normal likelihood on the log scale, an unbiased translational estimate, and a known standard error. The PK example additionally fixed volume, assumed a one-compartment bolus model and linear elimination, and omitted residual and population variability. The PD example treated \(E_{\max}\) and the concentration as known. If these assumptions are not credible, the resulting interval may omit important uncertainty or be systematically misleading.
Useful checks include confirming that posterior draws respect parameter constraints, evaluating whether predicted profiles have plausible units and behavior, and examining whether a small number of extreme draws dominates summaries. Compare the prior and posterior to see whether the data meaningfully changed the parameter distribution; an almost unchanged posterior may reflect weak information, while an unexpectedly narrow posterior can signal an over-precise likelihood or an omitted uncertainty source. Simulation summaries should also be stable when the number of valid posterior draws is increased. These checks assess computation and plausibility; they do not establish that the translational model is biologically correct.
| Approach | What it summarizes | Important qualification |
|---|---|---|
| Plug-in prediction | Profile evaluated at a single parameter estimate | Does not show parameter uncertainty in the profile |
| Bootstrap interval | Sampling uncertainty represented by resampling and refitting | Its interpretation depends on the resampling design and fitted procedure |
| Bayesian posterior prediction | Predictions averaged over the posterior, optionally including an observation model | Conditional on the prior, likelihood, and model assumptions |
A bootstrap interval, covered in the earlier material on bootstrap intervals for translational parameters, and a Bayesian posterior interval are not interchangeable by definition. They arise from different constructions and may differ when data are sparse, parameters are constrained, or prior information has a substantial role. A plug-in curve is useful as a concise central scenario, but it should not be presented as quantifying uncertainty. Bayesian prediction is especially useful when uncertainty in a parameter distribution must be carried into a human profile; its limitation is that the answer can depend on assumptions not tested by the data.
Common Pitfalls
- Reporting a parameter interval as a future-observation interval. A posterior interval for clearance alone does not include assay error or patient-to-patient variability. Add those components only when they are part of the prediction target and model.
- Using the posterior mean as if it were the whole posterior. This discards uncertainty and can distort nonlinear predictions. Evaluate the model over posterior draws, then summarize the predictions.
- Calling a pointwise band a whole-profile guarantee. Pointwise quantiles describe separate times. A simultaneous statement about an entire curve requires a summary constructed for that curve-level target.
- Interpreting posterior probability without its conditions. A 90% posterior interval has 90% posterior probability conditional on the specified model, prior, and data; it does not certify the model or guarantee 90% coverage in every repeated-sampling setting.
Check Your Understanding
Answer each question, then compare with the concise solutions below.
- If a human parameter has posterior \(N(2,0.09)\) on the log scale, what are the posterior median and standard deviation on the log scale?
- Why does a posterior interval for model-predicted concentration not necessarily describe future measured concentrations?
- At fixed \(C>0\), does \(E(C)=E_{\max}C/(EC_{50}+C)\) increase or decrease as \(EC_{50}\) increases?
- What does a pointwise 90% posterior interval at each time fail to guarantee about an entire concentration–time curve?
Solutions: (1) The median on the original parameter scale is \(\exp(2)=7.389\); the log-scale standard deviation is \(\sqrt{0.09}=0.300\). (2) It includes uncertainty in the modeled concentration due to parameter uncertainty, but not necessarily residual assay error or individual variability; these require additional model components. (3) It decreases, since the denominator increases with \(EC_{50}\) while the numerator is fixed and positive. (4) It does not guarantee that 90% of entire future curves lie within the pointwise bounds simultaneously; each time point is summarized separately.