Three trial designs get sold as one product. Basket, umbrella and platform trials, bundled together under "master protocol" and pitched to you at conferences and in CRO decks as the modern, efficient way to run a study. Here is the part the deck skips over: a single-asset biotech can sponsor exactly one of the three.
That is the strategy conversation, and almost nobody has it in that order. The founder's opening question is usually "should we run a master protocol?", which is the wrong question: for two of these three designs you were never eligible to be the sponsor in the first place.
So before we get to who runs what, be clear on what the three actually are. They get muddled constantly, sometimes by people with something to sell.
Strip the jargon and each design answers a different question about what you are varying.
Basket trials hold the drug constant and vary the disease. One drug tested across several conditions that share a common molecular alteration. One drug, many diseases. VE-BASKET is the textbook case: vemurafenib across a range of BRAF V600-mutant non-melanoma cancers under a single protocol, 208 patients, run to completion by Roche (NCT01524978) [1][2].
Umbrella trials hold the disease constant and vary the drug. One condition, patients sorted by biomarker into several drug sub-studies under one protocol. One disease, many drugs. Lung-MAP does this in non-small-cell lung cancer (NCT03851445, run by SWOG); Beat AML does it in acute myeloid leukaemia (NCT03013998) [3][4].
Platform trials add time. A standing, adaptive, multi-arm structure where interventions enter and leave over the trial's life, each tested against a common control. The protocol outlives any single arm. GBM AGILE runs this way in glioblastoma (NCT03970447); I-SPY2 has done it in breast cancer for years [5].
The master protocol itself is the shared spine underneath all three: one screening and operational engine feeding many questions at once, instead of a fresh trial for each. Park et al.'s 2019 systematic review is the clean reference for the formal definitions (PMID 31533793) [1].
Memory hook, because these get confused in the room: basket varies the disease, umbrella varies the drug, platform varies over time.
One caution about NCI-MATCH, the trial you will hear cited as the poster child for all of this. It is a genomic-screening master protocol running many sub-studies, closer to a basket-of-baskets than a clean example of any one category. Its value here is its numbers, which come shortly.
Let's be clear about who actually runs these. Look at every umbrella and platform trial in the wild, and the sponsor is almost never a single-asset company.
NCI-MATCH is sponsored by the National Cancer Institute (NCT02465060). Lung-MAP by SWOG, a cooperative group. Beat AML by a purpose-built consortium, Beat AML, LLC. TAPUR by ASCO, a professional society (NCT02693535). GBM AGILE by the Global Coalition for Adaptive Research, a nonprofit [6][7]. Where a company does sponsor an umbrella, it is a large one drawing on its own multi-asset pipeline: Roche running more than a dozen parallel Morpheus trials, GSK and Amgen master protocols off their own pipelines [23].
The reason is structural. An umbrella needs many drugs to sort patients into; a platform needs a rotating cast of arms to keep the protocol fed. You have one drug. You cannot fill an umbrella with one arm, nor sustain a platform with an arm that never rotates out. It is the same mismatch we flagged in the biotech playbook piece: tools built for a portfolio company, borrowed by one with a single shot on goal.
CTTI, the FDA-Duke public-private partnership, mapped this in its Master Protocol Value Proposition Guide. Its worked examples of platform sponsors are I-SPY (QuantumLeap Healthcare Collaborative, a 501(c)(3)), DIAN-TU (Washington University), and the Duchenne platform trial (I-ACT for Children, with PPMD) [8]. Not one is a single-asset biotech. In CTTI's own breakdown, the role reserved for an individual drug's developer (the IMP developer, in CTTI's terms) is fee-payer: you pay user fees to run your arm inside someone else's trial.
That is the honest picture. Single-asset biotechs do turn up in these trials, names like Kazia, Kintara, Vigeo and Syndax, but every time they appear as arm-contributing collaborators, never as the overall sponsor [8].
Lu et al. said it plainly back in 2021. Master protocols fit programmes with "multiple experimental treatments and/or multiple diseases... without a clear priority", and are the wrong tool "if a clinical development program team has a clear focus on one compound and/or one indication". Their image for that mismatch: "like using a swiss army knife to perform a heavy paper cutting job" (PMID 34160785) [9].
So for two of the three designs, here is the blunt version. You are not the sponsor. You are an arm.
Read that as a redirect, then. When someone floats an umbrella or a platform in your disease area, find the cooperative group or consortium already running one and get your asset accepted into it. You can rent an arm inside someone's platform; you cannot finance the whole building for a single tenant.
Free download
The TPP Template
Target and minimally-acceptable profiles side by side, every claim linked to the evidence that has to support it.
Get the template →Now the design that is genuinely yours to sponsor. The basket is the one master-protocol design a single-asset biotech can own outright. It asks only for one drug and several diseases; if your target recurs across tumour types, your asset fits that shape without any portfolio behind it.
The registry proves the door is open. Puma Biotechnology, a single-asset company, ran its own neratinib basket, SUMMIT, for the better part of a decade (September 2013 to January 2023), enrolling 582 patients before it too was terminated (NCT01953926). Debiopharm ran its own Debio 1347 basket, FUZE, across roughly 105 planned sites in more than 20 countries (NCT03834220). Istari Oncology ran its own lerapolturev basket, LUMINOS-103 (NCT04690699) [10][11][12]. Nobody stopped them at the door.
But open is not free, and here the sponsor-type question bites. When a consortium runs a basket, a single slow-accruing cohort is one line item among many and the trial keeps moving. Run one alone and every cohort is your cohort, with no other arm to carry the trial while one struggles. You are self-insuring a single cohort where a consortium spreads that risk across a whole syndicate of them.
The registry outcomes track that exposure. SUMMIT stayed open for ten years and reached 582 patients; FUZE terminated at 63, LUMINOS-103 at 4. The registry does not say why any of them stopped, and I will not invent a reason. But the shape is the point: the lone-sponsor baskets carried the whole accrual risk on one asset, and several did not go the distance.
That said, size is not the whole story here. Novartis, hardly resource-constrained, still terminated its own iptacopan basket at 19 patients (NCT05086744) [13]; set that against Roche's VE-BASKET, which completed cleanly at 208, and the fragility looks identical in kind. Scale just absorbed it. What actually decides the outcome is how many arms you have to spread the accrual risk across.
So if you sponsor a basket alone, size your runway to the SUMMIT reality, a decade in which one cohort's accrual can decide the whole trial, not to the efficiency line in the brochure. And if a full master protocol is off the menu, other designs are not: a TwICS design gives a lean oncology sponsor a randomised comparison without a portfolio to fill.
Why does spreading risk across many arms matter so much? Because a master protocol is mostly a sorting machine the treatment is a small attachment on the end. The screening funnel, not the therapy, is where most patients are lost, and NCI-MATCH quantified that better than any trial in the literature.
Start with biomarker prevalence, the number everyone quotes. In the full NCI-MATCH cohort of 5,540 patients after exclusions, 37.6% had a molecular alteration that could, in principle, assign them to a treatment arm. Strip out the eligibility exclusions and the match rate falls to 26.4%. Now the number that actually matters: the rate of patients assigned to treatment was 17.8%, or 985 of 5,540 (Flaherty et al., J Clin Oncol 2020, PMID 33048619). Nearly a third of the theoretically matchable patients still could not be treated. The science was not the bottleneck. There simply was not enough parallel subprotocol capacity running at once: arms were full, histology limits hit [14].
That fall, from 26.4% down to 17.8%, is the clearest picture in the evidence base of what portfolio depth buys you, standing infrastructure, enough arms open at once to catch the patients your screen keeps finding. The engine strains under its own success, too. As NCI-MATCH accrued, its median biopsy-to-result time rose from 14 days in September 2015 to 36 days a month later, because, in the investigators' words, "the robust accrual overwhelmed the initially planned resources" [14].
The trial-management view from STAMPEDE and FOCUS4 (Schiavone et al. 2019) fills in the rest. Every biomarker assay had to clear a formal validation gate before its sub-study could open, and one added comparison meant a drug-supply tender that put more than 200 days onto time-to-activation. The initial build took five years to get 80 centres to randomise their first patient; once the spine existed, later comparisons needed only three to four months (PMID 31138317) [15]. The economics only work at that scale: Park et al. modelled ten interventions run as one platform trial at $104.95M, versus a median 57.5% more and 311.9% longer to run them separately (PMID 35819785) [16].
So the spine is a fixed cost. A consortium amortises it across many arms and reuses it for a decade; a single-asset sponsor pays the whole build for one drug. Model the funnel, then, not the headline prevalence of your biomarker: the rate that decides your trial is screen-to-assigned, and it sits well below the number on the slide.
There is one more cost hiding inside these designs. Worth stating in plain English, even if the statisticians own the fine print.
Basket trials squeeze power out of small cohorts by borrowing information across sub-baskets, treating cohorts as related so a signal in one lends strength to another. That borrowing is where the bill lands. Chu and Yuan showed in 2018 that naive Bayesian hierarchical borrowing across cohorts that are not truly alike can inflate the Type I error rate "from the nominal value of 10% to over 50%" (PMID 29499621) [17]. Calibrated methods exist to control it; naive borrowing is the trap. Benchmarked on the real VE-BASKET data, where one cohort held just seven patients, Daniells et al. found Type I error running between 21.6% and 42.1% against a nominal 10% (PMID 37614070) [18]. And principle and practice part company: Li et al.'s 2022 cross-industry survey found that "Most of the organizations did not attempt to control the study level family-wise type I error rate (FWER) in master protocol trials" [19].
Know exactly what is settled on the regulatory side and what is not. The FDA's master-protocol guidance for oncology drugs is final (issued February 2022, docket FDA-2018-D-3292). Its general, cross-therapeutic-area master-protocol guidance is still in draft; a second iteration was reissued in June 2026, with comments open into August (docket FDA-2023-D-5259) [20][21]. In the EU, the operative document is the adopted 2022 Complex Clinical Trials Q&A (EMA/298712/2022), and, as with the EMA's posture on single-arm trials, its line is operational feasibility first: "Operational feasibility at multiple trial sites should always be a major focus in the choice of CCT design... Increased sponsor oversight will not be sufficient to justify a CCT design if clinical feasibility at the investigator site is compromised" [22].
No free lunch, in other words. Borrowing power from thin cohorts has a price, and you want to know which error you are accepting before a cohort reads out ambiguously, not after.
Single-asset biotechs are visibly inside umbrella and platform trials right now: Kazia, Kintara, Syndax and others have run their molecules in exactly these designs. So does that not blow a hole in everything above?
It does not, and the gap is in one word: "in" is not "sponsor". In every example the research turned up, the single-asset company is there as an arm: a collaborator contributing one drug and paying to run it, never the entity that holds and sustains the master protocol. The design is wide open to your drug. It is closed to your sponsorship, and once you see that, it stops being bad news. You do not build a platform; you get your asset adopted into one that already runs.
So before anyone in your building costs out a master protocol, run three questions in order.
First, which of the three is actually on the table: basket, umbrella or platform? If the person proposing it cannot answer that cleanly, you have your first finding already.
Second, for that design, are you the sponsor or an arm? If it is an umbrella or a platform, you are an arm, so find the consortium or cooperative group already running one in your indication and route your asset in rather than trying to build the building yourself.
Third, if it is a basket and you will sponsor it alone, size your runway and your risk to what the registry shows: a single-cohort trial where your accrual is the whole trial's fate, not one line in someone else's portfolio.
Get those three in order and you stop arguing about fashion. The real question surfaces on its own: does your one asset even clear the bar to sponsor the design, or should it be renting an arm inside someone else's? For a single-asset biotech weighing a master protocol against a plain, focused trial, that is the thing worth settling first.
A basket trial holds the drug constant and varies the disease: one drug tested across several conditions that share a molecular alteration. An umbrella trial holds the disease constant and varies the drug: one condition, several drugs, sorted by biomarker. A platform trial adds time, a standing structure where arms enter and leave over the trial's life. Basket varies the disease, umbrella varies the drug, platform varies over time.
Only one of the three: a basket trial, which needs just one drug and several diseases. Umbrella and platform trials both need a portfolio of drugs or a rotating cast of arms to sustain them, so a single-asset company is almost always a fee-paying arm inside someone else's trial, never the sponsor.
Cooperative groups, consortia and professional societies: NCI-MATCH is sponsored by the National Cancer Institute, Lung-MAP by SWOG, TAPUR by ASCO, and GBM AGILE by the Global Coalition for Adaptive Research. Where a company does sponsor one, it is a large, multi-asset developer drawing on its own pipeline, such as Roche's Morpheus programme.
You carry the whole accrual risk on one cohort instead of spreading it across a syndicate of arms. Puma Biotechnology's SUMMIT basket ran a decade and reached 582 patients; Debiopharm's FUZE terminated at 63; Istari's LUMINOS-103 terminated at 4. Size your runway to that reality, not to the efficiency pitch in the deck.
Get the monthly digest
The 5 things evidence leads need to know each month: regulatory moves, RWE developments and what they mean in practice. No pitch, one email a month.