Linking Biomarker Exposure-Response to Clinical Efficacy

Why Link Biomarker Response to Clinical Efficacy?

A treatment may change a biomarker before a clinical endpoint can be measured, or the endpoint may be costly, slow, or difficult to observe. A biomarker can therefore help describe the treatment’s biological response and may contain information about later clinical outcomes. The development question is not simply whether exposure changes the biomarker. It is whether biomarker change is related to the clinical endpoint in a way that is useful, reliable, and appropriately interpreted.

A linked analysis usually considers at least two relationships: exposure with biomarker response, and biomarker response with clinical outcome. These relationships can support prediction and help organize evidence about a treatment’s effects. They do not, on their own, show that changing the biomarker causes clinical benefit or establish that the biomarker can replace the clinical endpoint.

Specify the Link Before Interpreting It

Let \(X\) denote a selected exposure metric, \(M\) a biomarker change, and \(Y\) a clinical outcome. For a continuous biomarker and continuous clinical endpoint, a simple pair of models is:

Here, \(\alpha_1\) describes the exposure–biomarker association. In the clinical model, \(\beta_1\) describes the expected difference in \(Y\) associated with a one-unit difference in \(M\), comparing individuals with the same modeled exposure \(X\). The coefficient \(\beta_2\) describes the exposure association with \(Y\) conditional on \(M\). These are model-based associations unless additional causal assumptions are justified.

The form of the clinical outcome model must suit the endpoint. A continuous endpoint may use a linear model; a binary endpoint may use a logistic model; and a time-to-event endpoint may use an appropriate survival model. The coefficient’s scale then depends on the model: for example, a logistic regression coefficient is on the log-odds scale, not the probability scale. The same interpretive caution applies across model types.

Define the analysis population, exposure metric, biomarker summary, clinical endpoint, and timing before fitting the models. The biomarker should be measured in a window that is meaningful for the clinical outcome; if effects are delayed, a contemporaneous biomarker value may not represent the relevant response history. Baseline values, background therapy, disease severity, and other common causes of biomarker and outcome should be considered based on the question and causal assumptions. Earlier material on causal assumptions in individual-level exposure–response is directly relevant: conditioning on a post-treatment biomarker can change the estimand and can introduce bias.

A useful prediction target is the expected clinical outcome for a specified exposure after using the exposure model to predict the biomarker. But uncertainty in both fitted models matters. Plugging a predicted biomarker into a clinical model gives a point prediction; it does not by itself provide a valid confidence interval for the complete prediction.

Quantifying Uncertainty in the Biomarker–Outcome Relationship

For a linear model with independent observations, a correctly specified mean relationship, and errors with mean zero and constant variance, the estimated slope is uncertain because the observed outcomes vary around the fitted line. A confidence interval expresses this sampling uncertainty. With normally distributed errors, the usual small-sample \(t\) interval also requires that the errors be normally distributed; sufficiently large samples may permit approximate inference under weaker distributional conditions.

For a simple linear regression of \(Y\) on \(M\), define \(S_{MM}=\sum_i(M_i-\bar{M})^2\), where \(\bar{M}\) is the sample mean of \(M\). The fitted slope is \(\hat{\beta}_1=S_{MY}/S_{MM}\), where \(S_{MY}=\sum_i(M_i-\bar{M})(Y_i-\bar{Y})\). If the residual sum of squares is \(SSE\), then the residual variance estimate is \(s^2=SSE/(n-2)\), and \(SE(\hat{\beta}_1)=s/\sqrt{S_{MM}}\). The \(100(1-\alpha)\%\) confidence interval is \(\hat{\beta}_1 \pm t_{1-\alpha/2,\,n-2}SE(\hat{\beta}_1)\).

The slope formula follows by choosing the intercept and slope that minimize the residual sum of squares. For any fixed slope, the minimizing intercept makes the residuals sum to zero, so the fitted line passes through \((\bar{M},\bar{Y})\). Write the fitted value as \(\bar{Y}+b(M_i-\bar{M})\). Differentiating the sum of squared residuals with respect to \(b\), setting the derivative to zero, and solving gives:

Association Is Not Surrogate Validation

A biomarker can respond strongly to exposure and still be a poor indicator of clinical benefit. The exposure–biomarker relationship answers whether the biomarker changes with exposure. The biomarker–outcome relationship asks whether biomarker values and clinical outcomes co-vary. Neither question alone establishes that treatment effects on the biomarker reliably predict treatment effects on the clinical endpoint.

Individual-level association and trial-level surrogacy are different evidentiary questions. An individual-level association means that, within the analyzed population and model, people with different biomarker values tend to have different outcomes. Trial-level surrogacy concerns whether treatment effects on the biomarker across relevant trials predict treatment effects on the clinical endpoint. Even a strong individual-level association may arise because disease severity affects both measurements, or because treatment exposure affects both through separate pathways.

Diagnostics, Alternatives, and Limitations

Before interpreting a linked model, check whether its assumptions and measurements are plausible. Useful checks include plots of clinical outcome against biomarker, residuals against fitted values and exposure, and observations with high influence. Look for curvature, unequal residual spread, outliers that dominate the slope, and differences in data availability across exposure or biomarker ranges. A narrow biomarker range can make the slope imprecise even when the fitted line appears steep. If measurement error in the biomarker is substantial, a naive regression can attenuate the estimated association; estimated exposure or biomarker uncertainty should not be ignored when it is material.

Timing and selection also matter. A biomarker measured after clinical deterioration may reflect the outcome rather than predict it. Restricting analysis to participants with both observed biomarker and clinical data can bias results if missingness is related to disease status or response. A model that omits a common cause of biomarker and outcome may report an association that is not attributable to the biomarker itself. These issues are reasons to state the target and assumptions, not reasons to relabel an association as causal.

The main alternatives answer different questions. Modeling the clinical endpoint directly as a function of exposure is often the clearest approach when the goal is exposure–efficacy characterization; it does not require the biomarker to explain the clinical response. A joint model can represent correlated biomarker and clinical outcomes and propagate uncertainty, but its conclusions still depend on its structure and assumptions. A mediation analysis targets a causal decomposition into pathways and requires additional causal assumptions beyond fitting a biomarker coefficient. Comparing treatment effects across trials is relevant to trial-level prediction but requires multiple informative trials and cannot be replaced by one individual-level regression.

A linked biomarker analysis is most useful as one part of an evidence chain: it can show whether the biomarker responds to exposure, whether it is associated with the clinical endpoint, and how uncertain those patterns are. It has limitations when the biomarker is measured imprecisely, the relationship changes across populations or time, the clinical endpoint is poorly represented by the model, or evidence is sparse at the exposures of interest. Keep the conclusion proportional to the analysis: prediction or association may be supported even when causal mediation or surrogate validity is not.