Biomarker Translation and Surrogate Uncertainty

Why a Biomarker Prediction Can Be Uncertain

A biomarker can provide an early signal about a drug’s pharmacologic effect, but translating a biomarker response into an outcome prediction involves two distinct questions. First, how precisely was the biomarker measured? Second, how precisely is the relationship between that biomarker and the outcome known? A precise assay does not guarantee a reliable outcome prediction if the relationship is uncertain, and a well-characterized relationship cannot remove substantial uncertainty in the biomarker value itself.

The distinction matters when animal biomarker responses are used to inform human outcomes. An animal measurement may be noisy, and the calibration relating biomarker to outcome may be estimated from limited data. Moreover, a biomarker associated with outcome is not automatically a validated surrogate: predicting an outcome from a biomarker does not establish that changing the biomarker causes the outcome to change, or that the relationship transports unchanged across species.

Definition: Biomarker measurement uncertainty is uncertainty about the relevant true biomarker value arising from assay error or sampling. Biomarker-to-outcome relationship uncertainty is uncertainty about the mapping from a true biomarker value to the outcome, including uncertainty in estimated relationship parameters and, for an individual outcome, residual outcome variability. These components describe different sources and should not be treated as interchangeable.

A Model That Separates the Sources

Consider a linear calibration relationship between a true biomarker value $B$ and an outcome $Y$:

Y=\alpha+\beta B+\varepsilon,\qquad \varepsilon\sim N(0,\sigma^2).

Here, $\alpha$ is the intercept, $\beta$ is the expected change in outcome per one-unit increase in the true biomarker, and $\sigma^2$ is the residual variance among outcomes at a given biomarker value. In calibration data, the biomarker may be observed with error: $W=B+e$, where $W$ is the assay result and $e$ is measurement error. For a human prediction, suppose the available information supports a true biomarker value with mean $m$ and variance $v_B$. The variance $v_B$ can reflect assay uncertainty in a measured value; for $k$ independent replicate assays with per-assay error variance $\tau^2$, the assay contribution to the variance of their mean is $\tau^2/k$.

Let $\hat\alpha$ and $\hat\beta$ be estimates from calibration data, with estimated covariance matrix $V$. The following decomposition assumes that the calibration estimates are approximately jointly normal, the biomarker uncertainty is independent of those estimates, and the residual variation is independent with constant variance. It also assumes that the linear relationship is appropriate over the biomarker range of interest and that the relationship can be used for the target population. Those are substantive assumptions, not consequences of the calculation.

Prediction-variance result: Under these assumptions, the approximate variance for an individual outcome prediction is $$ V_{\mathrm{pred}}\approx \sigma^2+ \begin{bmatrix}1&m\end{bmatrix} V \begin{bmatrix}1\\m\end{bmatrix} + (\hat\beta^2+V_{\beta\beta})v_B. $$ The first term is residual outcome variability; the second is uncertainty in the estimated relationship, evaluated at $m$; the third is uncertainty from the true biomarker value. For a prediction of the conditional mean outcome rather than an individual outcome, omit $\sigma^2$.

The covariance matrix $V$ contains the estimated variances of $\hat\alpha$ and $\hat\beta$, as well as their covariance. Keeping the covariance matters: intercept and slope estimates can move together, so ignoring their covariance can either overstate or understate uncertainty. The final term includes $V_{\beta\beta}$ because the slope is itself uncertain while the biomarker value is uncertain. If only the biomarker uncertainty is being approximated with a first-order calculation, this term is often simplified to $\hat\beta^2v_B$; the fuller expression is used in the examples below.

Why the Variance Adds Up This Way

Write the predicted outcome as $\hat\alpha+\hat\beta B+\varepsilon$, with $B$ varying around $m$. First condition on the fitted relationship and on $B$. The residual contributes $\sigma^2$. The remaining variance is the variance of the predicted mean $\hat\alpha+\hat\beta B$. Applying the law of total variance, first average over $B$ conditional on the fitted coefficients, then account for the uncertainty in those coefficients.

\begin{aligned} \operatorname{Var}(\hat\alpha+\hat\beta B) &=\operatorname{Var}_{\hat\alpha,\hat\beta}\!\left(E_B[\hat\alpha+\hat\beta B]\right) +E_{\hat\alpha,\hat\beta}\!\left[\operatorname{Var}_B(\hat\alpha+\hat\beta B)\right]\\ &=\operatorname{Var}(\hat\alpha+m\hat\beta) +E(\hat\beta^2)v_B\\ &=V_{\alpha\alpha}+2mV_{\alpha\beta}+m^2V_{\beta\beta} +(\hat\beta^2+V_{\beta\beta})v_B. \end{aligned}

The second line uses $E(B)=m$, $\operatorname{Var}(B)=v_B$, and independence of the biomarker uncertainty and the fitted coefficients. The last line uses $E(\hat\beta^2)=\hat\beta^2+V_{\beta\beta}$. Adding $\sigma^2$ gives the individual prediction variance. This calculation quantifies uncertainty under the stated model; it does not establish that a biomarker-to-outcome relationship is biologically valid or transportable.

Worked Example: Separating Assay and Relationship Uncertainty

Suppose a hypothetical animal calibration model predicts an outcome in outcome units from a biomarker measured in biomarker units. The fitted values are $\hat\alpha=2.0$, $\hat\beta=1.50$, and $\sigma=1.20$. The coefficient covariance matrix is

V= \begin{bmatrix} 0.64&-0.08\\ -0.08&0.04 \end{bmatrix}.

A hypothetical human biomarker measurement is $4.0$ biomarker units, with assay standard error $0.30$ units. Thus $m=4.0$ and $v_B=0.30^2=0.09$ square biomarker units. The predicted conditional mean is

\hat Y=\hat\alpha+\hat\beta m=2.0+(1.50)(4.0)=8.00\ \text{outcome units}.

The relationship-parameter contribution is

\begin{aligned} \begin{bmatrix}1&4.0\end{bmatrix} V \begin{bmatrix}1\\4.0\end{bmatrix} &=0.64+2(4.0)(-0.08)+(4.0^2)(0.04)\\ &=0.64-0.64+0.64=0.640\ \text{outcome units}^2. \end{aligned}

The biomarker contribution is

(\hat\beta^2+V_{\beta\beta})v_B =(1.50^2+0.04)(0.09)=0.2061\ \text{outcome units}^2.

For an individual outcome, add the residual variance, $\sigma^2=1.20^2=1.44$. The total is $0.640+0.2061+1.44=2.2861$ outcome units squared, so the prediction standard deviation is $\sqrt{2.2861}=1.512$ outcome units. A normal-approximation 95% prediction interval is $8.00\pm1.96(1.512)$, or approximately $5.04$ to $10.96$ outcome units. This interval is conditional on the model and transport assumptions; it is not evidence that the biomarker is a causal surrogate.

For the conditional mean outcome rather than an individual outcome, omit the residual variance: the standard error is $\sqrt{0.640+0.2061}=0.920$ outcome units. The corresponding approximate interval is $8.00\pm1.96(0.920)$, or $6.20$ to $9.80$ outcome units. These intervals answer different questions: one includes individual residual scatter, while the other describes uncertainty about the expected outcome at the target biomarker value.

Replicates Reduce One Component, Not All Components

If replicate assays have independent, mean-zero errors with common variance, averaging them reduces assay variance in proportion to the number of replicates. It does not reduce uncertainty in the fitted slope or intercept, residual biological variation, or uncertainty about whether the relationship transports to humans. The benefit is therefore limited by the other components.

Worked Example: The Value and Limit of Replicate Assays

Use the hypothetical calibration and biomarker values from the preceding example, but suppose the human biomarker is the mean of four independent assays, each with standard error $0.30$ biomarker units. The variance of the mean is $v_B=0.30^2/4=0.0225$ square biomarker units. The mean assay result remains $4.0$, so the predicted mean remains $8.00$ outcome units.

The parameter contribution remains $0.640$ outcome units squared. The new biomarker contribution is $(1.50^2+0.04)(0.0225)=0.05153$ outcome units squared. Including residual variance, the total is $0.640+0.05153+1.44=2.1315$ outcome units squared, giving a prediction standard deviation of $1.460$ outcome units. Replicates reduced the biomarker component from $0.2061$ to $0.05153$, but the total prediction standard deviation fell only from $1.512$ to $1.460$ outcome units. Further assay replication would have diminishing value unless measurement uncertainty remained a substantial share of total uncertainty.

This calculation assumes independent replicate errors. If replicates share a systematic bias, averaging them does not remove that bias; the effective reduction in uncertainty will be smaller than the $1/k$ variance rule suggests.

Measurement Error Can Also Distort the Relationship

There is a second consequence when biomarker measurement error occurs in the calibration data. Regressing outcomes directly on noisy measurements $W$, as if $W$ were the true biomarker $B$, can bias the estimated slope toward zero under classical measurement error. This is distinct from uncertainty in a human biomarker measurement at prediction time: here the calibration relationship itself is distorted.

Classical attenuation result: If $W=B+e$, with $e$ independent of $B$ and the outcome residual, the population slope from regressing $Y$ on $W$ is $$ \beta_W=\beta\frac{\operatorname{Var}(B)} {\operatorname{Var}(B)+\operatorname{Var}(e)}. $$ Since the fraction is less than one when measurement error has positive variance, the observed slope is attenuated relative to the true slope.

Worked Example: A Noisy Calibration Biomarker

Suppose hypothetical calibration data have true biomarker variance $1.00$ square biomarker units, assay-error variance $0.25$ square biomarker units, and a true outcome slope of $2.00$ outcome units per biomarker unit. Under the assumptions in the attenuation result, the slope estimated by treating the noisy assay value as exact is expected to be

\beta_W =2.00\frac{1.00}{1.00+0.25} =2.00(0.80) =1.60\ \frac{\text{outcome units}}{\text{biomarker unit}}.

Thus, the expected fitted slope is 20% smaller than the true slope in this hypothetical setting. A translation based on that naïve slope would understate the expected outcome change per biomarker unit. The result relies on classical, additive, independent error; systematic error, nonlinearity, or dependence between error and biomarker can produce different distortions. Replicate assays, assay validation, or a model that explicitly accounts for the latent true biomarker may be needed to address this problem.

Diagnostics, Alternatives, and Limits

Before interpreting a biomarker-based prediction, inspect the calibration data and the intended translation. Plot outcome against biomarker, check whether a linear trend is plausible over the prediction range, and examine residual spread for changes with biomarker value. Check assay precision and whether replicate differences support the assumed measurement-error variance. Predictions far outside the calibration biomarker range depend heavily on the assumed relationship and should be identified as extrapolations rather than presented as ordinary interpolation.

A useful comparison is between a plug-in prediction and a prediction that propagates uncertainty. A plug-in calculation uses $\hat\alpha+\hat\beta m$ and reports a single value; it is simple, but hides uncertainty in both the measured biomarker and the fitted relationship. Propagating uncertainty makes those components visible. Replicate averaging is appropriate for reducing independent assay noise, but it cannot repair a poorly estimated or nontransportable relationship. An errors-in-variables or latent-biomarker calibration model can address noisy calibration measurements, but it requires enough information about assay error and the biomarker distribution to identify the relationship. Directly measuring the clinical outcome may be preferable when the biomarker-outcome mapping is weak or poorly established.

The decomposition has limits. It relies on a specified relationship and on estimates of its covariance and residual variance; uncertainty in those quantities is not fully represented when they are treated as fixed plug-in values. A normal-approximation interval may be inaccurate with sparse calibration data, strong nonlinearity, highly uncertain biomarker values, or predictions near a boundary. Most importantly, the calculation does not include every possible source of translation error. A relationship observed in animals may differ in humans because of biology, outcome definition, or study context. The uncertainty terms above do not automatically quantify such differences.

Common pitfall: “The assay is precise, so the predicted human outcome is precise.” This ignores uncertainty in the estimated slope and intercept, residual outcome variability, and whether the calibration relationship applies to humans. Report the prediction target, separate the quantified uncertainty components, and state the model and transport assumptions that remain unverified.

Check Your Understanding

Answer each question, then compare your reasoning with the solutions.

If the assay standard error is halved while the calibration data stay the same, which uncertainty component changes, and by what factor does its variance change?
In the prediction-variance result, why does the covariance between the estimated intercept and slope appear?
For a prediction of the conditional mean outcome, which variance component is omitted, and why?
Under what assumptions does measurement error attenuate the calibration slope toward zero?
Does a narrow interval for an outcome predicted from a biomarker establish that the biomarker is a causal surrogate? Explain.

Solutions. (1) The assay-related biomarker variance changes; if the standard error is halved, its variance is quartered. Other components are unchanged under the stated model. (2) The prediction uses both estimates, and their sampling errors may be correlated; the variance of their linear combination includes $2mV_{\alpha\beta}$. (3) Omit $\sigma^2$, because it represents residual variation among individual outcomes around the conditional mean, not uncertainty in that mean. (4) The result assumes additive classical error independent of the true biomarker and outcome residual, with the relationship otherwise linear. (5) No. A narrow interval quantifies uncertainty under the specified prediction model; it neither proves causation nor establishes that the relationship transports to the intended human setting.