Why Biomarker Classification Matters
A biomarker can change after dosing, identify patients at higher risk, distinguish patients with different treatment effects, or serve as a proposed substitute for a clinical outcome. These roles are not interchangeable. A biomarker that demonstrates drug activity may help confirm target engagement, yet provide no evidence that patients feel or function better. A marker associated with poor outcomes may help describe prognosis without identifying who benefits more from a particular treatment.
The classification determines what evidence is relevant. A pharmacodynamic claim concerns a biological response to treatment. A prognostic claim concerns outcome risk. A predictive claim concerns variation in treatment effect across marker-defined groups. A surrogate claim is stronger: it proposes that a biomarker endpoint can stand in for a clinical endpoint when evaluating treatment benefit. The same biomarker may have more than one role, but each role requires its own justification.
Four Roles, Four Questions
A practical starting point is to ask what claim is being made. “The drug changes the marker” is a pharmacodynamic claim. “People with this marker value tend to have more events” is prognostic. “The treatment effect differs by baseline marker group” is predictive. “The treatment’s effect on this marker can stand in for its effect on clinical benefit” is a surrogate claim.
The distinction between prognostic and predictive is especially important. Prognosis concerns outcome risk within a context; prediction concerns a contrast between treatments. A marker can be strongly prognostic yet have no value for choosing between treatments if both treatments have the same relative or absolute effect across marker groups. Conversely, a marker can be predictive even if its groups have similar outcome risk under control, provided the treatment effects differ.
For a simple binary outcome, let \(R_{T,m}\) denote the event risk with treatment \(T\) in marker group \(m\), and let \(R_{C,m}\) denote the risk with control in that group. The treatment effect on the risk-difference scale in group \(m\) is \(R_{T,m}-R_{C,m}\). Evidence of prediction requires evaluating whether these treatment-effect contrasts differ across groups, rather than comparing treated patients in one group with treated patients in another.
The scale must be specified: effect modification on a risk-difference scale need not match effect modification on a risk-ratio scale. The relevant scale should be chosen for the clinical question and stated when interpreting results.
Worked Example: A Baseline Marker That Is Prognostic and Potentially Predictive
Hypothetical data. In a randomized trial, 100 participants in each baseline-marker group receive treatment and 100 receive control. The clinical event counts are:
| Baseline marker group | Treatment events / participants | Control events / participants |
|---|---|---|
| High | 20 / 100 | 30 / 100 |
| Low | 8 / 100 | 10 / 100 |
Prognostic comparison. In the control group, event risk is \(30/100=0.30\) for the high-marker group and \(10/100=0.10\) for the low-marker group. The high-marker group has 20 percentage points higher observed control risk, so the marker is prognostic in this trial context, subject to uncertainty and the population studied.
Treatment-effect comparison. In the high group, the risk difference is \(20/100-30/100=-0.10\), or 10 percentage points fewer events with treatment. In the low group, it is \(8/100-10/100=-0.02\), or 2 percentage points fewer events. The difference in risk differences is \(-0.10-(-0.02)=-0.08\), an 8-percentage-point contrast suggesting a larger absolute benefit in the high group.
This pattern is compatible with predictive value on the risk-difference scale, but the point estimates alone do not establish it. A treatment-by-marker interaction analysis should quantify uncertainty in the contrast, and the trial must have adequate information for that comparison. Randomization supports treatment comparisons within baseline-defined groups; it does not eliminate sampling uncertainty or justify claims beyond the studied treatment and population.
Evidence Needed for Each Classification
The marker’s measurement must first be fit for its intended use: the assay should measure the stated quantity with adequate reliability, and the sampling time and units should be defined. Beyond that foundation, the evidence differs by role.
- Pharmacodynamic: show that marker change follows intervention or exposure in a way consistent with biological activity. A suitable comparison, such as randomized treatment versus control, helps distinguish drug-related change from natural variation, disease course, or measurement drift. This supports a biological-response claim, not automatically a clinical-benefit claim.
- Prognostic: show that marker status is associated with future clinical outcome in a clearly specified population and context. The association should be estimated with uncertainty and, when the aim is individual risk prediction, evaluated for predictive performance in relevant data. A prognostic association does not by itself establish that changing the marker changes outcome.
- Predictive: compare treatment effects across groups defined by a marker measured before treatment assignment. The key evidence is an adequately precise treatment-by-marker interaction on a prespecified, clinically interpretable scale. Separate significance tests within groups are not a valid substitute for testing whether the effects differ.
- Surrogate: establish substantially more than a relationship between marker values and outcomes. The treatment effect on the marker must provide dependable information about the treatment effect on the clinical endpoint in the intended setting. Evidence should address the relevant treatments, populations, and outcome context. A marker is not made a surrogate merely because it responds to treatment or predicts outcome.
The first three claims can be useful without making the fourth. A pharmacodynamic marker can support dose or mechanism decisions, and a prognostic marker can support risk stratification, while clinical outcomes remain necessary to establish clinical benefit. A predictive marker may inform treatment selection if its interaction evidence is credible and clinically useful. Substituting a biomarker endpoint for a clinical endpoint requires a stronger evidentiary basis because the decision then relies on the biomarker as a measure of treatment benefit.
Worked Evidence Assessments
Worked Example: Drug-Related Biomarker Change Is Not Clinical Benefit
Hypothetical data. In a randomized study, the mean change in a biomarker from baseline to day 14 is \(-18\) units among 40 participants receiving an active drug and \(-4\) units among 40 receiving placebo. Assume the same assay and visit schedule were used in both groups, and lower values represent the hypothesized biological response.
The estimated treatment contrast in mean change is:
Thus, the active-treatment group had a mean change 14 units lower than the placebo group over this period. Given the randomized comparison and the stated measurement conditions, this is evidence of a treatment-related pharmacodynamic response. Without standard errors or confidence intervals, its precision cannot be assessed from the supplied summaries.
No clinical outcome data are supplied. Therefore, these results do not establish that the drug improves symptoms, prevents events, or changes the overall benefit-risk balance. They also do not establish that the biomarker is a surrogate. The justified conclusion is limited to evidence of a difference in short-term biomarker change.
Worked Example: A Strong Outcome Association Does Not Prove Surrogacy
Hypothetical data. A treatment-control study measures a biomarker and a clinical event. The mean biomarker change is \(-25\) units with treatment and \(-5\) units with control. Clinical events occur in 12 of 200 treated participants and 13 of 200 controls.
The treatment contrast in mean biomarker change is \((-25)-(-5)=-20\) units, showing a larger mean reduction with treatment. The event risks are \(12/200=0.060\) and \(13/200=0.065\), giving a treatment-control risk difference of \(0.060-0.065=-0.005\), or 0.5 percentage points fewer events in the treatment group.
For illustration, the approximate standard error of the risk difference is:
An approximate 95% interval is \(-0.005\pm1.96(0.0242)\), or about \(-0.0524\) to \(0.0424\). This interval is wide and includes both a potentially meaningful reduction and an increase in event risk. The data do not precisely determine the clinical effect.
The biomarker appears treatment-responsive, but this one comparison does not show that biomarker change can stand in for clinical benefit. In particular, a treatment effect on the biomarker plus a noisy or uncertain clinical comparison is not validation of a surrogate. The appropriate conclusion is that the treatment changed the biomarker; the clinical benefit remains uncertain in these hypothetical data.
Justification: Why a Predictive Claim Requires an Interaction
Suppose a marker separates participants into groups 0 and 1, and treatment is randomized within each group. The treatment effect in group 0 is \(\Delta_0=R_{T,0}-R_{C,0}\); in group 1 it is \(\Delta_1=R_{T,1}-R_{C,1}\). A claim that treatment effect depends on marker group is a claim that these effects differ. Subtracting the group-specific effects gives:
If this contrast is zero, the estimated risk-difference treatment effect is the same in the two groups, even if their outcome risks differ greatly. If it is nonzero, the estimated treatment effects differ on this scale. Thus, comparing the interaction contrast directly addresses prediction; comparing marker groups only among treated participants mixes baseline risk differences with treatment-effect differences.
A credible predictive analysis also requires that the marker definition and analysis scale be clear, the marker be measured before treatment assignment, and uncertainty in the interaction be reported. A large subgroup difference can occur by chance, particularly when many markers or cut points are explored. Clinical usefulness further depends on whether the treatment-effect difference is large enough to alter a decision after considering harms, costs, and uncertainty.
Diagnostics, Alternatives, and Limitations
Several checks help reveal when a classification is being overstated:
- For a pharmacodynamic claim, inspect assay performance, timing, baseline imbalance, missing samples, and whether a similar change occurs in the comparison group.
- For prognosis, check whether the marker-outcome relationship applies to the stated population and context. Association alone does not guarantee accurate individual risk predictions.
- For prediction, assess the interaction estimate and its uncertainty rather than relying on “significant in one subgroup, not significant in another.” Those two separate results do not test the difference between subgroup effects.
- For surrogacy, ask whether evidence concerns clinical treatment effects, not only correlations among biomarker and outcome measurements. A treatment may affect the marker without its clinical effect being represented by that change.
The main alternative to relying on a biomarker as a surrogate is to measure the clinical outcome directly. This usually provides more direct evidence about patient benefit, though it may require longer follow-up or larger studies. Biomarkers can still be valuable for early biological assessment, risk stratification, or treatment selection when supported for those specific roles. The limitation is not that biomarkers are inherently uninformative; it is that evidence for one role cannot be automatically transferred to another.
A common mistake is to reason, “The drug changed the biomarker, and people with favorable biomarker values had better outcomes, so the biomarker proves clinical benefit.” The first statement supports pharmacodynamic activity; the second may support prognosis. Neither alone shows that treatment effects on the biomarker reliably represent treatment effects on the clinical outcome. Keep the claim, comparison, and evidence aligned.
Check Your Understanding
Answer each question, then compare with the brief solutions below.
- A drug reduces a laboratory marker relative to placebo, but clinical outcomes are not measured. Which role is supported, and which is not established?
- In control participants, event risk is 25% in a high-marker group and 10% in a low-marker group. What role does this pattern support, and what additional comparison is needed to claim prediction?
- Why is “significant treatment effect in one marker subgroup, not significant in the other” insufficient evidence that the marker is predictive?
- What is the treatment-effect difference across groups if risk differences are \(-0.06\) in group 1 and \(-0.02\) in group 0?
- Why does a biomarker’s association with clinical outcome not, by itself, establish it as a surrogate endpoint?
Solutions. 1. It supports a pharmacodynamic response if the comparison and measurement are credible; clinical benefit and surrogacy are not established. 2. It supports a prognostic association in that context. To support prediction, compare treatment effects between marker groups, preferably using a prespecified interaction on a stated scale. 3. The difference between one significant and one nonsignificant result is not itself a test that the effects differ; estimate and assess the interaction contrast. 4. The contrast is \(-0.06-(-0.02)=-0.04\), so the group 1 risk-difference treatment effect is 4 percentage points more favorable. 5. Association does not show that treatment-induced changes in the biomarker reliably represent treatment effects on the clinical outcome.