Bayesian basket trials: a faster path for rare and orphan indications

A Bayesian basket trial tests one drug across several biomarker-defined or mechanistically related diseases under a single master protocol, and uses a hierarchical model so each disease cohort borrows evidence from the others. For rare and orphan indications, this can replace several unaffordable Phase 2 trials with one. It works only when the diseases are plausibly similar and the design's error rates are simulated and shared with FDA in advance.

What is a basket trial, and how does it differ from umbrella and platform trials?

A basket trial is a master protocol that tests a single drug or combination in multiple populations defined by cancer type, histology, biomarker, or prior therapy. FDA's March 2022 master protocols guidance defines the term this way and notes that basket substudies are usually single-arm cohorts with overall response rate as the primary endpoint [1].

An umbrella trial is the inverse: several drugs tested within one disease, allocated by biomarker [1]. A platform trial runs multiple therapies against a common control and adds or drops arms over time. Woodcock and LaVange laid out this taxonomy in 2017, and FDA cites their review as the source of the terminology [2].

What does Bayesian borrowing actually do to a small cohort?

Without borrowing, each basket is its own Phase 2 trial with its own sample size. Berry and colleagues compared Simon's two-stage design, an independent Bayesian design, and a hierarchical Bayesian design across four patient groups. Under the null hypothesis, the hierarchical model reduced mean sample size by an additional 4 to 7 patients per group beyond the independent Bayesian design [3].

The gain is not free. When the drug worked in all groups, power rose to at least 98% in every group. When it worked in only one group, power for that group fell to 65%, because the inactive cohorts dragged its estimate down [3]. Borrowing rewards homogeneity and punishes the lone responder.

FDA's January 2026 draft guidance on Bayesian methodology calls hierarchical models the main way to implement exchangeable models, and uses basket trials as the example: treat the disease-level treatment effects as a random sample from one common distribution. It cautions that exchangeability is not always appropriate, for instance when a drug is expected to work better in a biomarker-positive subgroup [4].

Why does this matter for rare and orphan indications?

Orphan designation under 21 CFR Part 316 requires documentation that the disease affects fewer than 200,000 persons in the United States [5]. Designation brings tax credits for qualified clinical trials, user fee exemption, and potential seven-year market exclusivity [6]. The incentives are well known. The trial economics are the harder problem.

On the first-in-class biologic I led for a rare pediatric leukemia, the target biology pointed to other hematologic malignancies and immune-mediated diseases. That mapping moved the potential U.S. addressable market from roughly $200M per year into the multi-billions. The open question was how to generate clinical evidence in several small populations at once.

Separate Phase 2 trials per indication were not affordable for a virtual company. Each needed its own protocol, sites, enrollment period, and budget. A single basket protocol with a shared safety database and a hierarchical analysis was the only structure that let us plan a credible path across more than one indication. One limit was plain: nonclinical efficacy justified including an indication, but it did not justify assuming the cohorts were exchangeable.

What has FDA actually approved on basket-trial evidence?

On May 23, 2017, FDA granted accelerated approval to pembrolizumab for microsatellite instability-high or mismatch repair deficient solid tumors, the first approval based on a biomarker rather than tumor site. The FDA approval summary reports 149 patients across 15 tumor types in five single-arm trials and a 39.6% objective response rate [7].

On November 26, 2018, FDA granted accelerated approval to larotrectinib for solid tumors with NTRK gene fusions. Efficacy rested on 55 patients across 12 tumor types from three single-arm trials, with a 75% overall response rate [8]. The NEJM publication reported 17 unique fusion-positive tumor types, with 71% of responses ongoing at one year [9].

Both were accelerated approvals with confirmatory obligations, and both pooled single-arm data rather than running a formal Bayesian primary analysis. FDA's October 2022 draft guidance on tissue-agnostic development states that these applications face the same standards as tissue-specific drugs [10]. A basket design changes the shape of the evidence, not the bar.

Where does borrowing go wrong?

Heterogeneity can wash out a real signal or manufacture a false one. In the vemurafenib basket study of 122 patients with BRAF V600 mutations, response rates were 42% in non-small-cell lung cancer and 43% in Erdheim-Chester disease or Langerhans'-cell histiocytosis, while other cohorts produced only anecdotal responses. The authors concluded that BRAF V600 is targetable in some, but not all, nonmelanoma cancers [11].

Freidlin and Korn simulated five to ten subgroups and found insufficient information in the outcome data to tell whether borrowing was appropriate. Models that borrowed more strongly had markedly lower power when the drug worked in only a minority of subgroups, and the same response data could be made to borrow heavily or hardly at all by changing the hyperprior [12].

Multiplicity is the other trap. Cunanan and colleagues showed that five independent Simon two-stage baskets at a 5% error rate each give roughly a 23% chance that an inactive drug appears active in at least one basket, rising to about 40% with ten baskets [13]. Borrowing can reduce that, but FDA's draft guidance notes that borrowing itself typically inflates the Type I error rate relative to the nominal level [4].

What does FDA expect from a sponsor proposing a Bayesian basket design?

FDA's December 2020 guidance on complex innovative trial designs asks for a rationale for any borrowing, an explanation of how priors were constructed, and an evaluation of operating characteristics including the chance of erroneous conclusions. Where Type I error control is not applicable, sponsors should propose alternative characteristics and test sensitivity to the prior [14].

The same guidance points to existing meeting pathways: pre-IND meetings for early-phase designs with novel elements, Type C meetings for statistical detail, and End of Phase 2 meetings for overall strategy. Because review often involves extensive computer simulation, meeting packages need more lead time than usual [14].

For a basket protocol, the master protocols guidance expects each substudy to carry its own objectives, a rationale for that population, and a statistical analysis plan with sample size justification and stopping rules. If any substudy may support a marketing application, FDA expects an independent data monitoring committee and blinded radiologic review [1].

FDA's announcement of the draft Bayesian guidance called these methods especially valuable for rare and pediatric indications where populations are small [15]. That is encouraging, not a shortcut. The sponsor still owns the biology argument for exchangeability and the simulation report that shows what happens when it is wrong.

This article is part of the Market Positioning Solutions practice at Hamilton Atlas Group — see the full capability list there.

References

1. U.S. Food and Drug Administration. Master Protocols: Efficient Clinical Trial Design Strategies to Expedite Development of Oncology Drugs and Biologics. FDA, 2022.
2. Woodcock J, LaVange LM. Master Protocols to Study Multiple Therapies, Multiple Diseases, or Both. N Engl J Med. 2017;377(1):62-70.
3. Berry SM, Broglio KR, Groshen S, Berry DA. Bayesian hierarchical modeling of patient subpopulations: Efficient designs of Phase II oncology clinical trials. Clin Trials. 2013;10(5):720-734.
4. U.S. Food and Drug Administration. Use of Bayesian Methodology in Clinical Trials of Drug and Biological Products. FDA, 2026.
5. U.S. Government Publishing Office. 21 CFR Part 316 - Orphan Drugs. Code of Federal Regulations, 2024.
6. U.S. Food and Drug Administration. Designating an Orphan Product: Drugs and Biological Products. FDA, 2024.
7. Marcus L, Lemery SJ, Keegan P, Pazdur R. FDA Approval Summary: Pembrolizumab for the Treatment of Microsatellite Instability-High Solid Tumors. Clin Cancer Res. 2019;25(13):3753-3758.
8. U.S. Food and Drug Administration. FDA approves larotrectinib for solid tumors with NTRK gene fusions. FDA, 2018.
9. Drilon A, Laetsch TW, Kummar S, et al. Efficacy of Larotrectinib in TRK Fusion-Positive Cancers in Adults and Children. N Engl J Med. 2018;378(8):731-739.
10. U.S. Food and Drug Administration. Tissue Agnostic Drug Development in Oncology. FDA, 2022.
11. Hyman DM, Puzanov I, Subbiah V, et al. Vemurafenib in Multiple Nonmelanoma Cancers with BRAF V600 Mutations. N Engl J Med. 2015;373(8):726-736.
12. Freidlin B, Korn EL. Borrowing Information across Subgroups in Phase II Trials: Is It Useful? Clin Cancer Res. 2013;19(6):1326-1334.
13. Cunanan KM, Gonen M, Shen R, et al. Basket Trials in Oncology: A Trade-Off Between Complexity and Efficiency. J Clin Oncol. 2017;35(3):271-273.
14. U.S. Food and Drug Administration. Interacting with the FDA on Complex Innovative Trial Designs for Drugs and Biological Products. FDA, 2020.
15. U.S. Food and Drug Administration. FDA Issues Guidance on Modernizing Statistical Methods for Clinical Trials. FDA, 2026.