Experiment: combinatorial dose-titration of the HOXA9/MEIS1 program to find its minimal cut set
Use a gene-dimmer-switch technique to turn down several leukemia genes at once, by controlled amounts, in every combination, and find the smallest group that has to be dimmed together before the leukemia program switches off.
The same thing, for a scientist
A pooled dual- and triple-guide attenuated-CRISPRi screen across candidate program nodes in KMT2A::AFF1 models, with graded knockdown per node, single-cell transcriptomic and phenotypic readout, and response-surface fitting to distinguish independent-threshold, aggregate-threshold and multiplicative models of program failure. The enabling technology exists — attenuated-guide libraries titrate expression continuously and reveal sharp gene-specific thresholds [E1] — and the combinatorial extension is the novelty. The critical comparison is at matched total suppression, not at matched guide number.
Design
Nodes. HOXA9, MEIS1, PBX3, MEF2C, HOXA5-HOXA8, plus MEN1 and DOT1L as upstream comparators and two neutral genes as controls. Node selection should be justified from the existing dependency ranking in ana-partner-dependency-vs-essentiality rather than chosen by reputation.
Library. Attenuated guides at four graded activity levels per node, in single, dual and triple combinations. Full factorial is infeasible; use a sparse but balanced design so that every node appears at every level in combination with every other node at least a specified number of times, and fit the surface rather than testing cells.
Models. SEM and MV4-11 CRISPRi lines, plus a KMT2A::AFF1 PDX-derived CRISPRi model if obtainable, and a KMT2A-germline line as specificity control.
Readout. Pooled guide abundance over time as the primary fitness readout; single-cell RNA-seq on a subset to place cells on the program-suppression continuum and to detect differentiation; colony replating on selected combinations as the self-renewal readout that actually matters.
Validation of knockdown. Every claim about aggregate suppression depends on knowing the actual suppression achieved, so per-node knockdown must be measured directly at each guide level rather than assumed from the library design. This is the step that is easiest to skip and most fatal to skip.
Analysis — fitting the surface, not testing a combination
Fit three pre-specified models to the phenotype as a function of the measured per-node suppression vector: independent thresholds, weighted aggregate threshold, and multiplicative residual activity. Compare by cross-validated predictive accuracy on held-out combinations, which is a stronger test than in-sample fit and guards against a flexible model winning by flexibility.
Report the fitted minimal cut set with an uncertainty set, and separately report the best cut set restricted to nodes with existing chemical matter, because a genetically optimal set with no drugs is not actionable.
Pre-specify the comparison of interest: phenotype at three nodes suppressed 40% each versus one node suppressed 90%, at matched measured total suppression. Everything else is secondary.
The pharmacological bridge, which is the part that decides anything
A genetic result does not transfer automatically, because fractional occupancy of a protein by a drug is not the same perturbation as reduced protein level — an occupied transcription factor may still scaffold, and a partially depleted one cannot. The experiment should therefore include a small pharmacological arm: a menin inhibitor, a DOT1L inhibitor and a CDK9 inhibitor at fractional doses in a matched three-by-three-by-three grid, asking whether the surface shape reproduces.
If the genetic and pharmacological surfaces disagree, that disagreement is the most interesting result the experiment can produce and should not be treated as a failure.
| design | Pooled dual- and triple-guide attenuated-CRISPRi screen at four graded knockdown levels per node across HOXA9, MEIS1, PBX3, MEF2C, HOXA5-HOXA8 plus MEN1 and DOT1L comparators, in SEM and MV4-11 CRISPRi lines and a KMT2A::AFF1 PDX-derived model where obtainable, with a sparse balanced combination design and response-surface fitting; plus a small pharmacological arm using fractional doses of a menin inhibitor, a DOT1L inhibitor and a CDK9 inhibitor in a matched grid. |
|---|---|
| readout | Pooled guide abundance over time; single-cell RNA-seq on a subset to place cells on the program-suppression continuum; colony replating on selected combinations as the self-renewal endpoint; directly measured per-node knockdown at every guide level. |
| controls | Neutral-gene guides; KMT2A-germline line as specificity control; HOXA9 single-node deep knockdown as the positive control from published shRNA data; degron-based complete fusion removal as a calibration ceiling; measured rather than assumed knockdown depth. |
| cost tier | high |
| timeline | 12-18 months: library design and validation 4 months, screen 4 months, single-cell and replating validation 4 months, pharmacological arm 3 months |
| go no go | If measured per-node knockdown does not span a usable range at the four designed levels, stop and re-engineer the library before screening; an unmeasured suppression axis makes the entire response surface uninterpretable. |
| prespecified primary comparison | Phenotype at three nodes suppressed 40% each versus one node suppressed 90%, at matched measured total suppression. |
| model comparison plan | Independent-threshold, weighted-aggregate-threshold and multiplicative models fitted and compared by cross-validated predictive accuracy on held-out combinations, not by in-sample fit. |
| decision value | Either establishes that combination design should optimise coverage across a set rather than potency at one target, or shows that the maintenance program has a genuine single-node dependency, which would strongly simplify the field's target strategy. |
tests hypotheses
[ "hyp-minimal-cut-set-fractional-multinode-inhibition" ]
collaborators needed
[ "a functional genomics group with attenuated-guide CRISPRi capability", "a computational group for sparse response-surface design and fitting" ]
- This is the most expensive experiment in this beat and the one with the highest chance of an uninterpretable result, because sparse response surfaces over rewiring systems are hard to fit. The go/no-go on measured knockdown depth is the main protection.
- A positive result recommends multi-agent sub-toxic regimens in infants, which is a large clinical step to take on the back of a cell-line response surface.
- Is a sparse balanced triple-combination design at four levels per node achievable within a library size a single screen can support?
- Does fractional pharmacological occupancy behave like fractional expression knockdown for a transcription factor, and if not, does the genetic result transfer at all?
- E1strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.paper
Attenuated-guide CRISPRi libraries titrate gene expression across a continuum in human cells including K562, with deep-learning-derived mismatch rules and a compact library covering around 2,400 essential genes; single-cell RNA-seq of cells staged along the 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 2020What this evidence cannot show
Single-gene titration; the dual and triple extension increases library size multiplicatively and is the main feasibility risk.
Evidence that argues against this (1)
- The design's governing metaphor does not license its comparison. Minimal cut set analysis is Boolean — the standard reference states that 'events either occur or they do not occur' and defines a cut set by the *simultaneous occurrence* of basic events — so 'fractional suppression across cut-set members' has no counterpart in the source method and the prediction must stand on dose-response biology alone. Concretely, the experiment needs an operational definition of 'equal total inhibitory effort' that is not supplied by the analogy: summed percentage knockdown, summed occupancy, and summed toxicity-equivalent dose give different rankings, and the result will depend on which is chosen. That choice should be pre-registered before the titration runs. Vesely WE et al. 1981
- E2strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.paper
No single Hoxa gene appears necessary for Mll-AF9 leukemia, with Hoxa9-null mice developing leukemia at similar incidence and time course, and the authors proposed a Hox code minimally defined by the Hoxa5-a9 cluster — the node set this experiment titrates.
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 and initiation; the node list it suggests should be supplemented from human dependency data.
Evidence that argues against this (1)
- The cut-set members proposed are not independent channels, which is the configuration the same handbook flags as the dominant failure mode of the method: 'The primary failures ... do not necessarily have to be independent. A single, more basic cause may result in multiple failures which fail the system', and such common cause failures 'can often dominate the random hardware failures.' HOXA5-A9 and MEIS1 are co-regulated outputs of one fusion complex at a shared set of loci. Partial knockdown of several of them is therefore closer to partial knockdown of their shared driver than to redundant-channel attack, and a positive result would be difficult to attribute. The control the design needs, and does not currently have, is titration of the shared upstream node (the fusion itself, or menin) across the same effort range. Vesely WE et al. 1981
- E3strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.paper
HOXA9 depletion is selectively lethal across seven human KMT2A-rearranged lines among seventeen tested (p=0.007), establishing that at least one node has a strong single-gene phenotype in the human maintenance setting and providing a positive control for the screen.
HOXA9 is required for survival in human MLL-rearranged acute leukemias · Faber J, Krivtsov AV, Stubbs MC, et al. · Blood 2008What this evidence cannot show
shRNA rather than CRISPRi, and the achieved knockdown depth was not titrated, which is why this result does not by itself settle the independent-versus-aggregate question.
Evidence that argues against this (1)
- A well-chosen single node already produces the phenotype this experiment is designed to show requires several. Menin-MLL1 inhibition — one interface — evicted MLL1 and the fusion oncoprotein from chromatin at maintenance loci, collapsed the HOXA cluster and Meis1, induced differentiation markers and prolonged survival in patient-derived xenografts. Together with the HOXA9 shRNA lethality recorded in this evidence item, the realistic prior is that the experiment will find a single-node arm that matches or beats the distributed arm, and the design should be powered to detect that outcome rather than treating it as a failure mode. Heikamp EB et al. 2022
- E4strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.paper
Degron-mediated removal of the fusion protein itself collapses the target program with elongation changes within 15 minutes, providing an upper-bound reference phenotype against which any combination of downstream node suppressions can be calibrated.
MLL::AF9 degradation induces rapid changes in transcriptional elongation and subsequent loss of an active chromatin landscape · Olsen SN, Godfrey L, Healy JP, et al. · Molecular Cell 2022What this evidence cannot show
A different fusion; used as a calibration ceiling, not as a claim about KMT2A::AFF1.
Evidence that argues against this (1)
- Degron removal is COMPLETE and near-instantaneous; a drug gives partial, reversible occupancy. Using the degron phenotype as the upper-bound reference therefore sets a bar that no pharmacological combination can reach, and the interesting question — what fraction of occupancy suffices — is the one a complete-removal experiment cannot address. Journal of Clinical Oncology 2025
- related to →KMT2A::AFF1 (MLL-AF4), t(4;11)(q21;q23)
missing control arm: titrate the shared upstream node (the fusion, or menin) across the same effort range, because the proposed cut-set members are co-regulated rather than independent channels
combinatorial attenuated-guide CRISPRi response surface with a pharmacological bridge arm
node selection should follow the existing dependency-versus-essentiality ranking