Fault trees, minimal cut sets and common-cause failure — reliability engineering on a buffered oncogenic program
also called minimal cut set · beta-factor · common-cause failure · defence in depth
Engineers who design systems that must not fail have exact tools for finding the smallest combination of parts whose simultaneous failure brings the system down, and for the uncomfortable fact that backup systems tend to fail together for shared reasons. The leukemia program is a fault-tolerant system, and neither tool has been pointed at it.
The same thing, for a scientist
The HOXA9/MEIS1 output of a KMT2A fusion is buffered. HOXA9 is required for survival of established human KMT2A-rearranged leukemia yet dispensable for murine Mll-AF9 leukemia initiation, with Hoxa9-null leukemias merely more immature, and the original authors concluded that no single Hoxa gene is likely necessary and proposed a Hox code minimally defined by the Hoxa5-a9 cluster [E1]. Meis1 knockdown delays but does not prevent leukemia [E2]. Deleting the entire HoxA cluster produces a haematopoietic effect much stronger than Hoxa9 inactivation alone [E3]. Every one of those sentences is a reliability-engineering statement about redundancy, and the discipline has two formal tools for exactly this situation: minimal cut set analysis, which finds the smallest set of components whose simultaneous failure brings the system down, and common-cause failure analysis, which quantifies the degree to which nominally independent redundant channels fail together. The first generates a combination-design hypothesis; the second generates the sharpest available attack on the field's current combination logic.
Minimal cut sets — and why fractional multi-node inhibition may beat deep single-node inhibition
A minimal cut set is the smallest collection of component failures that produces system failure. In a redundant system, no single component is a cut set of size one; that is what redundancy means, and it is exactly what the Hoxa9 knockout result says 1.
The drug-discovery consequence is not obvious and is worth stating carefully. If the program fails only when a set of nodes is jointly degraded, then total inhibitory effort is better spent spread across the set than concentrated on one member — three nodes at 40% each may cross the failure threshold where one node at 90% does not. That is testable, because CRISPRi libraries of systematically attenuated guides can titrate expression continuously and have already been shown to reveal sharp, gene-specific expression thresholds for cellular behaviours 4. Nobody appears to have run the combinatorial version in a leukemia program.
For an infant, the toxicity arithmetic makes this more than a curiosity: three agents at sub-toxic fractional doses against one agent at maximum tolerated dose is a genuinely different safety profile.
Common-cause failure — the beta-factor, and the attack on orthogonality
Reliability engineering's most durable empirical finding is that redundant systems almost never fail through independent simultaneous failures; they fail through a common cause that defeats all channels at once — a shared power supply, a shared maintenance error, a shared design flaw. The beta-factor is the fraction of a component's failure rate attributable to common cause, and it dominates the reliability of any well-designed redundant system.
Translated: the graph's hyp-menin-plus-engager-orthogonal-pre-emption argues that combining mechanistically orthogonal agents extends durability, and lists among its own falsifiers the discovery of a single mechanism conferring cross-resistance to menin inhibition, venetoclax and CD19-directed therapy simultaneously. Reliability engineering says that is not an exotic failure mode to be hoped against — it is the expected dominant failure mode of any redundant design, and the right response is to go looking for it deliberately rather than to treat it as a residual risk. A candidate exists: BET-inhibitor resistance in MLL-AF9 emerges from leukaemia stem cells, is cross-resistant to genetic BET depletion and is not efflux-mediated 5, which is the signature of a state rather than a lesion.
Where the engineering frame is wrong
Fault trees assume a fixed system architecture whose components fail but do not adapt. A leukemia rewires: knocking down one node changes the dependency structure of the others, so the tree is not static and a cut set measured today may not be a cut set under drug. That is a real and serious disanalogy, and it means the combinatorial titration experiment must measure the dependency structure under perturbation rather than assuming it from baseline.
Second, reliability engineering optimises for the system not failing. Here we want the system to fail, which inverts every sign but also removes the usual engineering safety margins from the reasoning — a cut set that works 95% of the time is a triumph for a reactor and a relapse for a child.
| domain | engineering |
|---|---|
| source concept | fault tree analysis, minimal cut sets, common-cause failure and the beta-factor, defence in depth |
| principal disanalogy | Fault trees assume a static architecture. A leukemia rewires under perturbation, so the dependency structure must be measured under drug rather than inferred from baseline. |
mapping
[
{
"bio": "the HOXA9/MEIS1 self-renewal program",
"analog": "a fault-tolerant system that must fail for therapy to succeed"
},
{
"bio": "HOXA9, MEIS1, PBX3, MEF2C, the Hoxa5-a9 code and other program nodes",
"analog": "redundant components, none of which is a cut set of size one"
},
{
"bio": "partial knockdown giving partial program suppression",
"analog": "degraded but functioning operation — the defining behaviour of a redundant system"
},
{
"bio": "the smallest set of nodes whose joint partial inhibition collapses the program",
"analog": "the minimal cut set"
},
{
"bio": "three agents at fractional dose versus one at maximum tolerated dose",
"analog": "attacking a cut set rather than over-engineering an attack on one component"
},
{
"bio": "a quiescent or primitive leukemia stem cell state resistant to several mechanistically distinct agents at once",
"analog": "common-cause failure — the shared vulnerability that defeats every redundant channel simultaneously"
},
{
"bio": "combining mechanistically orthogonal drugs",
"analog": "defence in depth, whose real reliability is set by the beta-factor and not by the number of layers"
}
]generated hypotheses
[ "hyp-minimal-cut-set-fractional-multinode-inhibition", "hyp-common-cause-failure-defeats-orthogonal-combination" ]
- If the program's minimal cut set is large, then no achievable combination collapses it and the ceiling on this entire therapeutic approach is set by redundancy rather than by drug quality.
- If a common cause exists, orthogonal combination therapy delivers the toxicity of three agents and the durability of one, which would be the worst outcome available and is currently nobody's working hypothesis.
- What is the minimal cut set of the KMT2A::AFF1 output program, and does it have three members or ten?
- Is there a single cell state conferring cross-resistance to menin inhibition, BCL-2 inhibition and T-cell engagement at once, and has anyone looked for one deliberately rather than treating it as a residual risk?
- Does the dependency structure measured at baseline survive perturbation, or does each knockdown redraw the tree?
- E1strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.paper
Mll-AF9 knock-in mice with or without Hoxa9 both showed myeloid expansion at 8 weeks and developed myeloid leukemia with similar incidence and time course, with Hoxa9-null leukemias more immature in phenotype; Mll-AF9 overexpressed Hoxa5, Hoxa6, Hoxa7, Hoxa9 and Hoxa10, and the authors concluded that none of these as a 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 2003What this evidence cannot show
Mouse Mll-AF9 and leukemia initiation, not human KMT2A::AFF1 maintenance. HOXA9 knockdown does kill established human KMT2A-rearranged lines, so initiation and maintenance differ and the redundancy claim is strongest for initiation.
Evidence that argues against this (1)
- The independence assumption the entry needs is the one the source discipline says you cannot make. NUREG-0492: '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 failure analysis matters because such failures 'can often dominate the random hardware failures.' HOXA5-A9 and MEIS1 are not independent redundant channels; they are co-regulated outputs of one fusion complex binding one set of loci, which is the textbook common-cause configuration. The entry is right that the discipline has two tools here, but the two tools point in opposite directions for this system: the cut-set tool suggests spreading effort across the HOXA nodes, and the common-cause tool says those nodes will fail together anyway because they share an upstream cause, so the leverage is upstream. Vesely WE et al. 1981
- E2strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.paper
shRNA knockdown of Meis1 in murine Mll-AF9 leukemia cells decreased proliferation and survival, reduced clonogenic capacity, increased monocytic differentiation and significantly delayed — but did not prevent — establishment of leukemia in transplantation recipients.
A role for MEIS1 in MLL-fusion gene leukemia · Kumar AR, Li Q, Hudson WA, et al. · Blood 2008What this evidence cannot show
shRNA knockdown is partial by construction, so 'delayed but not prevented' may reflect incomplete depletion rather than genuine buffering — which is precisely the ambiguity a titration experiment resolves.
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. "Delayed but not prevented" from shRNA is also the expected result of incomplete knockdown in a transplant that selects for cells retaining expression, so it does not cleanly establish that the node is outside the minimal cut set. Oncofusion knowledge graph (internal analysis) 2026
- E3strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.paper
Mice lacking the entire HoxA cluster in mesodermal lineages showed mild skeletal phenotypes but a striking haematopoietic effect much stronger than that seen for Hoxa9 inactivation alone, affecting haematopoietic stem cells and the erythroid lineage, indicating that several Hoxa genes are jointly necessary for normal haematopoiesis.
Additive and global functions of HoxA cluster genes in mesoderm derivatives · Di-Poi N, Koch U, Radtke F, Duboule D · Developmental Biology 2010What this evidence cannot show
Normal haematopoiesis in mice, not leukemia. It demonstrates that the cluster is jointly required in the physiological setting, and the leukemic dependency structure could differ.
Evidence that argues against this (1)
- Deleting the entire HoxA cluster in MESODERMAL lineages is a developmental experiment in normal haematopoiesis, not a test of what an established leukaemia needs. It shows the cluster is jointly required for building the system, which is a different claim from joint requirement for maintaining a transformed clone. Oncofusion knowledge graph (internal analysis) 2026
- E4strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.paper
CRISPRi libraries of systematically attenuated guide RNAs titrate gene expression across a continuum in human cells including K562, and single-cell RNA-seq readout of cells staged along that continuum revealed sharp transitions in cellular behaviours at gene-specific expression thresholds; a compact library titrating around 2,400 essential genes was constructed.
Titrating gene expression using libraries of systematically attenuated CRISPR guide RNAs · Jost M, Santos DA, Saunders RA, et al. · Nature Biotechnology 2020What this evidence cannot show
A methods paper establishing that graded knockdown and threshold detection are feasible. It does not do the combinatorial version, which is the gap.
Evidence that argues against this (1)
- WHERE THE ANALOGY BREAKS. Minimal cut set analysis is a Boolean method. The standard reference states plainly 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', and defines a minimal cut set as the smallest set of basic events whose simultaneous *occurrence* causes the top event. The formalism has no state between working and failed. So the analogy licenses the idea of a small jointly-necessary set — which is genuinely useful here — but it licenses nothing at all about *fractional* suppression, which is the prediction `hyp-minimal-cut-set-fractional-multinode-inhibition` derives from it. That prediction has to be argued from dose-response pharmacology on its own merits. Note too that the enabling CRISPRi paper cited here found sharp behavioural transitions at *gene-specific* thresholds, which is a per-node threshold model, not the aggregate-suppression model. Vesely WE et al. 1981
- E5strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.paper
BET-inhibitor resistance in MLL-AF9-immortalised murine haematopoietic progenitors emerged from leukaemia stem cells, was cross-resistant to JQ1 and to genetic BET knockdown, and was not efflux- or metabolism-mediated — the signature of a resistant cell state rather than a target lesion.
BET inhibitor resistance emerges from leukaemia stem cells · Fong CY, Gilan O, Lam EYN, et al. · Nature 2015What this evidence cannot show
A different drug class and a murine model. It is offered as the leading candidate for a common cause, not as evidence that one exists for the menin-venetoclax-immunotherapy triple.
Evidence that argues against this (1)
- The common-cause reading of BET-inhibitor resistance generated a prediction that combinations of orthogonal agents would underdeliver. Two menin-inhibitor combination trials published since contradict the strong form: an all-oral revumenib plus decitabine/cedazuridine plus venetoclax regimen gave composite complete remission of 71% with median CR/CRh duration of 10.5 months in relapsed/refractory AML (52% venetoclax-pretreated), and the frontline azacitidine plus venetoclax plus revumenib combination gave composite CR of 81.4% with universal centralised flow-MRD negativity. A resistant *cell state* shared across mechanistically orthogonal pressures should have blunted both. The common-cause tool remains the right question to ask; its strongest answer in this disease is now data-contradicted. Issa GC et al. 2026
the source formalism is Boolean (NUREG-0492: 'events either occur or they do not occur'), so minimal cut set analysis supplies no warrant for a fractional-suppression prediction
- analogous to →How much HOXA9 and MEIS1 is actually required? Partial dependence, a Hox code, and why this caps what any single target can do
the buffered HOXA9/MEIS1 program read as a fault-tolerant system with a minimal cut set
- analogous to →Which step is rate-limiting, and what compensates when you block it — the bite-back map for the causal chain
compensation as redundancy, formalised
common-cause failure is the expected dominant failure mode of any redundant design, not a residual risk
generated by minimal cut set analysis
generated by common-cause failure analysis and the beta-factor