Oncofusion

start / graph / hyp-minimal-cut-set-fractional-multinode-inhibition

Hypothesisspeculative confidence 1 of 5: Our own inference or analogy. Not shown to be true.Q12-analogy

Hypothesis: the fusion output program fails at a threshold on total cut-set suppression, so three nodes at 40% beat one node at 90%

No single gene in the leukemia program is essential on its own; knocking one down only partly works. If the program collapses when enough of a small group of genes is suppressed together, then partly suppressing three of them could work better, and be far less toxic, than nearly eliminating one.

The same thing, for a scientist

The HOXA9/MEIS1 program is buffered: no single Hoxa gene is necessary for Mll-AF9 leukemia and the originating authors proposed a Hox code minimally defined by the Hoxa5-a9 cluster [E1]; Meis1 knockdown delays but does not prevent leukemia [E2]; whole-HoxA-cluster deletion has a far stronger haematopoietic effect than Hoxa9 deletion alone [E3]. We hypothesise that program failure is governed by a threshold on aggregate suppression across a small minimal cut set rather than by the depth of suppression at any single node, and therefore that fractional multi-node inhibition is more efficient per unit of total inhibitory effort than deep single-node inhibition. The prediction is directly testable with attenuated-guide CRISPRi, which titrates expression continuously and has shown sharp gene-specific behavioural thresholds [E4]. The combinatorial titration required has, as far as we can find, never been performed on an oncogenic program in leukemia.

Go deeper
The precise form of the claim

Let s_i be fractional suppression of node i. Three candidate models:

  • Independent-node model. The program fails when any single s_i exceeds a node-specific threshold. Current drug development implicitly assumes this, which is why potency at one target is the optimisation objective.
  • Aggregate-threshold model (this hypothesis). The program fails when a weighted sum of s_i over the cut set exceeds a threshold. Spreading effort is then at least as good as concentrating it, and better once single-agent toxicity becomes limiting.
  • Multiplicative model. The program fails when the product of residual activities falls below a threshold, which strongly favours spreading, since 0.6 cubed is 0.22 while 0.1 at one node leaves the others untouched.

The experiment distinguishes these three by fitting a response surface, not by testing a single combination. That framing matters, because 'we tested a triple combination and it worked' would not distinguish them and is how this question usually gets answered badly.

Why the toxicity arithmetic is the real argument

In an adult with AML, a modest efficiency gain from spreading inhibition is interesting. In a one-year-old, three agents at sub-toxic fractional doses versus one at maximum tolerated dose is a categorically different proposition, because the dose-limiting toxicities of the candidate agents are largely non-overlapping — differentiation syndrome and QTc for menin inhibitors, cytopenias for venetoclax, marrow suppression for CDK9 inhibitors.

Non-overlapping toxicity is what makes spreading feasible. It is also, notably, the same property as mechanistic orthogonality viewed from the safety side, which means the cut-set argument and the orthogonality argument recommend overlapping but not identical combinations — and where they disagree is informative.

The honest problems

First, the nodes of the program are not all druggable, so a cut set identified genetically may have no pharmacological implementation. The experiment should therefore weight candidate nodes by existing chemical matter, and report the best druggable cut set alongside the best cut set.

Second, the system rewires. Knocking down HOXA9 may change the dependency on MEIS1, so the response surface is not separable and a full factorial titration across even five nodes at four levels is 1,024 conditions. The design has to be a pooled CRISPRi screen with dual or triple guides and computational surface fitting, not an arrayed factorial, and the resulting surface will be sparse.

Third, HOXA9 knockdown does kill established human KMT2A-rearranged lines with selectivity (p=0.007 across 17 lines) 5, which is evidence for an independent-node model in the maintenance setting even though the mouse initiation data argue for redundancy. Initiation and maintenance may simply have different architectures, and if so this hypothesis applies to maintenance only if the human maintenance data are wrong, or applies to both with different cut sets. That tension is unresolved and is the strongest reason the prior is not higher.

Facts and figures
statementFailure of the KMT2A::AFF1 output program is governed by a threshold on aggregate suppression across a small minimal cut set of program nodes rather than by depth of suppression at any single node; consequently, at equal total inhibitory effort, fractional suppression distributed across three or more cut-set members produces greater program collapse and greater loss of self-renewal than near-complete suppression of one member.
originanalogy
source analogyanalogy-engineering-fault-tree-minimal-cut-set
prior probability0.35
prior reasoningThe redundancy premise is well supported for leukemia initiation in mouse models but is contradicted for maintenance in human lines, where HOXA9 depletion alone is selectively lethal. If maintenance follows the independent-node model, the hypothesis fails in exactly the setting that matters therapeutically. We set 0.35 to reflect that the premise holds in one setting and is contested in the other, and we regard resolving that tension as the most valuable output of the experiment regardless of which model wins.
decision valueDetermines whether combination design should optimise for potency at one target or for coverage across a set, and whether a sub-toxic multi-agent regimen is a rational strategy in infants rather than merely a hopeful one.
predictions
[
  "A combinatorial attenuated-guide CRISPRi response surface over HOXA9, MEIS1, PBX3, MEF2C and the Hoxa5-a8 genes shows a threshold in aggregate suppression rather than independent per-node thresholds.",
  "At matched total suppression, triple fractional knockdown produces greater loss of colony replating and greater differentiation than single deep knockdown.",
  "The fitted minimal cut set has fewer than five members, and at least two of them have existing chemical matter.",
  "The pharmacological analogue — three agents at fractional dose versus one at maximal dose in a KMT2A::AFF1 model — reproduces the genetic result.",
  "The cut set differs between initiation and maintenance settings, reconciling the mouse and human data."
]
experiments
[
  "exp-cut-set-fractional-crispri-titration"
]
analyses
[
  "ana-partner-dependency-vs-essentiality"
]
falsifiers
[
  "The response surface shows independent per-node thresholds, with deep single-node suppression consistently outperforming distributed suppression at matched total effort.",
  "HOXA9 alone behaves as a cut set of size one in human KMT2A::AFF1 maintenance, as the shRNA data suggest, making the redundancy premise false where it matters.",
  "The fitted minimal cut set has so many members that no achievable combination reaches the threshold.",
  "The pharmacological analogue fails despite a positive genetic result, indicating that fractional occupancy by a drug is not equivalent to fractional expression knockdown — which would be an important negative in its own right."
]
How this could bite us back
  • A large minimal cut set would mean the program cannot be collapsed by any achievable combination, which caps the whole differentiation-therapy approach, not just one drug.
  • Three sub-toxic agents in an infant is three sets of long-term effects, three drug interactions and three sources of adherence failure. Sub-toxic is not non-toxic, and the toxicity arithmetic that motivates the hypothesis could turn against it once late effects are counted.
  • Distributed partial inhibition is also the classic recipe for selecting resistance, since no node is fully suppressed and every node is under selection at once.
Still open
  1. Do initiation and maintenance of KMT2A-rearranged leukemia have different dependency architectures, and is that the reconciliation of the mouse and human Hoxa9 data?
  2. Is fractional pharmacological occupancy equivalent to fractional expression knockdown for a transcription factor, or does sub-stoichiometric occupancy behave differently from reduced protein level?
  3. How many of the cut-set members have chemical matter, and what is the best cut set restricted to druggable nodes?
Every source behind this entry (5)
  1. E1strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.paper

    Mll-AF9 knock-in mice with and without Hoxa9 developed myeloid leukemia with similar incidence and time course, with Hoxa9-null leukemias more immature; the fusion overexpressed Hoxa5, Hoxa6, Hoxa7, Hoxa9 and Hoxa10, and the authors concluded no single gene is likely necessary, proposing a Hox code minimally defined by the Hoxa5-a9 cluster.

    Hoxa9 influences the phenotype but not the incidence of Mll-AF9 fusion gene leukemia · Kumar AR, Hudson WA, Chen W, Nishiuchi R, Yao Q, Kersey JH · Blood 2003
    PMID 14615372DOI
    What this evidence cannot show

    Mouse, initiation, and a different fusion partner. The phrase 'Hox code' in this abstract is the closest thing in the literature to a minimal cut set, and it was never operationalised.

    Evidence that argues against this (1)
    • The cut-set frame also assumes the members can be treated as independent channels, and the same handbook says they cannot be assumed so: 'The primary failures (component failures) on a fault tree do not necessarily have to be independent. A single, more basic cause may result in multiple failures which fail the system', and common cause failures 'can often dominate the random hardware failures.' HOXA5-A9, MEIS1 and the rest of the proposed cut set are not independent channels — they are co-regulated outputs of one fusion complex at one set of loci. Fractional inhibition of several co-regulated nodes is therefore not the engineering scenario in which distributed suppression wins; it is closer to attacking one channel several times. The honest form of the experiment is to titrate the shared upstream node and the downstream nodes separately and compare, which is not what `exp-cut-set-fractional-crispri-titration` currently specifies. Vesely WE et al. 1981
  2. E2strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.paper

    Meis1 knockdown in murine Mll-AF9 leukemia cells reduced proliferation, survival and clonogenicity, increased monocytic differentiation, and significantly delayed but did not prevent leukemia establishment in transplant recipients.

    A role for MEIS1 in MLL-fusion gene leukemia · Kumar AR, Li Q, Hudson WA, et al. · Blood 2008
    PMID 19109563DOI
    What this evidence cannot show

    Partial shRNA depletion; 'delayed not prevented' is exactly what incomplete knockdown of a non-redundant node also looks like, which is why titration rather than knockdown is required.

    Evidence that argues against this (1)
    • Murine MLL-AF9 in the myeloid compartment, and the index disease is human infant KMT2A::AFF1 B-lymphoid leukaemia — which, on this graph's own evidence, is HOXA-LOW in infants while these models run on HOXA activation. Incomplete shRNA knockdown in a transplant that selects for retained expression is also the standard confound for a "delayed but not prevented" result. Oncofusion knowledge graph (internal analysis) 2026
  3. E3strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.paper

    Deletion of the entire HoxA cluster in mesodermal lineages produced a haematopoietic effect much stronger than Hoxa9 inactivation alone, affecting stem cells and the erythroid lineage, indicating several Hoxa genes are jointly necessary.

    Additive and global functions of HoxA cluster genes in mesoderm derivatives · Di-Poi N, Koch U, Radtke F, Duboule D · Developmental Biology 2010
    PMID 20303345DOI
    What this evidence cannot show

    Normal haematopoiesis, and whole-cluster deletion is the extreme case rather than a titration.

    Evidence that argues against this (1)
    • Whole-cluster deletion in mesodermal lineages is developmental biology in normal haematopoiesis. Joint requirement for building the system is not joint requirement for maintaining a transformed clone, and no drug removes a gene cluster. Oncofusion knowledge graph (internal analysis) 2026
  4. E4strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.paper

    Attenuated-guide CRISPRi libraries titrate gene expression continuously across human cell models including K562, and single-cell RNA-seq of cells staged along the expression continuum revealed sharp transitions in cellular behaviour at gene-specific expression thresholds.

    Titrating gene expression using libraries of systematically attenuated CRISPR guide RNAs · Jost M, Santos DA, Saunders RA, et al. · Nature Biotechnology 2020
    PMID 31932729DOI
    What this evidence cannot show

    Single-gene titration. The combinatorial extension required here multiplies library complexity and is the main technical risk.

    Evidence that argues against this (1)
    • WHERE THE ANALOGY BREAKS. Minimal cut set analysis is Boolean. The NRC Fault Tree Handbook, the standard reference for the method, states that 'Boolean algebra is especially important in situations involving a dichotomy: switches are either open or closed, valves are either open or closed, events either occur or they do not occur.' A minimal cut set is defined as the smallest set of basic events whose *simultaneous occurrence* causes the top event; there is no representation in the formalism of a component that is 40% failed. So the source discipline supplies no warrant at all for the specific prediction at issue — that fractional suppression distributed across three cut-set members beats near-complete suppression of one. That prediction has to stand on dose-response biology, not on the borrowed vocabulary. Worse for the transfer, the CRISPRi paper cited here as the enabling technology found sharp transitions in cellular behaviour at *gene-specific* expression thresholds, i.e. per-node thresholds, which is the single-node threshold model rather than the aggregate-suppression model the hypothesis proposes. Vesely WE et al. 1981
  5. E5strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.paper

    shRNA HOXA9 depletion across 17 human AML and ALL lines caused proliferation arrest and apoptosis specifically in the 7 KMT2A-rearranged lines (p=0.007), with greater apoptosis in KMT2A-rearranged primary AMLs (p=0.01).

    HOXA9 is required for survival in human MLL-rearranged acute leukemias · Faber J, Krivtsov AV, Stubbs MC, et al. · Blood 2008
    PMID 19056693DOI
    What this evidence cannot show

    This is the principal evidence *against* the redundancy premise in the human maintenance setting, and is cited here for that reason. Reconciling it with the mouse initiation data is the central unresolved tension in this hypothesis.

    Evidence that argues against this (1)
    • A single-node intervention already achieves what the hypothesis says single-node interventions cannot. Menin-MLL1 inhibition — one node, one interface — evicted both MLL1 and the fusion oncoprotein from chromatin at maintenance loci, collapsed the HOXA cluster and Meis1, drove differentiation markers and prolonged survival in patient-derived xenografts of a related fusion family. Combined with the HOXA9 shRNA result recorded in this very evidence item, the picture is that near-complete suppression of a correctly chosen single node does collapse the program. The hypothesis is best re-stated as a claim about *equal total inhibitory effort*, which is a pharmacological statement about dose-limiting toxicity rather than a claim about network architecture — and that restatement needs a toxicity model, which the entry does not have. Heikamp EB et al. 2022
Connected to