One master protocol, one shared actionable mutation — parallel tumor-type "baskets" evaluated under a single statistical and regulatory chassis
Basket trials invert the classic one-drug-one-tumor-type paradigm: instead of asking "does this drug work in lung cancer," they ask "does this drug work wherever this specific molecular alteration occurs, regardless of organ of origin." That inversion starts upstream, at the diagnostic layer — a broad next-generation sequencing (NGS) panel run across every incoming tumor sample, agnostic to primary site, looking for one defined actionable event.
Conventional oncology trials enroll by organ: a Phase II NSCLC trial recruits only lung adenocarcinoma. That works when a driver alteration is common within one histology (EGFR in NSCLC, ~15%). It breaks down for ultra-rare, pan-cancer alterations like NTRK1/2/3 gene fusions, RET fusions, or BRAF V600E in non-melanoma tumors — each individually too infrequent within any single organ to power a conventional trial, yet biologically identical wherever they occur.
A basket ("bucket") trial solves this by screening broadly and enrolling narrowly on genotype:
• Central IRB-approved umbrella screening protocol accepts tissue from any solid tumor type • DNA-based hybrid-capture NGS (FoundationOne CDx, 324 genes) plus a dedicated RNA fusion assay (RNA-seq or anchored multiplex PCR, ArcherDx) — DNA-only panels miss ~25–30% of NTRK fusions because many involve large intronic breakpoints • MSK-IMPACT (505 genes, hybridization capture) and Illumina TruSight Oncology 500 are the two other most common feeder assays in US academic networks • Local testing results are confirmed centrally before basket allocation, per master protocol quality standards under ICH E6(R2) Good Clinical Practice
Because the fusion is vanishingly rare pan-cancer (~0.31%), practical basket trials pre-enrich: they screen preferentially in tumor types with elevated a priori fusion frequency (secretory breast carcinoma, infantile fibrosarcoma, mammary analogue secretory carcinoma of salivary gland — each 75–95% fusion-positive) while still keeping the door open to any histology that screens positive.
The pivotal larotrectinib program (NAVIGATE + SCOUT + a Phase I adult trial, pooled analysis Hong et al., Lancet Oncology 2020) enrolled 159 NTRK fusion-positive patients spanning 21 distinct tumor types identified this way — from infantile fibrosarcoma to cholangiocarcinoma to appendiceal cancer — despite no single histology accounting for more than ~15% of the pooled dataset.
Once a patient screens positive for the shared alteration, they are allocated to a basket defined by tumor histology — but every basket sits inside the same master protocol document, the same investigational new drug (IND) application, the same statistical analysis plan (SAP), and the same data standards. This administrative unification is what makes basket trials dramatically faster and cheaper than running N separate single-histology studies.
FDA's 2018/2022 guidance "Master Protocols: Efficient Clinical Trial Design Strategies to Expedite Development of Oncology Drugs and Biologics" formally defines three master protocol subtypes:
• Basket trial: one drug, multiple diseases/tumor types sharing a molecular target — the design used here • Umbrella trial: one disease, multiple drugs/biomarker-matched arms (e.g., Lung-MAP) • Platform trial: multiple diseases, multiple drugs, perpetual/adaptive entry and exit of arms (e.g., I-SPY2)
Operationally, a basket trial is built as a single overarching protocol with:
• One IND covering the investigational agent across all tumor types (versus a new IND per indication in the classical model) • One master statistical analysis plan (SAP) specifying the per-basket design (typically Simon two-stage or Bayesian) and the pooling/borrowing rule • Shared infrastructure: one IRB/central ethics submission, one data management system, one CDISC SDTM/ADaM dataset structure re-used basket to basket, one safety database feeding a common DSMB • Basket-specific appendices: eligibility criteria, sample size, stopping rules — each basket can have a distinct target ORR because baseline response expectations differ by histology (e.g., historical chemo ORR in cholangiocarcinoma vs. NSCLC)
New baskets are added by protocol amendment rather than by filing a new study — typically a 4–8 week regulatory cycle versus 12+ months to stand up an independent single-histology trial, because the safety database, manufacturing (CMC), and non-clinical package are already established under the existing IND.
ICH E20 (finalized 2022) codifies adaptive trial design principles — including master protocols — internationally, giving sponsors a harmonized regulatory vocabulary for basket, umbrella, and platform submissions across FDA, EMA, and PMDA.
Even inside a shared master protocol, each basket typically starts with its own frequentist futility gate: Simon's two-stage design (Simon, Controlled Clinical Trials 1989). It is the workhorse small-sample oncology design because it lets a basket stop early — after as few as 10–14 patients — if the drug is clearly inactive in that specific histology, without wasting a full-size cohort or inflating the trial-wide false-positive rate.
Simon's two-stage design tests H0: true ORR ≤ p0 (uninteresting response rate, often historical standard-of-care) against H1: true ORR ≥ p1 (clinically meaningful response rate) with two pre-specified enrollment stages:
Stage 1: enroll n1 patients. If the number of responders r ≤ r1 (a pre-computed threshold), stop the basket for futility — the drug does not appear active in that histology and no further patients are exposed. If r > r1, proceed to stage 2.
Stage 2: enroll an additional n2 patients (total N = n1+n2). Reject H0 (declare the basket a signal) if total responders exceed a second threshold r.
"Optimal" Simon design minimizes the expected sample size under H0 (i.e., minimizes patients exposed if the drug truly doesn't work); "minimax" Simon design instead minimizes the maximum total sample size. Basket trials typically use the optimal variant per-basket, since limiting exposure in an ultimately-futile histology is the dominant operational concern when N is small pan-cancer.
Worked example matching a basket like cholangiocarcinoma (p0=0.10 historical ORR to chemo, p1=0.30 target, α=0.05, power=0.80): n1=10, r1=1 (≤1/10 responders stops for futility); if the basket proceeds, n2 brings the total to N=29, r=5 (need ≥6/29 responders total to declare a signal).
Because every basket runs its own Simon boundary independently, a null result in, say, pancreatic cancer does not penalize or bias the ongoing sarcoma or thyroid baskets — each is its own frequentist experiment, sharing only infrastructure, not statistical inference, at this stage.
Simon's design treats each basket as an isolated experiment. But if the mechanism of action is truly tumor-agnostic — the drug blocks the same fusion kinase regardless of organ — then a response signal in one basket is genuine (if partial) evidence about the others. Bayesian hierarchical models (BHM), most influentially the calibrated BHM of Berry et al. (Clinical Trials, 2013) built for exactly this problem, formalize how much of that cross-basket evidence to borrow.
In a Bayesian hierarchical model, each basket k has a true response probability θ_k on the logit scale, drawn from a shared normal distribution: logit(θ_k) ~ N(μ, τ²), where μ is the overall pooled mean effect and τ is the between-basket standard deviation — the "exchangeability" or borrowing parameter.
Two limiting cases bound the model:
• τ → 0: baskets are assumed to have identical true response rates (full pooling / complete exchangeability). A tiny basket with 2/4 responders gets pulled hard toward the pooled estimate from all other baskets — high power, but risky if that basket is biologically genuinely different.
• τ → ∞: baskets are treated as fully independent (equivalent to separate Simon analyses, no borrowing). Small baskets get no help from the aggregate signal — safe, but underpowered for rare histologies with N as low as 4–6 patients.
Calibrated BHM (Berry 2013) selects τ by simulation to control the frequentist false-positive and false-negative rates of the pooled procedure across plausible true scenarios (all baskets equally active, one outlier basket inactive, etc.) — "calibration" means τ is not a subjective prior choice but tuned so the design's operating characteristics meet a pre-specified type-I error target basket by basket.
Extensions used in modern master protocols:
• EXNEX (EXchangeability-NonEXchangeability, Neuenschwander et al. 2016): mixture model that lets each basket probabilistically be "exchangeable" with the group or "stand alone," automatically down-weighting an outlier basket instead of forcing a single τ on all baskets • MEM (Multi-source Exchangeability Model, Hobbs et al. 2018, used in the BRUIN and other basket programs): pairwise exchangeability between every basket pair, estimated from the data itself
Borrowing directly changes what gets reported: a basket with a raw observed ORR of 40% (2/5 patients) might have a posterior (borrowed) ORR of 58% if four sibling baskets are running at 70%+ — the shrinkage estimate is what regulators and clinicians actually see in basket-trial publications, not the raw per-basket count.
The pivotal larotrectinib pooled analysis explicitly used a Bayesian approach only for supportive sensitivity analyses; the primary efficacy claim rested on the simple pooled ORR (75%, 95% CI 61–85%) across independent histologies — reflecting FDA's continued preference, circa 2018, for straightforward pooling over model-based borrowing when the totality of tumor types is still small.
The adaptive core of a basket trial is the decision layer that sits on top of the statistics: a Bayesian posterior probability threshold (or a Simon stage-1 boundary) is checked at each interim look, and the trial itself — not just the analysis — physically reshapes. Futile baskets stop enrolling and release their slots; promising baskets expand into larger cohorts; and, increasingly, entirely new tumor-type baskets are grafted on mid-trial as new fusion-positive histologies are identified.
At each pre-specified interim analysis, every active basket is evaluated against a decision rule written into the master SAP before the trial started (never post hoc, to preserve interpretability):
• Go / expand: if the posterior probability that the true ORR exceeds a clinically meaningful bar (commonly 30%, chosen against historical control response rates) exceeds a high threshold (e.g., 0.85), the basket is declared active and expands to its full planned stage-2 sample size, or beyond, as an expansion cohort feeding pivotal-quality data.
• No-Go / close: if the posterior probability drops below a low threshold (e.g., 0.05), or the basket fails its Simon stage-1 boundary, enrollment closes immediately. Patients already enrolled continue follow-up for safety and long-term outcome, but no new patients are randomized into that histology.
• Continue / gray zone: baskets that are neither clearly active nor clearly futile continue to the next interim look without a definitive Go/No-Go call — this is common for ultra-rare histologies where even the maximum feasible N leaves genuine uncertainty.
An independent Data and Safety Monitoring Board (DSMB) reviews unblinded interim results and safety signals across all baskets simultaneously, since the drug's toxicity profile is shared even when efficacy diverges by tumor type — a basket can be efficacy-futile while still contributing safety data to the pooled label.
Adding a basket mid-trial: when a new fusion-positive histology is identified (e.g., the first NTRK fusion pancreatic cancer case in a screening network), sponsors file a protocol amendment defining the new basket's eligibility and sample size under the existing master SAP and IND — typically a 4–8 week cycle, versus starting an entirely new single-histology study. This is the single biggest driver of basket trials' cost and time efficiency relative to the classical model.
NCI-MATCH (started 2015) screened over 6,000 patients across 40+ molecularly defined treatment arms under one master protocol; only about 26% of screened patients matched to an available arm, illustrating both the reach and the diagnostic yield challenge of running many parallel baskets from a single central screening funnel.
Basket trial data feeds directly into a regulatory pathway that did not exist before 2017: tissue-agnostic (histology-independent) approval, where a drug is authorized based on a shared molecular alteration rather than a specific organ of origin. This closes the loop from the screening funnel in Stage 1 to a label that any oncologist, for any solid tumor, can act on once the biomarker is found.
Tissue-agnostic approvals in the US run through FDA's accelerated approval pathway (21 CFR Subpart H), which allows authorization based on a surrogate endpoint reasonably likely to predict clinical benefit — for basket trials, that surrogate is overall response rate (ORR) and duration of response (DoR), not overall survival, because no single basket accrues enough events for a survival endpoint on its own.
Regulatory precedents:
• Pembrolizumab (Keytruda), MSI-H/dMMR, May 2017 — the first tissue-agnostic approval, based on 149 patients across 15 tumor types pooled from five uncontrolled trials • Larotrectinib (Vitrakvi), NTRK fusion, November 2018 — pooled analysis of 55 patients (initial approval; expanded to 159 by 2020 update) across 17 tumor types, ORR 75% • Entrectinib (Rozlytrek), NTRK fusion, August 2019 — second-in-class TRK inhibitor, similar pooled-basket evidentiary structure • Selpercatinib (Retevmo), RET fusion, 2022 (LIBRETTO-001 basket) • Dabrafenib + trametinib, BRAF V600E, June 2022 — first tissue-agnostic approval for a targeted combination outside melanoma • Dostarlimab, dMMR, 2021; trastuzumab deruxtecan, HER2-positive solid tumors (DESTINY-PanTumor02), 2024
Because accelerated approval is conditional, sponsors commit to confirmatory trials or continued basket accrual to verify clinical benefit; FDA's Project Orbis enables simultaneous review with international partners (Australia's TGA, Health Canada, Swissmedic, UK MHRA, Singapore HSA), compressing global access timelines for these small, rare-biomarker populations.
Post-marketing, real-world evidence (RWE) registries and expanded-access programs continue feeding both efficacy follow-up and pharmacovigilance: adverse events flow into FDA FAERS and EMA EudraVigilance under ICH E2B(R3) case-reporting standards, coded to MedDRA preferred terms, while payers (CMS, and internationally NICE/ICER-style bodies) evaluate tissue-agnostic pricing against a fundamentally different cost-effectiveness question than a single-indication drug — value assessed across a portfolio of rare, heterogeneous tumor types rather than one population.
ASCO's TAPUR study, a real-world pragmatic basket trial launched in 2016, has since enrolled patients into more than 20 parallel drug-biomarker basket cohorts drawn from routine community oncology practice — demonstrating that the basket architecture generalizes beyond registration trials into post-approval, real-world evidence generation.