Assay Error in Preclinical PK and PD Data

Why Assay Error Matters in Translation

Preclinical PK and PD models are fitted to measurements, not to the underlying true concentrations or biomarker values. Those measurements can be noisy, and the size and form of the noise can depend on the value being measured. If this error is ignored or represented poorly, the model may give too much weight to imprecise observations, distort estimated parameters, or give an overly narrow range of human predictions.

The purpose here is to represent the measurement process explicitly. Assay error is distinct from animal-to-animal variability: it concerns variation in a measurement around the value in the sample being assayed. It is also distinct from model discrepancy, which concerns systematic failure of the model to represent the biological process. A measurement-error model does not, by itself, resolve either biological variability or model discrepancy.

The conditional framing matters: the assay distribution describes how the recorded value could vary if the same sample were measured repeatedly. A typical assumption is that measurements are conditionally independent given their latent values and any shared batch effects included in the model. Calibration, dilution, and sample handling should be sufficiently controlled that the specified distribution is a credible approximation. Replicate measurements and quality-control samples can help assess that assumption.

Choosing an Error Scale and Variance Model

An additive error model is appropriate when the assay’s absolute standard deviation is roughly constant across the measurement range:

Here \(\sigma_a\) is the additive assay standard deviation, in concentration units. This model can assign non-negligible probability to negative results, so it is generally unsuitable near zero unless that approximation is acceptable over the range used.

If relative error is roughly constant, a proportional model is often more appropriate. One option is a normal model on the original scale with standard deviation proportional to the true value. A combined model allows both a background absolute component and a concentration-dependent component:

The parameter \(\sigma_p\) is dimensionless. The variance rises with concentration, while the additive component can remain important at low concentrations. This is a heteroscedastic model: the conditional variance changes with \(C_i\). For strictly positive measurements with approximately constant relative variation, a log-scale model is another option:

Under this parameterization, \(C_i\) is the conditional median of \(Y_i\), not its arithmetic mean. In fact, \(E(Y_i\mid C_i)=C_i\exp(\sigma_{\log}^2/2)\). If \(C_i\) is intended to be the arithmetic mean instead, the log-scale normal distribution needs a mean correction: use \(\ln(Y_i)\mid C_i\sim N(\ln(C_i)-\sigma_{\log}^2/2,\sigma_{\log}^2)\).

Likelihood, Weighting, and Results Below Quantification

An error model contributes to estimation through the likelihood: the probability density of each observed value given the model-predicted latent value. If the variance is known and differs across observations, a normal likelihood gives greater influence to precise observations. For independent observations with known variances \(v_i\), minimizing the negative log-likelihood is equivalent, apart from a constant, to minimizing the weighted squared residuals shown below.

The weights \(w_i\) are inverse variances, not inverse standard deviations. When variance parameters are estimated, or the variance depends on \(C_i\), the full likelihood also contains the variance-dependent normalization term. For the combined normal model, each observation’s negative log-likelihood contribution, up to a constant, is \(\tfrac{1}{2}\{\ln(v_i)+(Y_i-C_i)^2/v_i\}\). Omitting \(\ln(v_i)\) when \(v_i\) depends on fitted quantities can change the estimates.

A result reported as below the lower limit of quantification (LLOQ) is not an observed concentration equal to zero, nor is its exact value known to be the LLOQ. It is a censored observation: the result tells us that the assay value fell below a threshold. Under the combined normal error model, its likelihood contribution is the probability of falling below that threshold:

where \(\Phi\) is the standard normal cumulative distribution function. This uses the information actually reported without pretending that the unobserved result had a particular value. The normal model’s possibility of negative measurements still needs consideration, especially near zero; a positive-valued error model may be preferable when that feature is implausible.

How Assay Error Broadens a Human Prediction

Assay error can affect a human prediction indirectly by making parameters estimated from preclinical data less precise. The following calculation isolates that effect. It assumes a specified translational relationship, a known assay-error distribution, and no uncertainty in the relationship’s structure or inputs. Those assumptions make the assay contribution transparent; in an application, additional uncertainty sources may also need to be included.

Suppose a preclinical response is observed on the log scale and follows \(Z=\theta x+\epsilon\), where \(x\) is a known exposure-related input, \(\theta\) is an unknown parameter, and \(\epsilon\sim N(0,\sigma^2)\) represents assay error on the log scale. A proposed human prediction is the response at a specified input \(x_h\): \(R_h=\exp(\theta x_h)\). With one observation and known \(\sigma\), the likelihood-based estimate is \(\hat{\theta}=Z/x\), and its standard error is \(\sigma/|x|\). Since the prediction is linear in \(\theta\) on the log scale, the uncertainty in \(\ln(R_h)\) is \(|x_h|\sigma/|x|\).

Derivation and Practical Checks

The uncertainty calculation in the example follows directly from the observation model. Since \(Z=\theta x+\epsilon\), dividing by the known nonzero input gives \(\hat{\theta}=Z/x=\theta+\epsilon/x\). Therefore \(E(\hat{\theta})=\theta\) and \(\mathrm{Var}(\hat{\theta})=\sigma^2/x^2\). Multiplying by the known human input gives the estimated log response \(x_h\hat{\theta}\), with variance \(x_h^2\sigma^2/x^2\). This derivation relies on the specified linear relationship, known inputs, and the normal assay-error model; it does not account for uncertainty in any of those assumptions.

In practice, evaluate whether the error model describes the assay across the concentrations and biomarker values that inform the translation. Useful checks include plotting replicate residuals against predicted values, checking whether their spread rises with concentration, examining observed coefficients of variation by concentration range, and reviewing quality-control results across batches and dilution levels. A variance pattern that widens with concentration suggests constant-variance error may be inadequate. Systematic trends or batch shifts suggest that a simple independent error model may also be inadequate.

Results at or below the LLOQ deserve particular attention. A large fraction of censored measurements can leave the data with limited information about low concentrations; the model’s predictions there may then depend strongly on assumptions about the error distribution and PK/PD shape. Examine the fraction censored by time, dose, and species, and check whether the fitted model can reproduce that pattern. Do not infer that a model is reliable at low concentrations simply because it fits observations above the LLOQ.

Common Pitfalls and Limitations

• Using equal weights despite unequal assay precision. If variance changes with concentration, equal weighting treats observations with different precision as equally informative. Use a justified variance model and its likelihood.
• Confusing assay error with animal variability. Replicate assay measurements of the same sample inform measurement error; measurements from different animals can also vary biologically. These sources require distinct representation when both matter.
• Replacing a below-LLOQ result with zero or half the LLOQ. That creates a numerical observation not supplied by the assay. Use the probability of the censored event under the stated model.
• Reporting an assay-only interval as total human uncertainty. The interval in the worked example excludes other sources by construction. State which uncertainties it includes, and do not present it as a comprehensive prediction interval.

No single error model is correct for every assay. Replicate and validation data may be sparse, variance can vary by batch, and the nominal assay specification may not describe performance in a particular matrix. If the error model is poorly identified, human predictions can appear more certain than the evidence warrants. Sensitivity to plausible error structures and transparent reporting of censoring are therefore important, especially when preclinical observations near the quantification limit drive the prediction.