1. What Is Biomarker Surrogacy?
A biomarker is a measurable characteristic that can provide information about a biological process, disease state, pharmacologic activity, or response to treatment. In drug development, biomarkers can be used for diagnosis, patient selection, pharmacodynamic assessment, dose selection, safety monitoring, or prediction of clinical outcomes.
A biomarker becomes a candidate surrogate endpoint when it is intended to stand in for a clinical endpoint. The fundamental question is stronger than whether the biomarker responds to treatment:
This distinction is central. A biomarker can be biologically relevant and strongly associated with a clinical outcome without being a valid surrogate for treatment effects.
A candidate surrogate sits between treatment and clinical outcome. Surrogacy concerns whether information about treatment effects on the biomarker reliably captures treatment effects on the clinical outcome.
2. Biomarkers Have Different Roles in Drug Development
Not every biomarker is intended to replace a clinical endpoint. Its role depends on the scientific question and development stage.
| Biomarker role | Typical purpose | Does it require surrogate validation? |
|---|---|---|
| Diagnostic biomarker | Identifies or characterizes a disease or condition | Not necessarily |
| Prognostic biomarker | Provides information about future outcome independent of treatment | Not necessarily |
| Predictive biomarker | Identifies patients more likely to respond differently to a treatment | Not necessarily |
| Pharmacodynamic biomarker | Measures a biologic response associated with drug exposure | Not necessarily |
| Intermediate endpoint | Measures a process expected to occur before the clinical outcome | Often requires stronger validation when used as a substitute |
| Surrogate endpoint | Intended to substitute for a clinical endpoint in treatment-effect evaluation | Yes, to the extent required for the intended use |
A useful distinction is that association is not the same as surrogacy. A biomarker may predict who will experience an event without accurately capturing how much a treatment changes that event.
3. Why Association Alone Is Not Enough
Suppose a treatment \\(T\\) affects a biomarker \\(S\\), and the biomarker is associated with a clinical outcome \\(Y\\). A simple representation is:
This diagram suggests a pathway through which treatment could influence the outcome. But real biological systems can contain additional pathways:
The second pathway represents a treatment effect on the clinical outcome that is not mediated through the biomarker.
There may also be common causes of both biomarker and outcome, feedback mechanisms, measurement error, or multiple biological pathways. Consequently, observing that patients with favorable biomarker values tend to have favorable outcomes does not establish that changing the biomarker will produce the corresponding change in outcome.
4. What Makes Surrogacy Model-Based?
Model-based drug development provides a framework for representing relationships among exposure, biomarkers, disease processes, and clinical outcomes simultaneously.
A simplified model-based chain might be written as:
Each arrow can be represented using a mathematical model. For example, pharmacokinetics can describe concentration over time, a PD model can describe biomarker response, and a disease-progression model can describe how the biomarker or latent disease state relates to a clinical outcome.
The advantage is that the biomarker is not evaluated in isolation. Instead, its position in the broader drug-disease-response system can be examined quantitatively.
5. Linking Drug Exposure to a Biomarker
Consider a simple indirect-response model in which drug exposure inhibits a biomarker-generating process. A possible model is:
Here, \\(S(t)\\) is the biomarker concentration, \\(C(t)\\) is drug concentration, \\(k_{\mathrm{in}}\\) is the baseline input rate, \\(k_{\mathrm{out}}\\) is the turnover rate, and the inhibitory function describes drug action.
A model of this type can answer questions such as:
- How rapidly does the biomarker respond to drug exposure?
- What exposure produces a specified magnitude of biomarker change?
- Is the biomarker response delayed relative to concentration?
- How long does the biomarker take to return toward baseline?
- How does biomarker response vary across dose levels?
These are important pharmacologic questions, but they still do not establish that the biomarker is a surrogate for a clinical outcome.
6. Linking the Biomarker to the Clinical Outcome
The next modeling step is to connect the biomarker to the outcome of interest. For a continuous clinical outcome, one possibility is:
where \\(S_i\\) is the biomarker, \\(X_i\\) represents additional covariates, and \\(Y_i\\) is the clinical outcome.
For a binary outcome, a logistic model could instead be used:
For time-to-event outcomes, the biomarker can enter a hazard model:
Longitudinal biomarkers can also be incorporated through joint models, latent disease-state models, or other dynamic approaches.
7. Individual-Level Versus Trial-Level Surrogacy
Surrogacy is often evaluated at more than one level.
Individual-level association
Individual-level association asks whether individuals with different biomarker values tend to have different clinical outcomes. A correlation or regression relationship may be useful for describing this association.
Trial-level association
Trial-level surrogacy asks whether the treatment effect on the biomarker predicts the treatment effect on the clinical endpoint across randomized trials.
For trial \\(j\\), let \\(\Delta S_j\\) denote the treatment effect on the biomarker and \\(\Delta Y_j\\) the treatment effect on the clinical outcome. A simple meta-analytic relationship is:
A strong individual-level association does not automatically imply a strong trial-level association. This is one of the most important conceptual distinctions in surrogate-endpoint evaluation.
8. How Is Surrogacy Quantified?
Several statistical frameworks have been proposed for evaluating surrogate endpoints. The appropriate measure depends on the outcome type, study design, estimand, and scientific objective.
| Concept | What it asks | Typical evidence |
|---|---|---|
| Individual-level association | Do biomarker values distinguish individuals with different outcomes? | Correlation, regression, joint modeling |
| Trial-level association | Do treatment effects on the biomarker predict treatment effects on outcome? | Meta-regression, treatment-effect models |
| Surrogate explained variation | How much variation in outcome treatment effects is explained by biomarker treatment effects? | Measures such as \\(R^2\\)-type quantities |
| Residual treatment effect | How much treatment effect remains after accounting for the biomarker? | Direct-effect or mediation models |
| Predictive performance | How accurately does biomarker information predict clinical benefit? | External validation and predictive assessment |
No single numerical measure should be interpreted without considering the underlying causal model, study population, measurement quality, and intended use of the biomarker.
9. Surrogacy and Mediation Are Closely Related—but Not Identical
Mediation analysis asks how much of a treatment effect operates through an intermediate variable. Suppose treatment \\(T\\) affects biomarker \\(S\\), which in turn affects outcome \\(Y\\).
Conceptually, the total treatment effect can be decomposed into pathways involving the biomarker and pathways that do not involve it.
This framework can be useful for understanding mechanism. However, demonstrating that a substantial fraction of an effect is mediated through a biomarker is not automatically equivalent to establishing that the biomarker is a valid surrogate endpoint for future treatment effects.
Surrogacy is ultimately a predictive question: can treatment effects on the biomarker reliably stand in for treatment effects on the clinical endpoint?
10. Biomarker Measurement Error Matters
A biomarker can be biologically meaningful but measured imprecisely. Let the true biomarker be \\(S_i^*\\), while the observed measurement is:
Measurement error can weaken observed associations and make it more difficult to estimate treatment effects accurately.
Sources of variability may include:
- Analytical assay variability.
- Sample collection and handling.
- Biological within-person variability.
- Timing relative to dosing.
- Laboratory batch effects.
- Missing or below-quantification-limit measurements.
Longitudinal modeling can help separate some components of variability, but it cannot automatically recover information that was never measured adequately.
11. Why Time Matters
Many biomarkers change dynamically after treatment. A single post-dose measurement may therefore fail to capture the relevant pharmacologic signal.
Suppose biomarker response follows:
The magnitude and timing of the response depend on \\(\Delta S\\) and \\(k_S\\). Two treatments could produce the same biomarker value at one time point while having substantially different trajectories.
Model-based approaches can represent:
- Baseline biomarker values.
- Time-dependent treatment effects.
- Delayed pharmacodynamic responses.
- Turnover and recovery.
- Repeated measurements within individuals.
- Relationships between biomarker trajectories and later clinical outcomes.
This is especially important when the clinical consequence of a biomarker change depends on the duration or cumulative magnitude of the response.
12. Joint Models for Biomarkers and Clinical Outcomes
A powerful approach is to model the longitudinal biomarker and clinical outcome simultaneously.
For example, a longitudinal submodel might be:
where \\(m_i(t)\\) is the underlying latent biomarker trajectory.
A survival submodel can then relate the latent trajectory to event risk:
The treatment effect can therefore be modeled jointly with the biomarker trajectory and clinical event process.
Joint models can be particularly useful when biomarker measurements are noisy, longitudinal, and subject to informative dropout or when the underlying biomarker trajectory is more scientifically meaningful than any single observed measurement.
13. Biomarkers in Disease-Progression Models
In model-based drug development, a biomarker may represent an observable indicator of an underlying disease process rather than being the final clinical outcome itself.
A conceptual disease-progression model might be:
where \\(D(t)\\) represents a latent or observed disease state. A biomarker could then provide information about \\(D(t)\\), while the clinical outcome depends on the disease state.
Alternatively, the biomarker itself may be part of the causal pathway between treatment and outcome. The appropriate model depends on the biology and the measurement process.
14. Worked Example: A Hypothetical Biomarker Surrogate
Consider a hypothetical randomized trial program evaluating a treatment for a chronic disease. The treatment reduces a biomarker \\(S\\), and the clinical endpoint is a disease-related event.
Suppose three completed trials provide estimated treatment effects on the biomarker and the clinical endpoint:
| Trial | Biomarker treatment effect \\(\Delta S\\) | Clinical treatment effect \\(\Delta Y\\) |
|---|---|---|
| Trial 1 | −0.40 | −0.18 |
| Trial 2 | −0.20 | −0.10 |
| Trial 3 | −0.60 | −0.27 |
Step 1: Examine the treatment-effect relationship
A simple trial-level model is:
The observed pattern suggests that larger reductions in the biomarker are accompanied by larger reductions in the clinical outcome in these hypothetical trials.
Step 2: Ask the stronger question
The key question is not merely whether the three points are correlated. It is whether this relationship would reliably predict treatment effects in new trials.
Step 3: Consider uncertainty
Only three trials provide limited information about the trial-level relationship. Treatment-effect estimates also have sampling error. A fitted relationship could therefore appear strong while providing poor prediction in future settings.
Step 4: Consider alternative mechanisms
The biomarker could correlate with clinical benefit without capturing all treatment pathways. Another treatment might reduce the biomarker while producing little or no improvement in the clinical endpoint if its biological mechanism differs.
15. How Is a Candidate Surrogate Validated?
Surrogate validation generally requires evidence that extends beyond a single trial. The precise validation strategy depends on the disease, treatment class, endpoint, biomarker, and intended regulatory or scientific use.
A model-based validation program can include:
- Biological plausibility. Establish a credible mechanistic relationship between treatment, biomarker, and outcome.
- Measurement characterization. Quantify assay performance, reproducibility, timing, and sources of variability.
- Individual-level evidence. Evaluate whether biomarker trajectories or values are associated with clinical outcomes.
- Trial-level evidence. Evaluate whether treatment effects on the biomarker predict treatment effects on the clinical endpoint across randomized studies.
- External validation. Test predictive performance in trials or populations not used to construct the model.
- Context evaluation. Determine whether the relationship is consistent across relevant treatments, populations, doses, and disease stages.
A surrogate may be useful for one treatment class or development setting but not necessarily for another. Surrogacy is therefore often context-dependent.
16. Surrogacy Is Better Viewed as a Spectrum
The evidence supporting a biomarker can vary considerably. A useful conceptual spectrum is:
| Evidence level | Interpretation |
|---|---|
| Biological rationale | The biomarker is plausibly related to the mechanism of disease or treatment. |
| Prognostic association | Biomarker values are associated with future clinical outcomes. |
| Pharmacodynamic response | The biomarker changes in response to drug exposure. |
| Mediation evidence | Part of the treatment effect appears to operate through the biomarker. |
| Trial-level surrogacy | Treatment effects on the biomarker predict treatment effects on the clinical outcome across trials. |
| Predictive validation | The biomarker successfully predicts clinical treatment effects in independent settings. |
These categories should not be treated as interchangeable. A biomarker can have compelling biological and prognostic evidence while still lacking evidence sufficient for substitution of a clinical endpoint.
17. Common Pitfalls in Biomarker Surrogacy
- Confusing correlation with causation. A strong biomarker-outcome association does not establish that changing the biomarker changes the outcome.
- Using individual-level association as proof of trial-level surrogacy. These address different questions.
- Ignoring treatment mechanism. Surrogacy relationships may not generalize across drugs with different mechanisms.
- Ignoring measurement error. Noisy biomarker measurements can distort estimated relationships.
- Ignoring time. A single measurement may not represent the relevant longitudinal exposure or response.
- Overfitting trial-level relationships. A model developed from few trials may provide unstable predictions.
- Ignoring uncertainty in treatment effects. Estimated treatment effects on both biomarker and clinical endpoint contain sampling error.
- Extrapolating beyond the evidence base. A surrogate validated in one disease stage, treatment class, or population may not transfer automatically.
- Assuming complete mediation. Treatment may influence the clinical outcome through pathways that are not represented by the biomarker.
18. A Practical Model-Based Surrogacy Workflow
- Define the clinical endpoint. Specify the outcome for which the biomarker is being considered as a surrogate.
- Define the biomarker estimand. Decide whether the relevant quantity is a change from baseline, slope, exposure-response metric, trajectory, or another summary.
- Map the causal structure. Identify plausible pathways from treatment to biomarker to clinical outcome.
- Characterize the biomarker. Evaluate assay performance, timing, missingness, and within-person variability.
- Build the exposure-response model. Describe how drug exposure affects the biomarker.
- Model the biomarker-outcome relationship. Use an appropriate continuous, binary, count, longitudinal, or time-to-event model.
- Evaluate treatment-effect relationships. Examine whether treatment effects on the biomarker predict treatment effects on the clinical endpoint.
- Quantify uncertainty. Propagate uncertainty from PK, biomarker, and clinical-outcome models.
- Validate externally. Test predictions using independent trials or datasets.
- Define the intended use. Clearly distinguish exploratory biomarker use from use as a potential surrogate endpoint.
19. How Surrogacy Fits Into Model-Based Drug Development
Biomarker surrogacy is particularly valuable when it can connect early pharmacologic evidence to later clinical outcomes.
A broader model-based drug-development framework might look like:
This framework can integrate information from preclinical studies, early-phase trials, dose-ranging studies, biomarker studies, and confirmatory trials.
When supported by appropriate evidence, such models can help identify informative dose ranges, characterize delays between exposure and response, quantify uncertainty, simulate trial scenarios, and evaluate whether biomarker information can improve prediction of clinical outcomes.
The important distinction is that model integration does not eliminate the need for empirical validation. A sophisticated mechanistic model can improve understanding while still depending on assumptions that must be tested against clinical evidence.
20. Using a Surrogacy Model for Prediction
Suppose a model estimates that the treatment effect on a biomarker is \\(\Delta S\\), and a validated trial-level relationship predicts the clinical treatment effect as:
A future study could then use the observed or predicted biomarker effect to generate a prediction of the corresponding clinical effect.
But the prediction should include uncertainty. Conceptually:
This distribution should reflect uncertainty in the historical relationship, the new biomarker effect, measurement, and any relevant model parameters.
21. Surrogacy and Clinical Development Decisions
A candidate surrogate can serve different purposes during development. For example, a biomarker may be useful for demonstrating target engagement or pharmacodynamic activity without being sufficiently validated to replace a clinical endpoint in a confirmatory trial.
Therefore, questions should be separated:
| Question | What the evidence needs to establish |
|---|---|
| Does the drug affect the biomarker? | Evidence of pharmacologic activity |
| Does the biomarker relate to disease? | Prognostic or biological relevance |
| Does treatment-induced biomarker change relate to clinical benefit? | Evidence of treatment-effect surrogacy |
| Can the biomarker predict future clinical effects? | External predictive validation |
| Can the biomarker replace the clinical endpoint? | Evidence appropriate for the intended clinical or regulatory use |
These are progressively stronger claims and should not be treated as equivalent.
22. Where the Modeling Can Go Next
More advanced analyses can extend biomarker surrogacy models in several directions:
- Hierarchical models for multiple trials, populations, and treatment classes.
- Joint longitudinal-survival models for dynamic biomarkers and time-to-event outcomes.
- Mechanistic disease models linking biomarkers to latent disease processes.
- Bayesian models for incorporating prior biological knowledge and propagating uncertainty.
- Exposure-response models linking PK exposure directly to biomarker treatment effects.
- Dynamic treatment-effect models when treatment effects change over time.
- Simulation to assess how biomarker uncertainty affects clinical-trial operating characteristics.
- Causal inference frameworks to distinguish prognostic relationships, mediation, and treatment-effect surrogacy.
These approaches can provide a richer quantitative representation of the treatment-disease system, particularly when biomarker measurements are repeated and clinical outcomes occur over long follow-up periods.
23. Key Takeaways
- A biomarker is not automatically a surrogate simply because it is biologically relevant or associated with clinical outcomes.
- Surrogacy concerns whether treatment effects on a biomarker reliably predict treatment effects on a clinical endpoint.
- Individual-level biomarker-outcome association and trial-level treatment-effect surrogacy address different questions.
- PK/PD models can connect drug exposure to biomarker response and provide a quantitative description of pharmacologic activity.
- Clinical-outcome models can connect biomarker trajectories or latent disease states to events, continuous outcomes, or other endpoints.
- Joint longitudinal and survival models can be useful when biomarkers are repeatedly measured and clinical outcomes occur over time.
- Measurement error, missing data, timing, and within-person variability can materially affect surrogacy analyses.
- A treatment can affect a clinical outcome through pathways that are not captured by the biomarker, so complete mediation should not be assumed.
- Trial-level evidence is particularly important when the intended use is to predict clinical treatment effects from biomarker treatment effects.
- Surrogacy relationships may depend on the treatment mechanism, disease population, endpoint, biomarker measurement, and development context.
- External validation is important because a model that describes historical trials may not necessarily predict new trials.
- Model-based drug development can integrate PK, biomarkers, disease progression, and clinical outcomes into a coherent quantitative framework.
- The ultimate value of a surrogate is determined by its ability to provide reliable information about clinically meaningful treatment effects for its intended use.
Where to Go Next
A natural progression is to study PK/PD biomarker models, followed by indirect-response models, disease-progression models, longitudinal biomarkers, joint models for longitudinal and time-to-event data, mediation analysis, and trial-level surrogate-endpoint validation.
The next tutorial can build directly on these ideas by showing how biomarker trajectories are incorporated into longitudinal and survival models, including how treatment effects, latent disease states, and clinical events can be modeled jointly.