Basket Trials: Review of Current Practice and Innovations for Future Trials

Abstract

Advances in biology and immunology have elucidated genetic and immunologic origins of cancer. Innovations in sequencing technologies revealed that distinct cancer histologies shared common genetic and immune phenotypic traits. Pharmacologic developments made it possible to target these alterations, yielding novel classes of targeted agents whose therapeutic potential span multiple tumor types. Basket trials, one type of master protocol, emerged as a tool for evaluating biomarker-targeted therapies among multiple tumor histologies. Conventionally conducted within the phase II setting and designed to estimate high and durable objective responses, basket trials pose challenges to statistical design and interpretation of results. This article reviews basket trials implemented in oncology studies and discusses issues related to their statistical design and analysis.

INTRODUCTION

Efforts to ascertain the genetic and immunologic origins of cancer extend treatment options for patients beyond cytotoxic chemotherapies to biomarker-targeted treatments that act on distinct mechanisms of tumorigenesis. Advances in sequencing technologies coupled with large-scale public projects such as The Cancer Genome Atlas led to the recognition of recurrent alterations that occur across cancer types of distinct organs of origin and histology. Several recent genetic and immune phenotypic characterizations of cancer transcend traditional criteria for cancer classification on the basis of tissue histology.¹ A few tissue-agnostic therapies have demonstrated robust clinical efficacy spanning multiple tumor types.^2,3 The therapeutic translation of the microsatellite instability high phenotype pioneered a new regulatory pathway for biomarker-guided cancer treatments.^4,5

CONTEXT

• Key Objective

• To elucidate key statistical considerations of basket trials in oncology settings.

• Knowledge Generated

• Designs of basket trials are reviewed, and pertinent regulatory guidance on their application to oncology trials is discussed. Additionally, methodology developed for basket trials is presented alongside issues related to their statistical design and analysis.

• Relevance

• Innovations in molecular biology and immunology have revealed that distinct cancer histologies share common genetic and immune phenotypic traits. Technological developments have made it possible to target these alterations, yielding novel classes of cancer therapeutics that span traditionally distinct cancer subtypes. Basket designs accommodate enrollment of multiple tumor histologies providing an important tool for prospective clinical inquiry of target therapies.

Endeavoring to extend the promise of precision medicine by way of effective anticancer therapies for patients, drug developers in oncology have increasingly emphasized molecularly targetable features that span multiple tumor types. Currently, there are three biomarkers and four drugs for which histology-agnostic approvals have been granted by the US Food and Drug Administration (FDA).^6-8 Evaluated over the course of three phase I/II trials enrolling patients with neurotrophic tyrosine receptor kinase (NTRK) fusion-positive tumors, larotrectinib yielded impressive efficacy across 17 tumor types—overall response rate was 80%, with 71% ongoing after 12 months—leading to FDA approval in 2018.⁹ Entrectinib was tested in three phase I/II trials in patients with NTRK fusion-positive tumors leading to an FDA approval in 2019.¹⁰ Dostarlimab was approved for treatment of microsatellite instability high cancer in 2021¹¹ after a phase I trial yielded promising results. In 2020, the FDA approved pembrolizumab for solid tumors with high tumor mutational burden. Several additional tumor-agnostic therapies are in development, including therapies targeting mutations in the ret proto-oncogene and neuregulin 1 gene.^12,13

These recent innovations in pharmacology and regulatory policy, however, challenge the statistical assumptions that provide the foundations for analysis and trial design using traditional techniques. The tumor-agnostic designation implies that a treatment is effective for a biomarker subpopulation without regard to tissue of origin. Yet, patient prognosis for progression-free survival and overall survival vary to a considerable extent across tumor types and clinical stages with standard-of-care chemotherapies.^14,15 Moreover, the statistical null expectation for experiencing a future overall response following treatment with standard of care may differ substantially among tumor histologies, especially in the refractory setting. In the presence of enrollment to many heterogeneous tumor types, statistical estimation of the drug effect may be limited by pooled analyses; yet, enrollment may be too diffuse to draw conclusions that pertain to histology-specific activity.¹⁶ For example, trials of biomarker-guided treatments targeting human epidermal growth factor receptor (HER)2 amplification and BRAF-mutant tumors have demonstrated heterogeneous efficacy across enrolled tumor types.¹⁷ Contemporaneous early-phase oncology trials have expanded in size and scope, progressively adopting broader inclusion criteria and frequently enrolling expansive dose-expansion cohorts with an aim to evaluate efficacy as an implicit objective,¹⁸ often without a prespecified statistical design.

To counteract these limitations, master protocols have evolved to extend clinical design to complex settings in oncology.¹⁹ Woodcock and LaVange define master protocols as methodologic innovations that facilitate formal, prospective coordinated inquiries of multiple therapies and/or multiple patient disease profiles within the same clinical trial. The FDA defines a basket trial as a master protocol study designed to test a single investigational drug or drug combination in different populations defined by disease stage, histology, number of prior therapies, genetic or other biomarkers, or demographic characteristics.²⁰ Referring to baskets as substudies, FDA guidance also notes that strong response signal seen in a substudy may allow for expansion to generate further data to support regulatory approval. This article reviews basket trials implemented in oncology studies. We emphasize issues related to their statistical considerations as well as briefly summarize guidance provide by US regulators.

BASKET TRIALS IN ONCOLOGY

In oncology, basket trials are often open-label, single-arm trials conducted in the phase II setting.²¹ Figure 1 summarizes basket trial methodology as well as describes challenges and innovations devised for characterizing tumor heterogeneity. Recent developments, however, have begun to expand their scope. In their systematic review, Park et al describe four basket trials incorporating random assignment and two basket trials designed for the phase III setting.²⁴ In a few instances, evidence from basket trials has led to regulatory approval, as discussed above.²⁵ Table 1 summarizes several pivotal basket trials implemented in the oncology setting. Enrolling a minimum of 54 patients and maximum of 498, most studies were designed to evaluate objective response rate as the primary end point. The primary statistical analysis for 54% (7/13) of studies pooled data among cancer histologies, whereas the remaining trials were designed for histology-specific analysis. The studies enrolled a median of 17 unique cancer types that ranged from 10 to 40, with each tumor type contributing 7.6 patients on average.

FIG 1.

Innovations in cancer biology have revealed that distinct cancer histologies share common genetic and immune phenotypic traits. Basket trials facilitate the study of biomarker-targeted therapies in multiple tumor types. Participants for whom the targeted feature is present are assigned to one or more therapies and followed for clinical outcomes. Basket trials are often open-label, single-arm trials conducted in the phase II setting. Basket trials circumvent the need to conduct separate trials for individual histologies. The methodology accommodates the study of rare tumor types, which can be included as baskets. Statistical evaluations of outcomes for rare tumors can be placed into the context of more prevalent cancers. Exceptional responses observed within individual tumor types can trigger pathways for rapid expansion and/or pursuit of pathways for accelerated regulatory approval. Robust responses observed among multiple tumor types can trigger the pursuit of pathways for tissue-agnostic regulatory review. Trialists choose to implement basket trials because they hold the expectation that the targeted feature is specific for delineating patients for whom the drug is efficacious in a manner that transcends tumor histology for the selected tumor types. Innovations in Bayesian statistical modeling on the basis of the multisource exchangeability model^22,23 can be used to quantify the extent of heterogeneity that is evident among tumor types as well as measure the evidence for tissue-agnostic efficacy. Bayesian posterior decision making, arising from evaluating the posterior probability that each tumor type exceeds the null hypothesized level of efficacy, can be applied at the final analysis or at interim futility analyses devised to halt further enrollment of underperforming histologies.

TABLE 1.

Summary of Select Oncology Basket Studies

A recently published meta-analysis demonstrated that the number of master protocols has increased dramatically, starting in 2013.²⁴ Most master protocol studies were conducted in oncology (91%; n = 76/83), and the majority used a basket design (59%; n = 49/83). Basket trials were most often conducted in phase I or II settings (96%; n = 47/49) and were open-label (94%; n = 46/49). The median sample size was 205 patients, and the majority did not comprise a control group or random assignment (90%; n = 44/49). Of the randomized basket trials included in this analysis that used a control group, two were oncologic studies (IMPACT II and SHIVA). In those studies, the control arm received standard-of-care therapy left to the choice of the treating physician.^36,37

A more recent meta-analysis published in 2018 identified eight published oncology basket studies, encompassing 1,176 patients.³⁸ Interestingly, this analysis demonstrated that common cancers might be under-represented when compared with rarer tumor types. The most commonly represented cancer types in these trials were ovarian and fallopian tube cancers (19%), colorectal cancer (12%), and sarcoma (11%), whereas lung cancer was only the fifth most common. The overall response rate was 25% for the combined analysis of all trials published to that point. Of note, most trials evaluated drugs as single agents and considered the presence of a single molecular alteration for enrollment. Yet, recent data have suggested the potential for rational combinations in targeted therapy to improve outcomes, along with the value of considering co-occurring alterations to improve the prediction of benefit.^39,40 We envision that the number of basket trials will continue to expand in the coming years, with a growing number of trials evaluating novel targets, rational combinations, and antibody-drug conjugates, along with presenting enhanced entry criteria on the basis of consideration of co-occurring mutations.

METHODOLOGY FOR BASKET TRIALS

Methods for design and statistical analysis of basket trials have evolved⁴¹ with their applications. Efforts to improve their potential for evaluating biomarker-guided therapies with tumor-agnostic potential comprise ongoing areas of methodology research. Traditional frequentist designs of clinical trials were devised to estimate averaged effects. Molecularly targeted therapies, however, may fail to offer sufficient efficacy to all cancer types. Therefore, basket trials pose challenges to the traditional paradigm for trial design and analysis, which assumes that individual patients who enroll in the same clinical study can be averaged.¹⁶ Cunanan et al was among the first to discuss the complexities inherent to applying basket trials to oncology studies.⁴² The authors also elucidated the scope of potential efficiency gains as well as losses from aggregation strategies devised to pool objective response data from similarly performing baskets.

In the absence of tumor-agnostic efficacy, trials designed for pooled analyses are sensitive to enrollment trends that are often uncontrolled by the design. Chance over-representations of an effective tumor histology may yield false-positive conclusions for sparsely enrolled tumor types. Conversely, an effective therapy may be deemed ineffective on the basis of imbalanced overenrollment of one or more ineffective histologies. Moreover, in the presence of heterogeneity, pooled estimates of outcomes may fail to describe the expected patient outcomes of any specific tumor type. Trials designed for basket-wise analyses often combine data from rare tumor types into a designated other basket. Yet, sparse enrollment may occur for even planned tumor types. Only five patients with cervical cancer enrolled in the SUMMIT trial,³⁵ whereas cervical, lung, prostate, and colon cancers were represented by three or fewer patients in the AKT inhibition trial.²⁶ The GARNET trial enrolled 11 tumor types contributing two or fewer patients.¹¹ Conclusions for such sparsely enrolled tumor types are impractical.

The challenges aforementioned compound further the difficulty of implementing efficient studies with sequential rules devised to adapt the trial to halt enrollment to ineffective tumor types when using standard designs. The SUMMIT trial evaluated Neratinib in patients with HER2- and HER3-mutant cancers comprising lung, breast, bladder, colorectal, biliary tract, endometrial, cervical, gastroesophageal, ovarian, and other.³⁵ Baskets were analyzed independently using the optimal Simon's two-stage design. Initial results supported the activity of neratinib only in the breast cancer cohort; yet, several baskets failed to reach enrollment targets by the time of interim analysis. This section discusses statistical considerations pertinent to the design and analysis of basket trials as well as highlights recent advances.

The Frequentist-Bayesian Dichotomy and Type I Error Control

Paradigms for statistical reasoning are defined by a dichotomy of frequentist versus Bayesian theory. Hypothesis testing in the frequentist paradigm uses P values that quantify the relative frequency of observing a statistical estimate as, or more extreme than that observed in the experiment. This conditional probability is conditioned on the particular experiment and null hypothesis. Having assumed a null hypothesis, which for basket trials is typically represented by an objective response rate that is too low to justify further study, two types of errors may occur. A type I error occurs when the trial data satisfy the criteria to reject the null hypothesis when the null is actually true. Conversely, a type II error occurs when the therapy is truly efficacious, but the trial data fail to reject the null hypothesis. Frequentists' decision making occurs through the application of P value thresholds (eg, P < .05). P value thresholds are calibrated, often under asymptotic assumptions, to control type I error rates.

By way of contrast, the Bayesian paradigm is founded on the perspective that probability should be defined in relation to one's existing belief. Before initiating an experiment, Bayesians require prior distributions for model parameters, which quantify existing knowledge or belief. A null hypothesis is not required. Bayesian inference rather occurs with respect to a posterior distribution. Posterior distributions are conditional on the observed data and prior, but not the experimental design. Moreover, posterior inference does not require the abstraction of asymptotic resampling, but rather conforms more naturally to human conceptions of probability. For example, using mathematical notation, Pr (π > 0.10 | Y) is the probability that objective response rate π exceeds 0.10 after having observed the data Y.

The Bayesian paradigm provides two approaches to decision making. Posterior decisions occur by applying thresholds to posterior probabilities. For example, after observing the trial data Y, one may make the decision that a trial yielded promising results if the posterior probability that objective response rate π exceeds 0.10 is > 0.95. This is denoted Pr (π > 0.10 | Y) > 0.95. This form of posterior decision making is conditional on the data, Y, observed so far in an experiment. Predictions of future outcomes Y* yet unobserved arise seamlessly under the Bayesian paradigm. Having observed interim data from a partially enrolled trial, the Bayesian paradigm provides the predictive probability that the trial would conclude successfully if the trial continues to full enrollment.^43,44 Bayesian decision making, whether posterior or predictive, requires calibration to control type I and type II errors at acceptable levels.

There are two types of type I errors that arise with basket trials: basket-wise and family-wise.^45,46 The basket-wise type I error rate, also called the marginal type I error rate, describes the rate at which a type I error occurs for an individual tumor type. Basket trials may enroll multiple tumor types for which the response rate is inadequate. The family-wise type I error rate considers all null baskets conjointly. It represents the false-positive rate for at least one of the null baskets, reflecting more stringent control.⁴⁷ Both frequentist hypothesis testing and Bayesian posterior and predictive decision making can be calibrated to control family-wise and basket-wise errors. An acceptable threshold for type I error will depend on the objective of the trial (eg, exploratory v confirmatory). Control for the family-wise type I error rate will provide better control against false positives than using the marginal type I error rate.

Strategy for Analysis

Before the evolution of master protocols, clinical trials were predominantly devised to examine the marginal, or population-level, response of a given intervention. In basket trial settings, the marginal response rate results from pooling outcomes observed for all tumor types, which represents the overall average among all indications. Although this approach may increase statistical power by leveraging the overall sample size, it assumes that the treatment understudy elicits agnostic (or global) effects. This fails to recognize the potential for heterogeneous treatment efficacy across tumor types. In the presence of heterogeneity and imbalanced enrollment (which is common among basket trials), this approach to analysis risks erroneous conclusions as outcomes from over-represented tumor types maybe extrapolated to individual histologies.

To confront this, trialists may prefer to rather analyze each tumor histology independently. Although this strategy avoids the assumption of global effects, it is limited by sample size and accrual, and one may need to adopt an analysis strategy that controls family-wise type I error among multiple hypothesis tests. With family-wise type I error control, any single basket with a false positive would be considered a failure of the design, which leads to more conservative decision making than basket-wise type I error control. Basket trials have often resulted in imbalanced enrollment among tumor types. Therefore, sample-size–dependent decisions (eg, interim monitoring for futility after so many participants have been enrolled) arise more naturally from independent analyses.

Necessitating the need to choose between pooling or splitting in advance of the trial, the primary analyses used in basket trials thus far are limited by the trialist's proficiency in accurately forecasting the presence versus absence of global effects. Trialists choose to implement basket trials because they hold the expectation that the targeted feature is specific for delineating patients for whom the drug is efficacious in a manner that transcends tumor histology for the selected tumor types. This expectation for tumor-agnostic efficacy is itself a hypothesis. One should not expect that it can be predicted from expert opinion in advance of a phase II study.

Devising a method for assessing the homogeneity of the basket-wise response rates at an interim analysis, Cunanan et al proposed the first method for efficient basket trial design.⁴⁸ More recent advances in statistical methodology facilitating model-based dynamic information-sharing have been developed to address this limitation. These approaches allow trialists to explicitly estimate the extent of heterogeneity evident from the trial data (integrating both small and large baskets without pooling).^49-52 A type of Bayesian model averaging using mixture priors on the basis of prespecified low-to-high expectations for response rates can accommodate different end points and null hypotheses across disease subtypes.⁵³ Multisource exchangeability models have been developed specifically for basket trials with sequential analysis.²² Using the Bayesian paradigm, multisource exchangeability models effectively model all possible pooling relationships among collections of tumor types to identify meta-subtypes with commensurate treatment efficacy. The analysis strategy explicitly quantifies the evidence for a tumor-agnostic label. The development of open-source software, such as the basket package in the R statistical software, has made these methods more readily accessible.²³ The application of information-sharing techniques to multi-indication studies has implications on a trial's operating characteristics. Recent work by Kaizer et al⁴⁷ has characterized these challenges and introduces weighted operating characteristics to calibrate basket trials to guard against drastic changes in operating characteristics when design assumptions are incorrect.

Interim Monitoring

Larger basket studies (> 100 patients) should implement sequential designs formulated with futility rules to terminate underperforming tumor types. For example, basket trials for both vemurafenib and neratinib used the frequentist Simon's two-stage design, which was applied independently within each basket. Devised with a single interim analysis that halts for futility if fewer than a predetermined number of successes are observed, the two-stage design provides a straightforward approach to sequential monitoring. If a sufficient number of overall responses is observed, enrollment continues to the maximum planned sample size.⁵⁴ Simon developed two approaches for designing two-stage trials that are calibrated to minimize the maximum sample size (minimax designs) or minimize the expected sample size (optimal designs).

Designs often fail to implement family-wise type I error control and fail to share information across baskets. Bayesian designs can leverage the use of either posterior probabilities or predictive probabilities and information sharing.^44,55,56 Predictive probability sequential rules are intuitive in that they facilitate decision making on the basis of forecasting the trial's final conclusion at interim analysis, quantifying the risk of ultimate failure. Lower predictive probabilities reflect higher risk of a negative trial if enrollment is continued. R software package ppseq provides tools for designing a clinical trial with sequential predictive probability monitoring.

REGULATORY GUIDANCE

Regulatory guidance insists that basket trials should include specific biological rationale for including each subpopulation. Moreover, statistical hypotheses, sample size justification, and stopping rules for futility need to be described for individual substudies within a detailed statistical analysis plan. Considering patient enrollment on the basis of a specific biomarker of interest, the guidance notes that the protocol should clearly specify how patients with more than one biomarker of interest will be assigned to substudies. To facilitate reviews of master protocols by institutional oversight committees, the FDA recommends the use of a central institutional review boards with adequate resources and appropriate expertise. A companion or complementary in vitro diagnostic device may be essential for the safe and effective use biomarker-target therapies. The FDA encourages frequent interactions with sponsors to help guide companion diagnostic development.⁵⁷

DISCUSSION

With the emergence of master protocol designs, innovations in clinical trial methodology toward precision medicine have been realized. Basket trials represent one type of master protocol devised to evaluate patients harboring a common actionable drug target across several histologies. These trials are designed with the presumption that the targeted feature is a stronger determinant of treatment response than histology. Basket trials facilitate efficient enrollment and provide a framework for studying rare tumor types with common genomic profiles. Statistical evaluations test for high and durable objective responses. Basket trials require tissue analysis and molecular profiling before registration. Thus, basket trials are practically feasible only at centers/institutions that can invest in this infrastructure.

Drug developers implementing basket trials tend to pursue avenues for streamlined regulatory pathways. Several drugs tested within this new paradigm have demonstrated exceptional efficacy signals, supporting the concept that tumor biology and/or immunologic factors can delineate actionable alterations across histologic types. This gained regulatory acceptance and culminated with the first approvals of histology-agnostic therapies.¹ Yet, with only four approved tissue-agnostic anticancer therapies to date, efficacy heterogeneity among cancer types has been evident more often than not among recently developed biomarker-guided anticancer therapies.

Designs of basket trials have predominantly used frequentist approaches. Innovations in statistical and design methodology have been proposed for early-phase basket trials to improve their efficiency and thereby conserve resources and reduce the number of patients being exposed to ineffective treatments. Bayesian approaches to sequential monitoring allow for smaller average sample sizes while maintaining acceptable statistical operating characteristics. Seamless phase I/II designs facilitate avenues for rapid trial and phase expansion that follow a prespecified design and thereby circumvent unplanned trial modifications via protocol amendment.⁵⁸ The assumption that a single biomarker is predictive of treatment benefit may be unrealistic in practice, given the complex and multidimensional nature of tumorigenesis, proliferation, and survival.

One should also note that it has become common to see phase I trials incorporate large dose-expansion cohorts as substudies. Often, these substudies enroll multiple tumor types and have an implicit efficacy objective, thus blurring the tradition border between phase I and phase II. Trials of pembrolizumab in melanoma and non–small-cell lung cancer incorporated large dose-expansion cohorts evaluating multiple doses.^59,60 Nivolumab was initially evaluated in five advanced solid tumor cohorts, but was later expanded through protocol amendments. Atezolizumab, an antiprogrammed death ligand-1 agent, was studied in dose-expansion cohorts comprising eight tumor types.⁶¹ Each of these trials effectively implemented a phase II basket trial spanning multiple indications with relatively large sample size. Authors have proposed frameworks to transition seamlessly from a pooled dose-escalation phase into a basket-style efficacy phase.⁶² The implementation of such methodology for seamless design may become more common with efforts to shorten drug development timelines.

AUTHOR CONTRIBUTIONS

Conception and design: Brian P. Hobbs, David S. Hong

Administrative support: David S. Hong

Provision of study materials or patients: David S. Hong

Collection and assembly of data: All authors

Data analysis and interpretation: All authors

Manuscript writing: All authors

Final approval of manuscript: All authors

Accountable for all aspects of the work: All authors

AUTHORS' DISCLOSURES OF POTENTIAL CONFLICTS OF INTEREST

Basket Trials: Review of Current Practice and Innovations for Future Trials

The following represents disclosure information provided by authors of this manuscript. All relationships are considered compensated unless otherwise noted. Relationships are self-held unless noted. I = Immediate Family Member, Inst = My Institution. Relationships may not relate to the subject matter of this manuscript. For more information about ASCO's conflict of interest policy, please refer to www.asco.org/rwc or ascopubs.org/jco/authors/author-center.

Open Payments is a public database containing information reported by companies about payments made to US-licensed physicians (Open Payments).