Selecting a biomarker threshold makes sample size efficiency compete with the breadth of the evidence base. In a heterogeneous population, an unselected Phase 2 design may require 40% to 60% more subjects than an enriched cohort to reach the same 80% power, while producing evidence that applies to fewer patients when the threshold is restrictive.
The Design Decision Behind Enrichment
Enrichment is a protocol-level choice about which biological population will generate the primary evidence. The decision affects treatment effect, event or response variance, screening yield, recruitment time, and the population to which the result can reasonably be generalized. A threshold can concentrate pathway-active patients, but it can also exclude patients whose response mechanism is less directly measured.
The relevant comparison is not simply enriched versus unselected enrollment. Teams need to compare candidate thresholds under a common operating model, including prevalence, assay performance, treatment effect by biomarker level, endpoint variability, dropout assumptions, and the number screened to obtain the randomized cohort. A threshold that increases the conditional response difference may still reduce overall efficiency if the eligible fraction becomes too small.
The central design question is therefore where additional biological selectivity stops compensating for lost breadth and added screening burden.
Predictive and Prognostic Information
A predictive biomarker modifies the expected treatment effect. A prognostic biomarker instead relates to outcome risk regardless of treatment assignment.
The distinction determines how enrichment should be interpreted. If a biomarker is prognostic only, selecting higher-risk patients may change observed outcomes without demonstrating greater treatment sensitivity. A predictive biomarker can support a treatment-by-biomarker interaction, but that interaction still requires an estimand, prespecified analysis, and adequate representation across biomarker strata.
Confusing the two mechanisms can produce an apparently favorable enriched result that does not establish pathway dependence. The protocol should specify whether the biomarker defines the intended treatment population, improves baseline outcome prediction, or serves both roles. Those claims have different implications for Type I error control, false negative risk, and subsequent enrollment assumptions.
Choosing a Defensible Biomarker Threshold
Consider a therapy aimed at an IL-17A-dependent pathway in moderate-to-severe psoriasis. Serum IL-17A levels could identify patients with greater pathway activity, but the threshold must reflect more than a convenient split of the distribution.
A usable threshold requires a biological rationale, an assay strategy, and an analysis plan. Investigators should examine whether response changes continuously with serum IL-17A, whether the assay separates pathway-active from less active disease, and whether the cutoff can be measured consistently during screening.
A binary biomarker threshold is an operational rule imposed on continuous biology.
- Biological rationale for the biomarker and its treatment interaction
- Assay precision and threshold reproducibility
- Prevalence of the biomarker-positive population at screening
- Generalizability beyond the enriched cohort
Threshold selection should be assessed across plausible distributions rather than at one assumed prevalence. The same cutoff can produce different recruitment demands and power when biomarker prevalence, assay misclassification, or treatment effect gradients change. Sensitivity analyses should therefore include both the statistical operating characteristics and the practical screening burden.
A Virtual Cohort Scenario: CRP and Response
C-reactive protein (CRP) provides a second illustration of threshold tradeoffs. Suppose treatment effect increases among patients with higher CRP, while the biomarker remains distributed across the target population.
We simulated 1,500 virtual patients per simulation run, assigning treatment response through a sigmoidal Emax relationship and evaluating the ACR20 outcome measure under alternative CRP thresholds.
ACR20 endpoint
- No enrichment produced an 8.2 ACR20 percentage points treatment effect and required 186 patients per arm for 80% power.
- Selecting CRP > 10 mg/L produced 12.4 percentage points, required 108 patients per arm, and yielded approximately 1.4:1 enrichment efficiency; this threshold included approximately 70% of population.
- Selecting CRP > 20 mg/L produced 16.1 percentage points, required 72 patients per arm, and yielded approximately 2.4:1 enrichment efficiency; this threshold included approximately 42% of population.
The scenario illustrates why the largest conditional effect is not automatically the best protocol choice. CRP > 20 mg/L reduces the randomized sample size further, but it also narrows the evidence base and increases screening demand because fewer eligible patients meet the biomarker threshold. CRP > 10 mg/L provides a less restrictive compromise between response enrichment and population coverage. Each option uses the same ACR20 outcome measure, yet the operating consequences differ materially.
From Simulation to Protocol
Simulation can compare enrichment stringency before the protocol fixes a threshold. For each candidate rule, the analysis should report treatment effect, power, sample size, biomarker prevalence, screening burden, and enrichment efficiency. Sample size re-estimation may be considered when accumulating blinded information can update nuisance assumptions without changing the scientific question.
Enrichment has a regulatory cost and is not costless from a regulatory perspective. Its appropriate use is proof of concept, not a workaround for uncertain biology or inadequate trial execution. A narrow Phase 2 cohort may support a targeted development hypothesis, but it does not automatically justify broad Phase 3 enrollment assumptions.
The FDA draft guidance on enrichment strategies in clinical trials is publicly available regulatory guidance that can inform the design discussion. Teams should define the intended population, biomarker rationale, and analysis consequences before treating an enriched result as evidence for wider use.
Enrichment Is Not the Only Tool
Enrichment may fail when a biomarker is prognostic rather than predictive, assay error obscures the pathway-active group, or treatment response is broad across biomarker levels. A restrictive cutoff can then exclude biologically relevant patients without creating a meaningful treatment contrast. Continuous biomarker modeling, prespecified interaction testing, covariate adjustment, or stratified randomization may retain information discarded by a binary rule. The appropriate design depends on the mechanism, endpoint, assay characteristics, and target population, not on the assumption that a smaller cohort is more informative.
The practical comparison should include designs that preserve biomarker information rather than treating enrichment as the default. A continuous model may detect a gradient, while stratification can balance biomarker groups without excluding them. These approaches still require control of Type I error and protection against a false negative interaction result.
For Phase 2, the decision is whether the selected cohort provides a credible proof of concept at an acceptable screening and regulatory cost. An enriched result can sharpen the development hypothesis, but the observed effect, threshold, and eligible-population fraction must remain explicit when planning the next trial.