BATTLE Biomarker-Matched Allocation

Why BATTLE Used Biomarker-Matched Allocation

Advanced non-small-cell lung cancer (NSCLC) is biologically heterogeneous: a treatment that may help one molecular subgroup need not help another. The BATTLE trial—Biomarker-integrated Approaches of Targeted Therapy for Lung Cancer Elimination—addressed this problem in patients with advanced NSCLC by combining pretreatment tumor biopsies, biomarker-defined groups, and adaptive randomization among targeted regimens.

The design’s central idea was not to assign every patient deterministically to a treatment labeled as the “match.” Instead, biopsy results placed a patient in a biomarker group, and that group helped determine the probabilities of random assignment to available regimens. As disease-control outcomes accumulated, allocation probabilities could favor regimens with more promising results in that group. This preserved a randomized allocation process while making it responsive to emerging evidence.

BATTLE was a phase II study in previously treated advanced NSCLC. It used pretreatment biopsies and four biomarker-defined groups involving EGFR, KRAS, VEGF, and RXR/Cyclin D1. The four treatment regimens were erlotinib, sorafenib, vandetanib, and erlotinib plus bexarotene. The primary outcome used for adaptation was disease control at eight weeks, generally classified as complete response, partial response, or stable disease at that assessment. The trial began with equal allocation and adapted allocation probabilities as outcome information accrued.

Biomarker Grouping and the Allocation Rule

The biopsy was consequential because it supplied the information needed to place a patient in the trial’s biomarker structure before treatment assignment. The group was not itself an outcome, and an association between a marker and response would not by itself establish that the marker caused a treatment effect. The design instead used biomarker information to organize treatment comparisons and tailor allocation.

Let \(G_t\) denote the biomarker group for the \(t\)th assigned patient, \(A_t\) the assigned treatment, and \(H_t\) the available trial history at that assignment. The history can include earlier patients’ assignments and observed outcomes, as well as other information used by the allocation rule. A valid sequential randomized assignment has probabilities \(P(A_t=a\mid G_t,H_t,C_t)\), where \(C_t\) represents any additional assignment covariates used by the protocol. For every treatment \(a\) eligible in that group, the probability must be positive.

Here \(Y_t(a)\) is the outcome the current patient would have under treatment \(a\), and \(\mathcal{A}_{G_t}\) is the set of treatments eligible for that patient’s group. The independence is conditional on the information governing assignment, not automatically on biomarker group alone. If earlier outcomes affect allocation and outcomes also change over calendar time, then assignment can be associated with current patients’ potential outcomes when the history is ignored. The conditional statement is the relevant randomization property.

Adaptation changes the probabilities, not the definition of randomization. A treatment can become more likely in a group because the history makes it look promising, while other eligible treatments retain nonzero probability. This overlap allows continuing randomized comparisons. The details of an allocation algorithm—such as how evidence is translated into probabilities or whether probabilities are bounded away from zero—must be specified in advance for a particular trial; the calculations below illustrate the principle and are not a reconstruction of BATTLE’s exact algorithm.

How Outcome Evidence Can Change Allocation

A simple way to understand the Bayesian updating behind response-adaptive allocation is to treat eight-week disease control as a binary outcome: success if disease control is observed, and failure otherwise. For a treatment in one biomarker group, suppose the unknown disease-control probability is \(p\), and use a Beta prior. After observing \(x\) successes among \(n\) patients, a Beta\((a,b)\) prior gives a Beta\((a+x,b+n-x)\) posterior. Its mean is \((a+x)/(a+b+n)\).

That posterior mean summarizes estimated disease control, but an adaptive design might use a different summary, such as the posterior probability that a treatment is best. Either way, the summaries are calculated within the relevant group. The adaptation rule then maps those summaries into assignment probabilities. The rule is a design choice: a high estimated response rate does not, by itself, dictate a unique allocation probability.

Why Conditional Randomization Matters

The justification for randomized comparisons can be stated directly. At each assignment, the randomization mechanism selects a treatment using its prespecified probability conditional on the available information. Because the selection is randomized within that information set, potential outcomes for the current patient are not used to choose the assigned treatment. Consequently, comparing outcomes by assigned treatment can support causal interpretation, subject to adherence, outcome ascertainment, and the trial’s analysis plan.

The role of history in that statement is important. If the allocation probability changes after favorable outcomes, then the probability of receiving a treatment varies across assignment times. A crude comparison that ignores the design may be misleading if prognosis also changes over time. The design and analysis must preserve the assignment history—for example, by accounting for the actual randomization scheme and any planned adjustment for time or stratification factors.

Positivity is also important. If a regimen’s probability becomes exactly zero in a group, later patients in that group cannot contribute randomized evidence about that regimen. Conversely, retaining positive probabilities supports comparison but may mean that some patients receive a treatment with a lower current estimated success rate. Adaptive allocation therefore balances learning about treatments against allocating more patients to options that appear promising.

A Short Justification for the Updating Step

For a binary disease-control outcome and a Beta\((a,b)\) prior, the likelihood from \(x\) successes and \(n-x\) failures is proportional to \(p^x(1-p)^{n-x}\). The prior density is proportional to \(p^{a-1}(1-p)^{b-1}\). Multiplying gives a posterior density proportional to \[ p^{a+x-1}(1-p)^{b+n-x-1}, \] which is the kernel of a Beta\((a+x,b+n-x)\) distribution. Its mean is therefore \((a+x)/(a+b+n)\). This derivation explains how accumulated group-specific outcomes can be summarized for an adaptive rule. It does not establish that a particular rule is well calibrated or that its resulting estimates have a desired operating characteristic; those properties require design evaluation, commonly by simulation.

Operational Requirements, Diagnostics, and Limitations

The biopsy-and-adaptation strategy creates practical dependencies that are especially important in an advanced-cancer population. Tissue must be obtained safely, specimens must be adequate for assay, biomarker results must be returned in time, and the group-classification rule must be applied consistently. Delayed or failed assays can disrupt assignment and may selectively exclude patients if biopsy success is related to prognosis or disease characteristics.

Useful monitoring includes the fraction of enrolled patients with usable biopsy results, time from biopsy to group assignment, group-specific enrollment counts, actual assignment probabilities, and outcomes by group and treatment. Investigators should check that recorded assignments agree with the probability schedule in force at each assignment and that every eligible treatment retains the required probability. They should also examine whether outcome availability or assessment timing differs across treatments or groups, since differential missingness can distort the apparent evidence driving adaptation.

Adaptive randomization is not a substitute for careful design. Small group sizes can make response estimates unstable; early random variation can steer allocation toward a treatment that is not truly better. Changing probabilities can complicate simple analyses that assume constant allocation. A biomarker may be prognostic—associated with outcomes regardless of treatment—or predictive—associated with differences in treatment effects. A response pattern within a group alone does not necessarily distinguish these roles, particularly without sufficiently informative comparisons and a suitable analysis.

Compared with fixed equal randomization, adaptive allocation may direct a greater share of later patients toward treatments with promising accumulated evidence, but it produces time-varying assignment probabilities and can reduce balance across treatment groups. Compared with deterministic biomarker matching, it retains randomized treatment comparisons and does not require certainty that each marker-treatment pairing is correct. Compared with a conventional treatment-by-biomarker factorial design, a biomarker-stratified adaptive design focuses allocation within biologically defined groups but may leave some treatment-group comparisons relatively sparse.

The design’s value depends on the clinical question and implementation. An eight-week disease-control endpoint is available sooner than many survival outcomes, making it useful for adaptation, but it is not identical to long-term benefit. A trial’s conclusions also depend on eligibility, assay definitions, treatment availability, endpoint assessment, and the statistical analysis. BATTLE illustrates a strategy for integrating biological information into randomized treatment allocation; it does not imply that every biomarker-guided design should use the same groups, endpoint, or adaptation rule.

Common Mistake