Biomarker-Stratified Trial Designs

Introduction

Modern clinical development increasingly asks not only whether a treatment works, but which patients are most likely to benefit. Biomarkers can help identify biologically defined populations and can therefore play an important role in clinical-trial design.

A biomarker may be a genomic alteration, protein expression measurement, pathway marker, immune marker, imaging characteristic, or another measurable biological characteristic.

When biomarker information is incorporated prospectively into trial design, the resulting study can be described as a biomarker-stratified trial.

The term encompasses several different designs, however. A trial may enroll patients regardless of biomarker status but stratify randomization by biomarker subgroup. Another trial may restrict enrollment to biomarker-positive patients. A more complex design may evaluate several biomarker-defined subgroups simultaneously and formally test whether treatment effects differ between them.

What Is a Biomarker-Stratified Trial?

Suppose a trial evaluates an experimental treatment versus control and patients are classified according to a binary biomarker:

The study therefore contains two biomarker-defined populations:

• Biomarker-positive: \(B=1\)
• Biomarker-negative: \(B=0\)

Treatment can then be represented by:

A biomarker-stratified design asks whether the treatment effect is consistent across these groups and, when scientifically justified, whether it differs according to biomarker status.

Why Use Biomarkers?

A biomarker can contribute to trial design for several reasons.

Prognostic vs. Predictive Biomarkers

One of the most important distinctions in biomarker research is between prognostic and predictive effects.

Prognostic Biomarker

A prognostic biomarker is associated with the clinical outcome independently of the treatment being evaluated.

For example, suppose biomarker-positive patients have longer progression-free survival than biomarker-negative patients in both treatment groups.

The biomarker may be prognostic.

Predictive Biomarker

A predictive biomarker modifies the relative effect of treatment.

For example, suppose the experimental treatment substantially improves outcome among biomarker-positive patients but provides little or no benefit among biomarker-negative patients.

That pattern suggests a potential treatment-by-biomarker interaction.

An Example of Prognostic and Predictive Effects

The Interaction Is the Key Question

Suppose the treatment effect among biomarker-positive patients is:

and among biomarker-negative patients:

The important scientific question is not simply whether:

or whether:

The central question is whether the treatment effect differs between the two biomarker groups.

That is an interaction question.

Formal Treatment-by-Biomarker Interaction

For a continuous outcome, a simple model might be:

Here:

• \(\beta_1\) represents the treatment effect when \(B=0\).
• \(\beta_2\) represents the biomarker-group difference under \(T=0\).
• \(\beta_3\) represents the treatment-by-biomarker interaction.

If:

there is evidence that the treatment effect differs according to biomarker status.

For Binary Endpoints

For a binary endpoint, logistic regression can be used:

logit{
  P(Y = 1)
}
=
β0 +
β1*T +
β2*Biomarker +
β3*T*Biomarker

The interaction coefficient \(\beta_3\) assesses whether the treatment effect on the log-odds scale differs between biomarker groups.

For Time-to-Event Endpoints

For survival endpoints, a Cox proportional-hazards model may include:

h(t) =
h0(t) *
exp(
  β1*T +
  β2*Biomarker +
  β3*T*Biomarker
)

The interaction coefficient provides a formal assessment of whether the treatment hazard ratio differs by biomarker subgroup under the model.

Why "Significant in One Group, Not Significant in the Other" Is Not Enough

Consider:

It is tempting to conclude that treatment works only in the biomarker-positive group.

That conclusion can be incorrect.

The two p-values answer different questions:

• Is treatment effective within the positive group?
• Is treatment effective within the negative group?

Neither directly answers:

That requires an interaction or another appropriate heterogeneity analysis.

Three Major Biomarker Trial Strategies

Biomarker-driven trials can be broadly organized into several designs.

All-Comers Design

In an all-comers design, both biomarker-positive and biomarker-negative patients are eligible.

The biomarker may be measured at baseline and used for subgroup analysis.

The primary analysis can target the entire population:

while biomarker-defined effects may be secondary or supportive analyses.

This design can be useful when the treatment is expected to benefit a broad population or when evidence for biomarker restriction is not sufficiently established.

Biomarker-Stratified Randomization

Suppose biomarker-positive patients are expected to have different baseline risk from biomarker-negative patients.

Simple randomization could, by chance, produce an imbalance in biomarker composition between treatment arms.

Stratified randomization reduces this risk by randomizing separately within biomarker strata.

For example:

The result is better treatment balance within each important biomarker stratum.

Stratified Randomization Does Not Prove Predictive Value

This distinction is easy to overlook.

If a trial stratifies randomization by biomarker status, that means investigators considered biomarker status important for treatment assignment balance.

It does not mean the biomarker has been established as predictive.

Enrichment Designs

An enrichment design limits enrollment to a selected biomarker population.

For example:

USUBJID
TRT01P
BIOMARK
BIOMARKN
BIOMARKGR
AVAL
PARAMCD
ADT
CNSR

coxph(
  Surv(time, event) ~
    treatment +
    biomarker +
    treatment:biomarker +
    age +
    stage,
  data = analysis
)

coxph(
  Surv(time, event) ~ treatment +
    strata(biomarker),
  data = analysis
)

fit <- glm(
  response ~ treatment * biomarker,
  family = binomial(),
  data = analysis
)

summary(fit)

treatment:biomarker

library(survival)

fit <- coxph(
  Surv(time, event) ~
    treatment * biomarker,
  data = analysis
)

summary(fit)

outcome ~ treatment + biomarker + baseline_covariates

outcome ~ treatment * biomarker + baseline_covariates

analysis$biomarker <-
  ifelse(
    analysis$biomarker_score >= 50,
    "Positive",
    "Negative"
  )

analysis$biomarker <-
  factor(
    analysis$biomarker,
    levels = c("Negative", "Positive")
  )

library(dplyr)

analysis %>%
  group_by(
    treatment,
    biomarker
  ) %>%
  summarise(
    N = n(),
    responders = sum(response == 1, na.rm = TRUE),
    response_rate =
      responders / N
  )

fit_logistic <- glm(
  response ~ treatment * biomarker,
  data = analysis,
  family = binomial()
)

summary(fit_logistic)

treatment:biomarker

library(survival)

fit_cox <- coxph(
  Surv(time, event) ~
    treatment * biomarker,
  data = analysis
)

summary(fit_cox)

forest_data <- data.frame(
  subgroup = c(
    "Overall",
    "Biomarker Positive",
    "Biomarker Negative"
  ),
  estimate = c(
    0.72,
    0.52,
    0.94
  ),
  lower = c(
    0.60,
    0.39,
    0.74
  ),
  upper = c(
    0.86,
    0.69,
    1.20
  )
)

analysis$treatment_biomarker <-
  interaction(
    analysis$treatment,
    analysis$biomarker
  )