When a Biomarker and a Clinical Endpoint Tell Different Stories
Phase II dose selection can become difficult when a biomarker changes steadily with dose but the primary clinical endpoint does not. The biomarker may show that the drug engages its biological target; it does not, by itself, establish that patients experience greater clinical benefit at higher doses. Conversely, a noisy or insensitive clinical endpoint may fail to reveal a real dose effect. The decision therefore depends on the quality and relevance of each endpoint, the dose-response models consistent with the data, and whether the conclusion holds under reasonable alternatives.
This tutorial works through a hypothetical case. The goal is not to force both endpoints into a single curve or to choose whichever looks more dose-responsive. It is to identify what each endpoint can support, quantify uncertainty, and make a recommendation that is proportionate to the evidence.
Set Up the Decision Before Comparing Curves
Suppose a randomized Phase II trial compared placebo with three doses of a hypothetical drug. The primary endpoint is improvement in a symptom score at Week 12; positive values mean greater improvement than placebo, in score points. A difference of 2.5 points has been specified as the smallest clinically meaningful average benefit. The biomarker is percent suppression of a disease-related pathway at Week 4; larger values indicate greater pathway suppression.
Before interpreting either pattern, the analysis needs a defined target: the treatment effect for the intended patient population, endpoint, and time point. The example assumes randomized, concurrent dose groups; consistent endpoint measurement; a prespecified primary endpoint and analysis; and no major treatment-related differences in missing endpoint data. It also assumes the biomarker assay is reliable and measured at a time suitable for detecting the pathway response. If these conditions fail, apparent disagreement might reflect bias, timing, or measurement error rather than different biological and clinical dose-response relationships.
The table gives hypothetical dose-group estimates and standard errors. The standard errors describe sampling uncertainty in each estimated treatment effect, not uncertainty about whether the endpoint is clinically relevant.
| Dose | Primary endpoint benefit (score points) | SE | Biomarker suppression (%) | SE |
|---|---|---|---|---|
| Placebo | 0.0 | — | 0 | — |
| 25 mg | 1.8 | 0.9 | 30 | 5 |
| 50 mg | 3.2 | 0.9 | 62 | 5 |
| 100 mg | 3.5 | 1.0 | 84 | 4 |
The biomarker rises across the tested doses. The primary endpoint improves from 25 to 50 mg, then changes little from 50 to 100 mg. That contrast is the decision problem; it is not a reason to treat either pattern as conclusive before considering uncertainty and endpoint meaning.
Worked Example: Quantifying the Apparent Clinical Plateau
Use approximate 95% confidence intervals of estimate \( \pm 1.96SE \) to describe the primary endpoint at 50 and 100 mg. Then calculate the 100 mg minus 50 mg contrast, assuming the randomized dose groups are independent.
At 50 mg, the interval is \(3.2 \pm 1.96(0.9)=3.2 \pm 1.764\), or approximately 1.44 to 4.96 score points. At 100 mg, it is \(3.5 \pm 1.96(1.0)=3.5 \pm 1.96\), or approximately 1.54 to 5.46 points. Both point estimates exceed the 2.5-point clinical-importance threshold, but both intervals include values below it.
The estimated incremental benefit at 100 mg is \(3.5-3.2=0.3\) points. With independent arms, its standard error is \(\sqrt{1.0^2+0.9^2}=\sqrt{1.81}=1.345\) points. The approximate 95% interval is \(0.3 \pm 1.96(1.345)\), or \(-2.34\) to 2.94 points.
The data do not establish that 100 mg provides additional clinical benefit over 50 mg, nor do they establish that the doses are equivalent. The wide interval allows both a meaningful advantage and no advantage, or even a disadvantage, for 100 mg. The biomarker's continuing increase cannot narrow this clinical contrast unless a justified relationship between biomarker change and clinical benefit is available.
Check Whether the Endpoints Can Answer the Decision
Endpoint relevance is not decided by statistical significance alone. A useful review asks whether the primary endpoint measures an outcome that matters to patients, whether its measurement and analysis are credible, and whether the biomarker has evidence connecting its change to clinical benefit. A biomarker may be useful as a pharmacodynamic or mechanistic measure without being a validated surrogate endpoint. Surrogacy requires evidence that treatment effects on the biomarker reliably predict treatment effects on the clinical outcome in the intended context; a within-trial association or dose trend alone is not enough.
Timing also matters. A Week 4 biomarker response and a Week 12 clinical outcome may differ because biological engagement precedes clinical response. That possibility should motivate a justified time-course analysis, not an assumption that the later clinical endpoint must eventually follow the biomarker. Exposure, adherence, baseline disease activity, and missing observations can also affect the two endpoints differently.
Worked Example: What the Biomarker Increment Does and Does Not Show
Compare biomarker suppression at 50 and 100 mg. Assuming independent dose groups, the estimated increment is \(84-62=22\) percentage points. Its standard error is \(\sqrt{5^2+4^2}=\sqrt{41}=6.403\) percentage points. An approximate 95% interval is \(22 \pm 1.96(6.403)\), or 9.45 to 34.6 percentage points.
Thus, this hypothetical trial supports a greater average biomarker response at 100 mg than at 50 mg. It does not show that the extra suppression improves symptoms. The primary-endpoint contrast in the previous example remains 0.3 score points with a wide interval. Unless the biomarker is a credible surrogate or there is other relevant clinical evidence, the additional 22 percentage points should be interpreted as additional pathway suppression, not as additional patient benefit.
Represent Model Uncertainty Explicitly
A dose-response model summarizes how response may vary between tested doses and can inform the expected response at a candidate dose. For a shallow or plateauing clinical response, an Emax model is one plausible shape: benefit rises with dose and approaches a maximum. A linear model over the tested range is another possibility. They encode different assumptions about response beyond the observed dose levels. The model that gives the most attractive high-dose prediction should not be selected merely because it supports that dose.
Model comparison should consider fit, predictive performance, biological plausibility, and whether the model reproduces the observed dose pattern. Sparse Phase II dose groups can leave multiple shapes plausible. A model-averaged estimate can summarize predictions across supported models, but it cannot repair an irrelevant endpoint or weak data. Its model weights should reflect a defensible basis, such as prespecified predictive performance, rather than preference for a particular decision.
Worked Example: Averaging Two Clinical Dose-Response Models
Suppose predictive assessment gives an Emax model weight of 0.6 and a linear model weight of 0.4. At 50 mg, their predicted primary-endpoint benefits are 3.1 points (SE 0.7) and 2.8 points (SE 0.9), respectively. At 100 mg, predictions are 3.4 points (SE 0.8) and 4.0 points (SE 1.1). Treat the model-specific predictive distributions as normal for this illustration.
At 50 mg, the averaged mean is \(0.6(3.1)+0.4(2.8)=2.98\) points. The variance, including between-model uncertainty, is \(0.6[0.7^2+(3.1-2.98)^2]+0.4[0.9^2+(2.8-2.98)^2]=0.6396\), giving an averaged standard deviation of \(\sqrt{0.6396}=0.800\) points.
At 100 mg, the averaged mean is \(0.6(3.4)+0.4(4.0)=3.64\) points. The variance is \(0.6[0.8^2+(3.4-3.64)^2]+0.4[1.1^2+(4.0-3.64)^2]=0.9544\), giving a standard deviation of \(\sqrt{0.9544}=0.977\) points. The linear model's higher prediction at 100 mg increases both the averaged mean and the between-model uncertainty.
These summaries do not make 100 mg the preferred dose. The predictions are model-dependent, the observed primary endpoint shows little estimated incremental benefit from 50 to 100 mg, and the biomarker's added suppression does not establish added clinical benefit. The example makes the disagreement between plausible clinical models visible rather than hiding it behind one selected curve.
Make the Recommendation Conditional and Auditable
For this hypothetical case, 50 mg is a defensible provisional Phase III dose if the decision objective is to select a dose with evidence of clinically relevant primary-endpoint benefit while avoiding an unsupported assumption that further biomarker suppression brings further clinical benefit. The 50 mg point estimate is 3.2 points, above the 2.5-point threshold, and the estimated clinical increment at 100 mg is small. This is a reasoned choice under uncertainty, not proof that 50 mg is optimal or that 100 mg is ineffective. If 100 mg has a separately established clinical advantage or if the primary endpoint is found to be insensitive, the recommendation could change.
A transparent decision should state what would change it. Relevant evidence includes confirmation that the primary endpoint and its analysis are reliable, consistency of the dose pattern under plausible model forms, a clinically credible link between biomarker change and outcome if the biomarker is to influence the dose choice, and whether key trial assumptions hold. The decision should also account for safety and tolerability evidence, without letting a biomarker substitute for either clinical benefit or safety assessment.
This approach differs from three tempting alternatives. Choosing the highest dose because it maximizes biomarker response assumes clinical relevance that may not exist. Choosing only the dose with the largest observed clinical mean ignores estimation uncertainty and model uncertainty. Selecting a composite score after seeing discordant results can obscure which endpoint drove the decision and create a new, unvalidated target. A prespecified primary clinical endpoint, supported by sensitivity analyses and explicit biomarker interpretation, keeps the reasoning inspectable.
Diagnostics, Failure Modes, and Limits
- Check the data and estimand. Review dose assignment, endpoint definitions, visit timing, assay performance, adherence, exposure, and missingness. A discrepancy caused by inconsistent measurement or differential dropout is not evidence of distinct biological dose-response curves.
- Inspect the observations, not only fitted curves. Plot dose-specific estimates and uncertainty, assess whether one dose group drives a model, and check residual patterns and influential observations. A smooth curve can conceal a poor fit or sparse support.
- Compare plausible shapes. Assess whether a plateau model, a linear trend over the observed range, or another scientifically justified form describes the primary endpoint. If predictions change substantially across plausible forms, report that dependence rather than presenting one curve as certain.
- Do not infer surrogacy from correlation. Patients with larger biomarker changes may also have larger clinical improvements for reasons unrelated to a causal biomarker-to-outcome relationship. Trial-level evidence across treatments and studies is more informative for surrogacy than a single pooled association.
- Recognize limited extrapolation. Predictions outside tested doses, to a different population, or at a different time point rely on assumptions not verified by these Phase II data. Model averaging represents uncertainty among included models; it does not cover every plausible model or bias.
The conclusion is therefore bounded: these data can support a provisional dose recommendation, but cannot prove that one dose is uniquely best, establish biomarker surrogacy, or guarantee the same response in a later population. The strength of the recommendation should match the strength and relevance of the evidence.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions below.
- In the hypothetical case, what does the increase from 62% to 84% biomarker suppression establish, and what does it not establish?
- Why does the 95% interval for the 100 mg minus 50 mg clinical contrast not establish equivalence?
- What uncertainty is omitted if the model-averaged variance includes only within-model variances?
- Name one finding that could make the provisional 50 mg recommendation less persuasive.
Solutions. (1) Under the stated assumptions, the estimated biomarker response is greater at 100 mg; it does not establish greater clinical benefit without a justified biomarker-outcome relationship. (2) The interval, approximately \(-2.34\) to 2.94 score points, includes differences in both directions and includes potentially important benefit; equivalence requires an appropriate margin and evidence that the contrast lies within it. (3) It omits between-model variation, represented by \((\mu_m-\bar{\mu})^2\), and can therefore understate predictive uncertainty. (4) Examples include a reliable clinical analysis showing meaningful added benefit at 100 mg, evidence that the primary endpoint is insensitive or biased, or validated evidence that the additional biomarker response predicts clinical benefit in this context.