Biomarker Bridging Across Populations

When Does a Biomarker Support a Population Bridge?

A pharmacodynamic biomarker can provide evidence about whether a drug produces a similar biological response in a source and target population. But an observed biomarker response is not, by itself, evidence that the populations will receive similar clinical benefit. A useful bridge therefore requires several linked questions: does exposure produce a comparable biomarker response; are biomarker measurements comparable; is the biomarker change clinically meaningful; and does it predict clinical outcomes in the target population?

Let \(S\) identify population, with \(S=s\) for the source and \(S=t\) for the target. Let \(E\) denote drug exposure, \(B\) a pharmacodynamic biomarker, and \(Y\) a clinical outcome. A biomarker bridge is an argument that evidence about the \(E\)-to-\(B\) relationship, together with evidence linking \(B\) to \(Y\), is sufficiently transportable to inform the target population. It does not automatically establish that changing \(B\) causes a change in \(Y\), or that \(B\) is a validated surrogate endpoint.

The earlier material on transporting exposure-response relationships provides the general framework for comparing exposure-response models across populations. Here the focus is the additional evidence needed when the response of interest is a biomarker. The source and target populations, relevant treatment conditions, and clinical question should be specified first. In particular, the exposure range observed in the source data must support predictions over the target population's relevant range; otherwise, similarity at the observed source exposures does not justify extrapolation.

Four Questions for the Evidence

Exposure-response consistency. Compare the conditional relationship between exposure and biomarker response, accounting for relevant baseline values and covariates. A similar slope alone is not enough if baseline biomarker levels, intercepts, exposure distributions, or model shapes differ in ways that change predicted responses. The comparison should use a common exposure definition, a clinically relevant exposure range, and an explicit model. For example, a linear model might be written as:

Here \(E^*\) is a reference exposure, \(X_i\) contains prespecified covariates, and \(\varepsilon_i\) is unexplained variation. The intercept describes the expected biomarker at \(E^*\), while the slope describes the expected biomarker change per exposure unit. This model assumes a linear relationship over the exposure range used for inference, an appropriate covariate specification, and residual behavior adequate for the intended estimation. Plots and residual diagnostics should be used to assess whether these assumptions are reasonable. A nonlinear pattern, substantial unexplained population differences, or poor exposure overlap can make a single slope comparison misleading.

Measurement comparability. A population difference in recorded biomarker values may reflect assay platform, sample handling, calibration, or laboratory procedures rather than biology. Comparability should be evaluated with appropriately collected paired samples or other direct cross-calibration evidence over the relevant measurement range. An acceptable average bias does not guarantee that the assays agree at high or low values; proportional bias and differences in precision can also matter.

Clinical relevance. The biomarker change should be interpreted against a clinically meaningful scale, such as a prespecified threshold associated with an important outcome or a validated pharmacological rationale. A statistically detectable change can be too small to matter clinically. Conversely, absence of a statistically significant difference between population estimates does not demonstrate that they are sufficiently similar.

Predictive validity. The biomarker should predict the clinical outcome of interest in the target population with performance adequate for the intended decision. Evidence can include calibration, discrimination, prediction error, and performance at a prespecified clinical decision threshold. A biomarker may be associated with outcome yet still predict poorly. Prediction must be evaluated in data not used to choose the model or threshold, or by a suitable validation procedure.

How the Pieces Connect

The logic connecting biomarker response to clinical outcome can be stated using the law of total probability. For a target-population prediction at exposure \(E\) and covariates \(X\), integrate over the target distribution of biomarker values:

The density \(f(B\mid E,X,S=t)\) describes the target biomarker response at that exposure and covariate profile. The term \(P(Y\mid B,E,X,S=t)\) describes how clinical outcome varies with biomarker, exposure, and covariates in the target population. To use a source exposure-response model in this calculation, its predictions for \(B\) must be applicable to the target. To use a biomarker-outcome relationship, that relationship must also be supported in the target or justified as transportable.

This decomposition explains why consistency in the exposure-to-biomarker relationship is necessary for some bridging arguments but cannot establish the whole argument. If the biomarker-outcome relationship differs in the target, or if the biomarker is only weakly informative about clinical outcome, a well-matched biomarker response may still fail to predict clinical benefit. Moreover, observational association between \(B\) and \(Y\) does not show that an intervention that changes \(B\) will change \(Y\). Such a causal or surrogate claim requires evidence beyond individual-level prediction.

Worked Examples: Measurement and Clinical Prediction

Diagnostics, Alternatives, and Limitations

A credible assessment examines the data and model at each link rather than relying on a single test. Inspect exposure and biomarker distributions by population, including whether target exposures are covered by source data. Check the fitted exposure-response form with plots, residual patterns, and alternative plausible specifications. Compare both predicted biomarker levels and their uncertainty at decision-relevant exposures. For assay data, inspect paired differences against the average measurement and assess bias and precision over the range of interest. For clinical prediction, use validation data and report performance measures with uncertainty, not only a p-value for association.

The main alternatives answer different questions. A direct comparison of population-specific biomarker responses tests whether those responses are similar under the studied conditions. A model-based transport analysis predicts response after accounting for measured population differences, but depends on correct model specification, adequate overlap, and sufficient measurement of relevant effect modifiers. A clinical-outcome analysis evaluates whether biomarker information predicts the outcome directly, which is more relevant to decisions but requires clinical data in the target population. These approaches can complement one another; none makes the other evidence unnecessary.

Several limitations deserve particular attention. Small target samples can leave substantial uncertainty even when point estimates look similar. Range restriction can conceal a different exposure-response shape. A population-specific assay shift can mimic a biological difference, while calibration in one range may fail elsewhere. A biomarker may be prognostic—associated with outcome regardless of treatment—without predicting treatment benefit. And a predictor that performs well in one population or under one treatment may not retain calibration under different care, baseline risk, or clinical-outcome definitions.

A biomarker bridge is therefore strongest when the population-specific exposure-response predictions are supported over a common range, measurement differences are acceptably small, biomarker changes have a defensible clinical interpretation, and the biomarker predicts the target clinical outcome well enough for the intended use. The conclusion should state which of these claims the data support and where uncertainty remains.