Oncofusion

start / graph / trialdesign-basket-and-umbrella-for-rare-fusions

Explainerstrong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.Q8-impact

Basket and umbrella designs: pooling rare fusions into one trial

also called basket trial · umbrella trial · fusion-agnostic trial design

A basket trial puts patients with many different cancers into one study because they share the same molecular fault; an umbrella trial does the reverse, splitting one disease into several arms by which fault each patient has. Both exist because no single rare group is big enough to study on its own.

The same thing, for a scientist

A basket design enrols one biomarker across many diseases and is the natural fit for a fusion family with 107 partners and a shared menin dependency. The founding oncology example, the vemurafenib BRAF V600 basket, is also the cautionary one: it showed response in some histologies and conspicuous non-response in colorectal cancer, proving that the same lesion in a different context is a different disease. Statistically, the modern form is a Bayesian hierarchical basket that borrows strength between baskets while shrinking towards the observed heterogeneity, so that a discordant basket is not dragged along by the others. For KMT2A, the natural basket axis is partner gene and the natural umbrella axis is age and lineage.

Go deeper
The founding example, and the warning inside it2 deeper

The vemurafenib basket enrolled patients with BRAF V600 mutations across multiple non-melanoma cancers and found response in some histologies and not others, most famously little activity in colorectal cancer where feedback EGFR activation rescues the pathway 1.

The lesson is not that baskets fail. It is that a basket is an experiment about context-dependence, and it should be designed to detect it. For KMT2A, the analogous question is whether menin inhibition works equally across partners and ages, and the graph already records that partner identity is strongly prognostic in AML, with subgroup 5-year EFS spanning 11% to 92% (disease-kmt2a-r-aml). A KMT2A basket that pools partners without a pre-specified heterogeneity analysis is set up to produce exactly the vemurafenib-in-colorectal error in reverse: a positive average concealing a subgroup with no benefit.

The statistical machinery that makes a basket honest

Bayesian hierarchical modelling of patient subpopulations lets each basket have its own response rate while shrinking those rates towards a common mean, with the degree of shrinkage estimated from the observed between-basket variability 2. If the baskets agree, the design behaves nearly like a single pooled trial and gains power. If one basket disagrees, shrinkage weakens and that basket is allowed to stand apart.

That is precisely the property a KMT2A trial needs. It buys the pooling that makes a 48-patients-a-year disease tractable without committing in advance to the assumption that all partners behave alike.

The cost is that hierarchical designs have poorly calibrated frequentist error rates unless simulated carefully, and regulators will want the operating characteristics under a range of true heterogeneity scenarios before agreeing.

What a KMT2A basket and umbrella would actually look like

Basket axis (one drug, many strata): menin inhibitor across KMT2A::AFF1, KMT2A::MLLT3, KMT2A::MLLT1, KMT2A::MLLT10, KMT2A::ELL, KMT2A::AFDN, rare partners pooled, plus NPM1-mutant and NUP98-rearranged as adjacent baskets. The shared dependency is the menin-KMT2A interface. About 48 KMT2A-rearranged infants a year internationally plus the far larger adult strata makes the aggregate viable (ana-infant-trial-feasibility-arithmetic).

Umbrella axis (one disease, many drugs): infant KMT2A-rearranged ALL assigned by co-lesion to menin inhibitor, RAS-pathway inhibitor (comut-ras-pathway-kmt2a-r), FLT3 inhibitor (comut-flt3-kmt2a-r), or CD19-directed immunotherapy. This is the design that matches the PI's 'therapies based off their custom biology' framing, and it is realistic because the co-mutational landscape of infant KMT2A-rearranged leukemia is unusually sparse, so the number of arms needed is small.

The infant-specific problem neither design solves: both still require an infant stratum large enough to say anything, and no amount of pooling across adult strata creates infant data. Pooling makes the trial feasible; it does not make the infant answer available unless borrowing is explicitly modelled (trialdesign-bayesian-borrowing-across-fusions).

Facts and figures
design classbasket and umbrella
statistical methodBayesian hierarchical model with adaptive shrinkage between baskets
key riska positive pooled result concealing a partner subgroup with no benefit
basket axis proposed
[
  "fusion-kmt2a-aff1",
  "fusion-kmt2a-mllt3",
  "fusion-kmt2a-mllt1",
  "fusion-kmt2a-mllt10",
  "fusion-kmt2a-ell",
  "fusion-kmt2a-afdn",
  "NPM1-mutant AML",
  "NUP98-rearranged"
]
umbrella axis proposed
[
  "menin inhibitor",
  "RAS-pathway inhibitor",
  "FLT3 inhibitor",
  "CD19-directed immunotherapy"
]
precedents
[
  "BRAF V600 vemurafenib basket",
  "NTRK tumour-agnostic programme"
]
How this could bite us back
  • Pooling partners to make a trial feasible imports the assumption the graph's own AML data contradict. If partner identity modifies menin-inhibitor response, a pooled positive result licenses a drug for partners in which it does not work, and infants with AFF1 disease are the largest single partner group that could be wrong about.
  • A basket that enrols adults and infants in the same strata will fill with adults and report an average that is an adult result with an infant label attached. The basket has to pre-specify a minimum infant stratum size, or it recreates the AUGMENT-101 problem with better statistics.
Still open
  1. Would regulators accept KMT2A partner gene as a basket axis the way they accepted tissue of origin for NTRK, and has anyone asked?
  2. What between-basket heterogeneity would a hierarchical model need to detect before it stopped borrowing, and is the partner-specific EFS spread in KMT2A-rearranged AML (11% to 92%) large enough to break pooling outright?
  3. Is the co-mutational landscape of infant KMT2A-rearranged leukemia sparse enough that an umbrella design would have only two or three meaningfully populated arms?
Every source behind this entry (3)
  1. E1established confidence 5 of 5: Multiple independent lines of evidence; treated as settled in the field.paper

    A histology-independent basket trial of vemurafenib in BRAF V600-mutant non-melanoma cancers found responses in some tumour types and notably limited activity in colorectal cancer, establishing both the feasibility of biomarker-defined baskets and the reality of tissue context-dependence.

    Vemurafenib in Multiple Nonmelanoma Cancers with BRAF V600 Mutations · Hyman DM, Puzanov I, Subbiah V, et al. · New England Journal of Medicine 2015
    PMID 26287849DOI
    What this evidence cannot show

    Non-randomised, many small histology cohorts with wide confidence intervals; some baskets closed for futility with very few patients, so absence of response in a small basket is weak evidence of true inactivity.

    Evidence that argues against this (2)
    • The documented basket failure that is closest to this graph's situation, because its eligibility included FUSIONS and not only point mutations. NCI-MATCH subprotocol R (EAY131-R) assigned 50 patients with BRAF non-V600 mutations or BRAF fusions to trametinib; only 32 were eligible and treated (one fusion, 13 class 2 and 19 class 3 mutations). There were no complete responses, one confirmed partial response (3%), median progression-free survival 1.8 months and median overall survival 5.7 months, and the authors state the subprotocol 'did not meet its primary endpoint'. Two operational lessons the entry does not currently carry: 36% of assigned patients never got treated, and a genotype-defined basket can fail outright rather than merely failing in one histology. Abstract read via PubMed; full text not read. Johnson DB et al. 2020
    • A second NCI-MATCH basket arm with the same shape. Subprotocol Z1A treated 47 patients with refractory solid tumours bearing NRAS codon 12, 13 or 61 mutations with binimetinib: objective response rate 2.1% (1/47), and 30% of patients discontinued because of drug-related toxicity. The single durable partial response was in a malignant ameloblastoma. The lesson for a KMT2A partner-axis basket is that a shared driver lesion plus a mechanistically matched drug is not sufficient, and that toxicity attrition can be as large as the response rate. Abstract read via PubMed; full text not read. Cleary JM et al. 2021
  2. E2established confidence 5 of 5: Multiple independent lines of evidence; treated as settled in the field.paper

    Bayesian hierarchical modelling of patient subpopulations allows a phase 2 oncology trial to borrow strength across subgroups while adapting the degree of borrowing to the observed between-subgroup heterogeneity, improving efficiency relative to independent subgroup analyses.

    PMID 23983156DOI
    What this evidence cannot show

    A methods paper with simulated operating characteristics; performance depends on the number of subgroups and the prior on between-group variance, and frequentist type I error is not controlled by default. It has not been applied to KMT2A partner strata.

    Evidence that argues against this (1)
    • Borrowing across subgroups assumes exchangeability, and infant KMT2A-rearranged disease is where that assumption is least safe: it differs from older-child disease in cell of origin, in HOXA output, in mutation burden and in outcome. Adaptive borrowing shrinks the amount borrowed when heterogeneity is OBSERVED, which requires enough patients per stratum to observe it — and the small strata are exactly the ones the method is being invoked to rescue. Statistics in Medicine 2006
  3. E3strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.paper

    Tumour-agnostic development of TRK inhibitors across 17 tumour types demonstrated that a biomarker-defined basket can support regulatory approval and can include infants.

    Efficacy of Larotrectinib in TRK Fusion-Positive Cancers in Adults and Children · Drilon A, Laetsch TW, Kummar S, et al. · New England Journal of Medicine 2018
    PMID 29466156DOI
    What this evidence cannot show

    Very large effect size; a basket succeeds most easily when the response rate is high enough that small cohorts are informative, which is not the situation in relapsed acute leukemia where CR+CRh rates are around 23%.

    Evidence that argues against this (1)
    • The TRK precedent is real but the entry should carry the counter-case alongside it: two of the NCI-MATCH fusion- and mutation-defined arms above returned response rates of 2-3% in exactly the design this entry advocates. The larotrectinib result is the tail of the distribution of basket outcomes, not its centre, and using it as the model for a KMT2A partner basket imports a survivorship bias. The comparison is direct because EAY131-R explicitly included fusion-positive patients and still failed. Both NCI-MATCH abstracts read via PubMed. Johnson DB et al. 2020
Connected to