Experiment: apply all three pressures and ask whether the survivors are the same cell
Expose the same leukemia to three completely different kinds of treatment, one after another, and look closely at the cells that survive each one. If they all look the same, there is a single hiding place that defeats every approach at once.
The same thing, for a scientist
A convergence test rather than a combination test. The same KMT2A::AFF1 starting population is split and driven to resistance independently under three mechanistically unrelated pressures — a menin inhibitor, venetoclax, and CD19-directed T-cell engagement — with single-cell transcriptomic and chromatin profiling of survivors from each arm. The primary estimand is convergence: do the three survivor populations occupy one state or three? Secondary arms add sequential triple pressure, prospective sorting of the candidate state from the untreated population, and BH3 profiling to characterise apoptotic priming. The design is symmetric: convergence supports a common cause, divergence vindicates the orthogonality strategy, and both are worth knowing.
Design
Arm A, parallel independent pressures. One KMT2A::AFF1 starting population (SEM and MV4-11, plus a PDX line if available) split three ways and driven to resistance separately under revumenib, venetoclax, and CD19 T-cell engagement (blinatumomab with donor T cells, or a CD19 CAR co-culture). Single-cell RNA-seq and ATAC on survivors from each arm plus the untreated parent.
Arm B, sequential triple pressure. Sequential exposure in all six orders, to test whether the order of pressures determines the survivor state — which distinguishes a pre-existing common state from one induced by whichever pressure came first.
Arm C, prospective isolation. Sort the untreated parent population for the candidate quiescent, primitive state (label retention or a quiescence surface signature derived from Arm A) and test sensitivity to all three agents before any exposure. If the state pre-exists and is triply resistant, that is the strongest version of the result.
Arm D, mechanism. BH3 profiling on sorted states to test the predicted shift in apoptotic priming away from BCL-2, and surface CD19 quantification to test the immune-visibility limb.
The analysis that decides it
Convergence is a quantitative claim and must be pre-specified as one. Embed all survivor and parent cells in a shared latent space, and compute the pairwise distance between arm centroids against a null built from technical replicates within arms. Report the ratio of between-arm to within-arm dispersion with a confidence interval.
Three outcomes, all pre-specified:
- Ratio near 1. Survivors converge; a common cause exists among these three agents.
- Ratio large. Survivors diverge; no common cause among these agents, and the orthogonality strategy is supported.
- Intermediate. Partial convergence, most likely a shared quiescence axis plus arm-specific components. This is the outcome we actually expect, and it needs its own pre-specified interpretation — the shared component should then be characterised and tested in Arm C rather than being described qualitatively and dropped.
What makes this hard
Driving resistance to a T-cell engager in vitro is the difficult arm: it requires sustained effector co-culture, and outgrowth can reflect T-cell exhaustion rather than leukemic resistance. Effector-side controls (fresh effectors added at each passage, effector function assayed) are mandatory, and without them the immune arm is uninterpretable.
Driving resistance in vitro also selects for culture adaptation in all arms equally, which the untreated parallel-passage control addresses. And the three agents kill at very different rates, so exposure must be titrated to matched selective pressure rather than matched concentration — otherwise the arms differ in how hard they selected, not only in how.
| design | One KMT2A::AFF1 starting population split and driven to resistance independently under a menin inhibitor, venetoclax and CD19-directed T-cell engagement, with single-cell RNA-seq and ATAC on survivors from each arm; plus sequential triple pressure in all six orders, prospective sorting of the candidate quiescent state from the untreated parent, and BH3 profiling of sorted states. |
|---|---|
| readout | Pre-specified convergence statistic: ratio of between-arm to within-arm centroid dispersion in a shared latent space, with a null built from technical replicates; plus sensitivity of the prospectively sorted state to all three agents, apoptotic priming by BH3 profiling, and surface CD19 density. |
| controls | Untreated parallel-passage arm to absorb culture adaptation; fresh effectors at each passage with effector function assayed, so that engager-arm outgrowth is not T-cell exhaustion; exposures titrated to matched selective pressure rather than matched concentration; PRC1.1-deficient cells as a known venetoclax-sensitive menin-resistant state serving as an internal negative control for convergence. |
| cost tier | high |
| timeline | 15-20 months: resistance derivation 6-9 months, profiling 4 months, sorting and BH3 arms 4 months |
| symmetric value | Convergence identifies the shared vulnerability and nominates a fourth agent; divergence removes the main theoretical objection to the orthogonal-combination strategy. Both outcomes are decision-relevant, which is unusual and is the main argument for the cost. |
| prespecified interpretation of intermediate result | Partial convergence is the expected outcome. The shared component must then be isolated and tested prospectively in the sorting arm rather than described qualitatively and set aside. |
tests hypotheses
[ "hyp-common-cause-failure-defeats-orthogonal-combination", "hyp-menin-plus-engager-orthogonal-pre-emption" ]
collaborators needed
[ "an immunology group able to run sustained engager or CAR co-culture selection", "a single-cell multi-omics facility", "a BH3 profiling laboratory" ]
- A divergent result would be read as vindication of combination therapy generally, which it would not be — it would show only that these three specific agents lack a shared escape. The write-up must resist that generalisation.
- If a common cause is found and it is the quiescent stem state, the countermeasure targets the compartment normal haematopoiesis depends on, so a positive result names a vulnerability that may not be safely druggable in an infant.
- Can resistance to a T-cell engager be derived in vitro in a way that reflects clinical escape, or does the immune arm require an in vivo humanised model to mean anything?
- Should lineage switch be included as a fourth pressure, given that it is the dominant clinical escape route from CD19-directed therapy in this genotype?
- E1strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.paper
BET-inhibitor resistance in MLL-AF9 models emerged from leukaemia stem cells, was cross-resistant to both pharmacological and genetic BET inhibition and was not efflux-mediated, establishing both the precedent for a state-based resistance unit and the experimental approach of driving resistance and profiling survivors.
BET inhibitor resistance emerges from leukaemia stem cells · Fong CY, Gilan O, Lam EYN, et al. · Nature 2015What this evidence cannot show
Within-class cross-resistance in a murine model. The cross-class question is the novelty here and has no precedent to build on.
Evidence that argues against this (1)
- RE-SCOPE THE SCREEN. The premise — that a single state confers cross-resistance to menin inhibition, BCL-2 inhibition and CD19-directed engagement — is contradicted by the clinical performance of the corresponding combination: an all-oral revumenib plus decitabine/cedazuridine plus venetoclax regimen gave composite complete remission of 71%, CR/CRh 60%, MRD negativity in 80% of responders and median CR/CRh duration of 10.5 months in relapsed/refractory AML including 52% venetoclax-pretreated patients, and the frontline triplet reached composite CR of 81.4% with universal centralised flow-MRD negativity. What remains genuinely unanswered, and what this screen is uniquely able to answer, is the *convergence* question: whether the rare survivors of sequential triple pressure occupy one transcriptional state or three. That is a question about the residual below the MRD floor, and the screen should be designed and reported as such rather than as a test of whether combinations work. Issa GC et al. 2026
- E2strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.paper
BET-resistant leukemias are uniformly characterised by rapid restoration of MYC transcription through activation and recruitment of WNT signalling components at a focal MYC enhancer, with remodelled regulatory landscapes, showing that survivor profiling can yield a specific, targetable mechanism rather than a vague state description.
Transcriptional plasticity promotes primary and acquired resistance to BET inhibition · Rathert P, Roth M, Neumann T, et al. · Nature 2015What this evidence cannot show
Cited for the methodological precedent that survivor profiling can be mechanistically productive, not for a claim about the three agents tested here.
Evidence that argues against this (1)
- BET inhibition's proximate output is MYC transcription, so restoration of MYC through WNT is the escape route specific to that target. Treating it as a common-cause template for a triple-pressure screen assumes the three pressures share a bottleneck, which is the hypothesis the screen exists to test. Oncofusion knowledge graph (internal analysis) 2026
- E3strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.paper
PRC1.1-deficient menin-inhibitor-resistant cells show reduced monocytic signatures, become BCL-2-dependent and are resensitised to menin inhibition by venetoclax, providing a known menin-resistant state that is specifically venetoclax-sensitive and therefore a built-in negative control for convergence.
Epigenetic regulation of noncanonical menin targets modulates menin inhibitor response in acute myeloid leukemia · Zhou X, Zhang L, Aryal S, et al. · Blood 2024What this evidence cannot show
If the menin arm converges on a PRC1.1-deficient-like state, the experiment should find it venetoclax-sensitive, which would be an internal validity check as well as a partial negative.
Evidence that argues against this (1)
- The PRC1.1-deficient state is the entry's model of a shared resistant state, but it is venetoclax-*vulnerable* and venetoclax-resensitised, i.e. it is a counterexample to common-cause cross-resistance rather than an instance of it — and the clinical triplet containing venetoclax is correspondingly effective. Searched for any published menin-inhibitor-resistant state demonstrated to be simultaneously venetoclax-resistant and immunotherapy-resistant; none found. Recorded so the gap is known to have been checked rather than skipped: the shared state this screen seeks has, as of this search, never been documented. Q12-analogy-counter (search record) 2026
- E4strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.paper
Post-immunotherapy lineage switch in KMT2A-rearranged disease occurs at a median of 1.5 months, frequently from documented MRD-negative remission, indicating that immune escape proceeds through a state transition originating in a compartment that survives deep cytoreduction.
Project EVOLVE: an international analysis of postimmunotherapy lineage switch, an emergent form of resistance · Project EVOLVE consortium · Blood 2025What this evidence cannot show
Clinical observation motivating the immune arm; in vitro engager resistance may not reproduce the in vivo switch phenomenon.
Evidence that argues against this (1)
- Pooling lineage switch with menin and BCL-2 escape as three faces of one state may be a category error. Whole-genome phylogenetic reconstruction of a post-CD19 lineage switch found the myeloid disease descended from the initial ALL via a genetically distinct subclone whose driver mutations were detectable before the switch — prior diversification followed by selection, not a shared plastic state. If lineage switch is clonal selection and menin escape is state-based, a screen that applies all three pressures to one pooled population will produce survivors whose convergence or divergence reflects the mixture of mechanisms rather than a common cause. The screen should score genotype as well as transcriptional state in the survivors. Coorens THH et al. 2023
re-scope: the cross-resistance premise is contradicted by triplet trial results; the screen's remaining unique contribution is the convergence question below the MRD floor
convergence test across three independently derived resistant populations
the same experiment can vindicate the orthogonality strategy if survivors diverge