Adaptive clinical trial designs allow prospectively planned changes based on accumulating data, including futility stopping, dose selection, sample-size modification, and population enrichment.
Bayesian methods can support interim decisions, evidence borrowing, response-adaptive randomization, subgroup analysis, and primary inference in drug and biologic trials.
Basket, umbrella, and platform trials use master protocols to evaluate multiple therapies, diseases, subtypes, or research questions within coordinated trial structures.
Regulatory acceptance of adaptive and complex innovative trial designs depends on prespecification, simulation, error control, data integrity, and early agency engagement.
Traditional clinical trials generally establish their principal features before enrollment and retain them through the end of the study. Sample size, treatment allocation, population, endpoints, and analysis methods are usually set in advance, even when important uncertainties remain. An adaptive clinical trial also begins with a detailed plan, but that plan permits prospectively defined modifications based on data collected during the study. The possible changes, the information used to make them, and the rules governing each decision are established before comparative results become available.1
This distinction separates legitimate adaptation from an unplanned response to disappointing or unexpected findings. Changes introduced after emerging results are known may create bias, inflate the probability of an erroneous conclusion, or make the final evidence difficult to interpret. An adaptive design instead anticipates selected uncertainties and establishes a controlled process for responding to them. The flexibility resides within the protocol and statistical strategy rather than in the discretion to revise the trial whenever circumstances change.
Adaptations may include stopping enrollment for futility or efficacy, revising sample size, discontinuing a treatment group, selecting a dose, changing randomization probabilities, or refining the population under study. The suitability of any adaptation depends on the development question, the timing and reliability of the interim data, and the effect of the decision on the trial’s statistical properties. Simpler adaptations may be evaluated analytically, while more complex designs often require extensive simulation to determine how they perform across a range of plausible scenarios.
1,2
The International Council for Harmonisation (ICH) is also developing E20, a guideline focused on confirmatory clinical trials that permit prespecified modifications based on interim analyses. As of June 2026, E20 remains a Step 2b draft rather than a final harmonized guideline. Its development nonetheless reflects broader regulatory interest in defining when adaptive trials can produce reliable and interpretable evidence within an overall development program.3
Adaptive trials form part of a wider change in clinical development. Bayesian statistical methods can update probabilities and support decisions as evidence accumulates. Master protocols can coordinate multiple therapies, populations, or research questions within a common framework. Basket, umbrella, and platform trials apply that framework in different ways. These concepts often appear together, but they are not interchangeable, and each addresses a different aspect of trial design.
Separating Adaptive Designs, Bayesian Methods, and Master Protocols
An adaptive design describes how specified features of a trial may change in response to accumulating data. Bayesian methodology describes a statistical framework that can support decisions or inference. A master protocol describes the organizational structure under which multiple questions or substudies are pursued. A single clinical program may combine all three, but none necessarily requires the others.
Bayesian analysis begins with a probability model that may incorporate relevant prior information. Current study data update that information to produce a posterior distribution, which expresses the remaining uncertainty about a treatment effect or another quantity of interest. Bayesian probabilities can be used at interim analyses to support decisions concerning futility, success, dose selection, subgroup performance, or treatment allocation. They can also provide the primary inferential framework for conclusions about effectiveness or safety.4
The Food and Drug Administration’s (FDA’s) January 2026 draft guidance is significant because it addresses Bayesian methodology specifically in drug and biological product trials, including its use for primary inference. Earlier applications of Bayesian methodology were especially visible in medical-device development, but the new draft provides drug- and biologic-specific recommendations concerning decision criteria, prior information, operating characteristics, and the interpretation of results. The document remains draft guidance and does not establish a preference for Bayesian over frequentist methods.4
Bayesian methodology is also not synonymous with adaptive design. A trial may use Bayesian probabilities to trigger an interim adaptation but employ frequentist methods for its final analysis. Conversely, a Bayesian primary analysis does not necessarily require any midtrial modification. Sponsors must identify the role of the Bayesian component clearly because the regulatory and methodological questions differ depending on whether it supports an operational decision, a statistical conclusion, or both.1,4
A master protocol provides an overarching structure for evaluating one or more interventions across one or more diseases, conditions, or subtypes. It may include separate substudies with distinct objectives while using common protocol elements, infrastructure, governance, or control information. Master protocols may incorporate adaptive features, but the ability to study multiple questions under one framework does not, by itself, make a trial adaptive.5,6
These distinctions matter during protocol development. Trial architecture, adaptation rules, and inferential methods must be selected for their scientific purpose and then integrated. Applying an innovative label without establishing how each component contributes to the intended evidence can create unnecessary complexity without improving the development strategy.
Matching Basket, Umbrella, and Platform Structures to the Scientific Question
Basket, umbrella, and platform trials fall within the broad master-protocol family, but they organize research questions differently. The appropriate structure depends on whether the program begins with a treatment hypothesis that spans diseases, several therapeutic hypotheses within one disease, or a continuing need to evaluate interventions over time.
A basket trial typically evaluates one therapeutic strategy across multiple diseases, conditions, or disease subtypes. In precision oncology, the organizing characteristic is often a molecular alteration shared by tumors arising in different tissues. The structure allows investigators to test whether activity associated with a target extends beyond a single conventional disease category while retaining separate evaluation of the participating cohorts.5,6
A phase II study of vemurafenib (now marketed as Zelboraf) illustrates this approach. The trial enrolled patients with several nonmelanoma cancers containing BRAF V600 mutations. Its histology-independent structure allowed a common targeted hypothesis to be examined across tumor types, but outcomes were assessed within disease-specific cohorts rather than assuming that the shared mutation would produce an identical response in every cancer.7
That distinction is central to the interpretation of basket trials. A biomarker may provide a rational basis for grouping patients, but the surrounding biology can vary by tissue, disease state, prior treatment, and co-occurring molecular features. A basket design can reveal where a common hypothesis appears promising and where it does not, but it does not establish that the biomarker has the same predictive significance in every setting.
Umbrella trials reverse the orientation. They evaluate multiple therapies or treatment strategies within a single disease, often assigning patients to substudies according to biomarkers or other disease characteristics. A common screening process can direct participants toward relevant treatment hypotheses, while the master protocol coordinates multiple substudies within the same disease-focused framework.5,8
Not every umbrella trial uses the same control group, randomization method, or adaptation strategy. The category describes the relationship among the disease, its subgroups, and the therapies under evaluation. Individual substudies may operate differently depending on their objectives and stage of development.
Platform trials emphasize a continuing infrastructure. Multiple interventions can be evaluated within the platform, and treatment groups may enter or leave according to protocol-defined procedures. Common systems, sites, governance, and data processes can support successive research questions without requiring every intervention to begin with an entirely separate trial structure. The European Medicines Agency (EMA) has recognized that platform trials may be intended to provide pivotal evidence and has identified methodological issues that require attention when they are used for confirmatory purposes.9
A platform may share control information among treatment groups, use adaptive allocation, or permit the introduction of new interventions, but these are possible features rather than universal requirements. The protocol must specify which participants contribute to each comparison, how changes over time will be addressed, and how the addition or removal of an arm affects the remaining analyses.
The structural choice should follow the scientific question. A basket trial is suited to a hypothesis that crosses diseases or subtypes. An umbrella trial coordinates several treatment hypotheses within one disease. A platform provides a durable framework for evaluating multiple interventions or successive questions. Combining these structures with adaptive or Bayesian methods can expand their capabilities, but added complexity is useful only when it serves a defined development objective.
Changing Development Strategy Through Planned Decisions
Drug development involves uncertainty about treatment activity, dose, population, sample size, and the value of continuing a study. A conventional design often resolves these questions through a sequence of separate trials or after the current study is complete. An adaptive design may address selected uncertainties through prespecified decisions during the trial, provided that the necessary data become available in time and the effects of each decision are understood.
Futility monitoring can allow enrollment to stop when interim evidence indicates that the trial has little probability of achieving its objective. An efficacy boundary may permit earlier stopping when a predefined level of evidence has been reached. Bayesian probabilities can support either type of decision, although other statistical approaches are also available.1,10
These options do not mean that adaptive trials necessarily end earlier. A trial may continue to its maximum sample size because the interim results do not cross any decision boundary. It may even require more enrollment than initially expected if a prespecified sample-size adaptation is triggered. Efficiency is therefore a possible result of the design rather than an inherent property of the adaptive label.
Dose selection offers another important application. An adaptive trial may discontinue doses that do not meet predefined criteria, focus later enrollment on more promising doses, or use accumulating information to select doses for subsequent studies. Bayesian modeling may support these choices by updating the probability that a dose meets specified efficacy or safety objectives.4
Population enrichment uses interim information to refine the population in which treatment continues to be evaluated. A design might begin with a broad population and permit enrollment to focus on a prespecified subgroup if accumulating evidence supports that decision. Because such changes can affect the interpretation and generalizability of the results, the biological rationale, subgroup definitions, decision thresholds, and statistical consequences require careful planning. An enrichment strategy should not become an unplanned search for a favorable subgroup after the overall findings appear weak.
Response-adaptive randomization changes treatment-assignment probabilities as outcomes accumulate. Participants entering later in the trial may have a higher probability of assignment to a treatment that appears more promising for their subgroup, although adequate information must still be collected for valid comparisons. The method requires outcomes or reliable intermediate measures to become available quickly enough to inform subsequent assignments.
The I-SPY 2 trial provides a well-documented example. This standing, multicenter phase II platform evaluates experimental regimens in combination with standard neoadjuvant therapy for high-risk breast cancer. Biomarkers classify patients into defined subtypes, and Bayesian adaptive randomization updates assignment probabilities according to accumulating evidence concerning regimen performance within those subtypes.11
The platform is designed to identify treatment and biomarker combinations that warrant further investigation. Its results can guide subsequent development, but identification of a promising regimen within the platform is distinct from establishing the evidence required for marketing authorization. The regulatory role of an adaptive trial must be defined at the outset because exploratory screening, dose selection, and confirmatory inference create different evidentiary demands.
Bayesian evidence borrowing can also influence development strategy. Prior information may come from earlier trials, related populations, external controls, nonconcurrent controls, or relevant real-world evidence. The current data then update that information rather than being analyzed in isolation.
The usefulness of borrowing depends on whether the external information is sufficiently relevant and compatible with the current trial. Differences in eligibility, endpoint assessment, background treatment, calendar time, or other features may limit comparability. The analysis must also show how strongly the prior information influences the result and how the conclusion changes under alternative assumptions. A larger historical data set does not automatically constitute a suitable prior.
Establishing Regulatory Confidence
Regulatory pathways for adaptive and complex trials have become more explicit. The FDA has issued final guidance on adaptive designs and on sponsor interactions concerning complex innovative designs. It also operates a paired-meeting program for qualifying complex innovative design (CID) proposals. Final oncology-specific master-protocol guidance is available, while broader master-protocol guidance and Bayesian guidance for drug and biological product trials remain drafts. ICH E20 also remains in draft form.1–4,6,8,12
The existence of these materials does not constitute blanket acceptance. A regulatory agency may recognize a methodology while determining that a particular implementation does not provide adequate evidence. The evaluation centers on whether the design addresses a meaningful scientific question and produces results that remain reliable, interpretable, and relevant to the intended regulatory decision.
Prespecification forms the foundation. The protocol and supporting documents should establish the timing of interim analyses, the data used for each decision, the available adaptations, and the criteria that trigger them. They should also explain who will conduct the analysis, who may receive unblinded information, how the decision will be implemented, and how the final analysis accounts for the adaptations. The level of detail required depends on the complexity and consequences of the potential changes.
Prespecification does not make an unsuitable design acceptable. A rule may be defined clearly yet still inflate type I error, reduce power under plausible conditions, introduce bias, or generate estimates that are difficult to interpret. Sponsors must evaluate how the complete design behaves rather than treating the adaptation algorithm as an isolated feature.
For many complex designs, simulation provides the principal means of assessing that behavior. The FDA asks CID sponsors to characterize operating properties, such as type I error, power, expected sample size or duration, and estimation performance under multiple parameter configurations. The simulation plan should explore scenarios in which the assumed treatment effect is present, absent, smaller than expected, or distributed differently across populations. It should also examine the consequences of assumptions that prove inaccurate.2,12
Simulation can guide design development before it becomes a regulatory submission issue. Alternative decision boundaries, allocation rules, sample-size limits, or borrowing strategies can be compared under the same scenarios. Investigators can identify conditions in which an apparently attractive design performs poorly and refine the protocol before enrollment begins.
A credible simulation package must correspond to the design that will actually be implemented. Changes to decision rules, data timing, population definitions, or statistical models may alter the operating characteristics and require additional evaluation. Assumptions, code, parameter settings, and outputs should be sufficiently transparent to permit review and reproduction.
Control of bias and error remains central, but it does not exhaust the regulatory evaluation. Regulators must also be able to understand which population the result applies to, what treatment contrast was estimated, how adaptations influenced the data, and whether the final analysis answers the clinical question originally posed. A design may control a specified error rate and still produce evidence that is not clinically interpretable.
Early interaction can help sponsors identify these issues before operational commitments make major revisions difficult. The FDA’s CID process allows discussion of technical questions related to modeling, simulation, and novel trial methods. Participation provides an avenue for focused feedback, but it does not constitute approval of the protocol or a commitment concerning the eventual application.
For global programs, sponsors must also consider differences among regulatory frameworks. FDA, EMA, and ICH materials address overlapping principles, but they differ in scope, terminology, and status. A design discussed with one authority should not be assumed to satisfy every other agency involved in a multiregional program.
Managing Operational and Platform-Specific Risks
The statistical plan cannot succeed if the trial cannot deliver sufficiently complete, timely, and reliable data for each decision. Adaptive designs often require closer coordination among clinical operations, data management, biostatistics, programming, laboratories, supply functions, and oversight bodies than a fixed trial with a single final analysis.
Interim information creates a particular risk. Knowledge that one treatment appears to be performing better, that an arm may soon close, or that the sample size may increase can influence recruitment, investigator behavior, endpoint assessment, treatment management, or business decisions. The FDA therefore asks CID sponsors to describe data-access plans and measures intended to protect trial integrity.12
The appropriate governance structure depends on the design. Independent data monitoring committees, separate unblinded statistical teams, adaptation committees, and operational firewalls may help restrict information, but no single arrangement suits every trial. Responsibilities and decision authority must be clear, and the people managing day-to-day operations should receive only the information needed to implement an approved adaptation.
Data timing is equally important. A decision based on delayed, incomplete, or inconsistently assessed outcomes may not reflect the patients enrolled at the time it is made. Biomarker-driven trials also depend on reliable classification within timelines that support enrollment and randomization. Data entry, query resolution, central review, assay performance, and endpoint maturity become components of the adaptation process rather than routine downstream activities.
Platform trials add further challenges because the infrastructure may persist while treatments, participants, and external conditions change. The protocol must govern the addition and discontinuation of arms, updates to common procedures, safety communication, responsibilities among collaborators, and the effect of amendments on ongoing substudies. Randomization systems, informed-consent materials, investigational-product supply, site training, contracts, and electronic data capture may all require coordinated changes when an arm enters or exits.
Multiplicity can arise when a platform evaluates several treatments, populations, endpoints, or hypotheses. The relevant error-control strategy depends on the relationships among the comparisons and their intended regulatory roles. EMA has identified multiplicity as a methodological consideration for platform trials, but this does not imply that every platform requires one universal familywise error calculation across all current and future substudies.9,13
Control-group strategy also requires close attention. Concurrent controls enroll during the same relevant period as the intervention group. Nonconcurrent controls entered the platform earlier and may have been treated under different conditions. Bayesian methods can incorporate nonconcurrent or external information, but the analysis must address whether the data remain comparable.10,13
Long-running platforms may experience changes in standards of care, diagnostic practices, background therapy, eligibility patterns, geographic participation, and patient characteristics. These temporal changes can confound comparisons if outcomes improve or worsen for reasons unrelated to the investigational treatment. Randomization against an appropriate concurrent control can reduce some of these concerns, while the use of earlier controls requires additional justification and sensitivity analysis.
The operational demands should influence design selection from the beginning. A technically sophisticated adaptation may offer little value if endpoint data arrive too slowly, sites cannot implement changes consistently, or governance arrangements permit interim information to affect conduct. Feasibility must encompass the full system required to execute the design, not only the statistical method.
Building Regulatory and Organizational Readiness
An adaptive strategy should begin with the decision the development program needs to support. Sponsors should identify which uncertainties may be resolved during the trial, which adaptations would be scientifically useful, and how the resulting evidence will contribute to later development or a regulatory submission. Starting with the decision prevents the methodology from becoming an end in itself.
Cross-functional planning should begin earlier than it might for a conventional fixed design. Clinical development and biostatistics must align the adaptation with the clinical question. Regulatory teams must determine how agencies are likely to view the proposed evidence. Operations and data-management groups must confirm that interim information will be available and reliable. Biomarker teams and laboratories must meet decision-relevant timelines. Supply functions must prepare for changes in enrollment or allocation. Programming and technology teams must implement the algorithm exactly as evaluated.
The infrastructure should match the adaptation. A blinded sample-size reassessment may require relatively limited changes to trial operations. Response-adaptive randomization requires timely outcomes, validated calculations, and randomization systems capable of applying updated probabilities. Population enrichment requires reliable classification and a plan for participants already enrolled outside the continuing subgroup. A platform that adds and removes arms requires durable governance, flexible systems, and sites prepared for repeated updates.
External partners may provide specialized statistical, operational, laboratory, or technology capabilities, but sponsors retain responsibility for ensuring that the components work together. A contract research organization (CRO) or statistical partner should be able to reproduce the adaptation algorithm, conduct the necessary simulations, and document implementation. Randomization and data systems must support planned changes without exposing sensitive interim information. Laboratories must deliver biomarker results within the time assumed by the design. Sites need training that can accommodate amendments while preserving consistency across substudies.
Partner evaluation should therefore extend beyond prior participation in a trial described as adaptive. Sponsors need to understand which design features the partner implemented, how decisions were governed, whether systems were validated for changing allocations or arm availability, and how traceability was maintained. Experience with a simple sample-size reassessment does not necessarily demonstrate readiness to operate a Bayesian platform with multiple biomarkers and changing treatment groups.
Regulatory engagement should proceed alongside this organizational planning. Early feedback may reveal that additional simulations, a different control strategy, tighter adaptation rules, or a revised evidentiary role is needed. Addressing those questions before the protocol, systems, supply plan, and site network are fixed can prevent a statistically defensible concept from becoming operationally unworkable.
Adaptive Does Not Mean Improvised
Adaptive designs formalize how a clinical trial may respond to accumulating evidence. Bayesian methods can support interim decisions, evidence borrowing, treatment allocation, subgroup analysis, or primary inference. Master protocols organize multiple research questions within a common framework, while basket, umbrella, and platform trials align that framework with different scientific objectives.
None of these approaches inherently guarantees a shorter trial, a smaller sample, or a lower development cost. Their value depends on whether the chosen design addresses a genuine uncertainty and whether the program can implement it without undermining statistical validity, operational integrity, or clinical interpretation.
Regulatory agencies have created clearer routes for discussing and evaluating innovative designs, but acceptance remains conditional. Sponsors must establish the scientific rationale, specify the decision rules, characterize operating performance, protect interim information, address multiplicity and control-group questions, and explain how the final results support the intended regulatory conclusion.
The strongest adaptive programs define the boundaries of flexibility before enrollment begins. They identify which decisions may change, the evidence that will trigger each change, the people authorized to act, and the methods that will preserve the credibility of the final result.
