The bite-back of striking a target: a wrong strike costs a cure and generates no data that it was wrong
also called cost of a false negative strike · asymmetry of removal
Deciding not to pursue a drug target is a decision with consequences, but unlike a failed experiment it never produces a result telling you it was a mistake. The target just stops being looked at, and nobody finds out.
The same thing, for a scientist
Seven candidates were removed from this project's consideration in cycles 21-22. Removal is an intervention and therefore requires bite-back, but its risk profile is unusual: the error is a false negative, it is silent, and it compounds because a struck entry is not re-examined. This entry states what the specific costs are for these seven, what evidence would have detected an error in each, and the two practices that make the error recoverable. It exists because the cycle-22 audit found that four of ten decisions rested on a reason that fails on reading, and in all four a different reason happened to hold — a list robust to its own reasoning being wrong is a list whose reasoning was not doing the work.
Why this asymmetry is worse than the usual one
A wrongly pursued target produces a failed assay, a negative dossier and a STOP entry in the register. The project learns. A wrongly struck target produces nothing at all: no assay is run, no entry is updated, and the gap is indistinguishable from the gap where a target genuinely never existed. There is no experiment whose result is "the thing you did not look at would have worked."
Three features make it compound:
- Struck entries are not re-read. The reason for the strike is recorded once and thereafter functions as a fact. Four of this cycle's ten decisions show what that costs: nobody had re-read the cited papers, and four citations did not say what they were said to say.
- A strike is cheap and a reversal is expensive. Removing a candidate takes a paragraph. Reinstating one requires arguing against a written decision, which in practice nobody does without new data, which nobody generates for a struck target.
- The instrument that does most of the striking has enumerable blind spots. Three DepMap-based strikes this cycle, and three documented failure modes of the essentiality axis — underpowered constraint, non-proliferative phenotypes, and guide placement on long proteins. BOD1L is the worked example of the third and it was one full-text read away from never existing as an entry.
What each strike costs if it is wrong
| Struck | Cost of the error being real | What would have detected it |
|---|---|---|
| PRMT1 | A biomarker-selected combination in the right disease, already shown to work in patient-derived xenografts: MS023 plus a FLT3 inhibitor, selected on FLT3 R972/973 methylation 1. This is the largest single cost on the list, and the strike does not claim to refute it — it declines it. | Reading the cited paper's conclusion, which is what the audit did. |
| RBBP5, ASH2L | A non-catalytic requirement. Corrected 2026-09-19: the claim that nobody has tested this in KMT2A::AFF1 is withdrawn. ASH2L siRNA has been run in RS4;11, a t(4;11) KMT2A::AFF1 line, and knocks down HOXC8 expression while reducing RBBP5, WDR5 and MLL occupancy at the HOXC8 promoter 5. The strike still replaces an inference with an inference, but the inference is no longer unopposed. | A knockdown in a KMT2A::AFF1 line with a HOX readout - which has now been done once, weakly 5. What has not been run is the same knockdown with a growth, HOXA9/MEIS1 or leukaemia-maintenance readout. |
| CTR9, CDC73 | Little, on these grounds: both are DepMap common-essential, which is an independent and sufficient reason. The error would be in the reasoning, and the reasoning has already been corrected. | The menin consistency test, now a standing rule. |
| CCNT1 | Confusion propagating to CDK9, which is the real risk here rather than losing CCNT1 itself. The alvocidib mortality argument, if it had travelled, would have discouraged the most mechanistically direct target in the fusion's machinery on a confounded comparison in a seven-years-older arm 2. | Reading the trial, including the authors' own conclusion. |
| ELL2 | Little as a target; the paralogue evidence is strong. The cost was in the framing: calling ELL2 'dispensable' would have licensed someone to treat ELL2 blockade as low-risk, when the published phenotype is a broken humoral response in a child who cannot spare one 3. | Reading the cited immunology paper. |
| MEIS1 (demoted) | MEIS1 was not removed but reclassified as a readout, which is arguably its correct role. Low cost. | — |
| KDM6A (reframed) | A reframing, not a removal, and it cuts both ways: high KDM6A tracks venetoclax tolerance, so KDM6A status is a therapy-selection variable that could be used now. | — |
The honest summary is that most of these strikes are cheap to be wrong about, and PRMT1 is not. A published, biomarker-defined combination in MLL-rearranged ALL is being set aside, and the register should record that as a resourcing decision rather than let it read as a scientific refutation.
Correction, 2026-09-19 (absence-vs-library sweep). The RBBP5/ASH2L row asserted that nobody had tested this requirement in KMT2A::AFF1 'in either direction'. Wu et al. did, in RS4;11 and THP-1 5: ASH2L knockdown lowers HOXC8 transcript and protein and lowers RBBP5/WDR5/MLL occupancy at the HOXC8 promoter. The paper is weak evidence - siRNA, a low-impact journal, one non-HOXA target gene, no growth or survival readout, and no fusion-versus-wild-type comparison - so the strike itself is not overturned. But the register said the experiment had not been run, and it had. The corrected statement is: tested once, at the level of a single target gene's transcription, never at the level of leukaemia maintenance.
The two practices that make the error recoverable
Neither is expensive and both are now recorded in concept-phase2-go-no-go-register.
Separate the load-bearing reason from the supporting reasons, and name the result that reopens the decision. A strike whose reasons are listed as an undifferentiated set cannot be reopened by the failure of any one of them, because there is always another. A strike with one named load-bearing reason reopens automatically when that reason fails. All seven entries in the register's struck_cycle22 now carry load_bearing_reason, reason_withdrawn and reopens_if.
Keep the struck entry and its evidence in full. The counter-evidence that argues for a strike lives inside the gene entry alongside the evidence that argued for inclusion, so a reader who arrives at gene-prmt1 sees both. Deleting the entry, or reducing it to its verdict, would make the strike unfalsifiable by removing the material an objection would have to engage with.
What neither practice fixes: nobody is scheduled to re-read these. The reopens_if conditions are conditions, not tasks, and a condition with no owner is a condition nobody checks.
| kind | essential-function |
|---|---|
| severity | Low for six of seven; material for PRMT1, where a published biomarker-selected combination in MLL-rearranged ALL is being declined rather than refuted. |
| mitigation | Two practices, both now recorded in concept-phase2-go-no-go-register: (1) each struck candidate records its load-bearing reason separately from its supporting reasons and names the result that would reopen it, so a load-bearing reason that later fails reopens the decision instead of being propped up; (2) struck entries and their full evidence are retained rather than reduced to a verdict, so an objection has something to engage with. Neither assigns an owner to actually check the reopening conditions, which is the residual hole. |
| error class | false negative, silent, non-self-correcting |
| detection cost | One full-text read per decision. That is what the cycle-22 audit cost and it changed the stated reasoning on four of ten. |
affects
[ "gene-prmt1", "gene-rbbp5", "gene-ash2l", "gene-ctr9", "gene-cdc73", "gene-ccnt1", "gene-ell2", "concept-phase2-go-no-go-register" ]
- This entry argues for making strikes harder to make and easier to reverse, and that has its own cost: a project that cannot close questions cannot allocate anything. Seven reopened questions and no resources is worse than seven closed questions, four of them closed for a reason that has since been corrected.
- Writing down the asymmetry can function as a substitute for acting on it. The cheapest test on this list — a single RBBP5 or ASH2L knockdown in SEM with a HOXA9 readout — is named twice in this graph and has never been run, and adding a third mention does not run it.
- The audit that generated this entry read twelve of fourteen papers at abstract level. If the lesson is that reading the source changes the conclusion, then the audit itself is only partly done, and the next reader should treat its abstract-level verdicts the way it treated the sweep's.
- Who owns the reopening conditions now recorded against each strike? A condition with no owner is never checked, and this is the part of the mitigation that does not work.
- Is declining the PRMT1/FLT3-methylation combination a resourcing decision or a scientific one? The register currently reads as though it were scientific and the audit says it is not.
- How many strikes exist elsewhere in this graph's 435 entries with no load-bearing reason named, and is there a cheap way to find them short of re-reading every entry?
- Should a strike expire? A decision with a review date behaves differently from one without, and nothing in this project currently expires.
- E1strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.paper
The PRMT1 strike sets aside a published, biomarker-defined combination in the right disease: MS023 killing paralleled baseline FLT3 R972/973 methylation in MLL-rearranged ALL cells, and MS023 plus the FLT3 inhibitor PKC412 eliminated MLL-r ALL cells more effectively than PKC412 alone in patient-derived mouse xenografts, with the authors concluding that PRMT1 inhibition is a promising strategy in this disease.
Targeting PRMT1-mediated FLT3 methylation disrupts maintenance of MLL-rearranged acute lymphoblastic leukemia · See PubMed record PMID 31395602 for the full author list (not transcribed here) · Blood 2019What this evidence cannot show
Abstract read, not full text. 'MLL-rearranged ALL' is not resolved to fusion partner in the abstract, so whether KMT2A::AFF1 was among the models is unknown without the methods; PKC412 (midostaurin) is a multi-kinase inhibitor; and a patient-derived xenograft result is preclinical. The point stands at the level it is made: something real is being set aside, and the register should say it is being set aside rather than refuted.
Evidence that argues against this (1)
- The countervailing consideration, and the reason the strike is still right for this project: a screen of Prmt1 shRNA across more than ten MLL and non-MLL fusions found transformation by most MLL fusions unaffected, with only MLL-GAS7 and MOZ-TIF2 dependent and MLL-AF9 recruiting KDM4C but not PRMT1. PRMT1 is not a general KMT2A-fusion dependency, PRMT1 is DepMap common-essential, and the first-in-human PRMT1 inhibitor trial terminated. The FLT3-methylation combination is a different programme in a different framing, not a defence of PRMT1 as a fusion-complex target. Cheung N et al. 2016
- E2strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.paper
The CCNT1 clinical argument, had it propagated to CDK9, would have discouraged the project's most mechanistically direct target on a confounded comparison: the alvocidib arm that carried 7 of 9 treatment-related deaths also had the highest response rate of three arms (28% vs 14% and 16%), was seven years older at median (62 vs 55), received a pan-CDK inhibitor with two cytotoxics, and was recommended by the authors for further clinical development.
A randomized trial of three novel regimens for recurrent acute myeloid leukemia demonstrates the continuing challenge of treating this difficult disease · See PubMed record PMID 30370956 for the full author list (not transcribed here) · American Journal of Hematology 2018What this evidence cannot show
Abstract read, not full text. Arms of 35, 36 and 19 under a Simon two-stage design per arm, never powered for head-to-head safety comparison; the age imbalance is stated in the abstract but no adjusted analysis is reported there. The treatment-related deaths are real and should not be dismissed as a regimen-level signal in a relapsed/refractory population — only as an attribution to P-TEFb inhibition, which the design cannot support.
Evidence that argues against this (1)
- The confound is real and the conclusion still needs a bar to clear. In this disease the bar is now set: AUGMENT-101 treated 94 patients with revumenib, reporting CR+CRh 22.8% and overall response 63.2% in the efficacy-evaluable KMT2A-rearranged population, with grade 3 or worse febrile neutropenia 37.2%, differentiation syndrome 16.0% and QTc prolongation 13.8% - the largest evaluation of a targeted therapy in these patients. A CDK9 or P-TEFb programme is being compared not against nothing but against an approved agent with registration-grade data, and the only clinical evidence for P-TEFb inhibition here remains a multi-CDK inhibitor given with two cytotoxics in 35 patients. Protecting the target from a confounded strike is right; it does not make the target competitive. Issa GC et al. 2025
- E3strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.paper
Describing ELL2 as 'dispensable' misstates the published phenotype: B-cell-conditional Ell2 knockout mice show curtailed humoral and recall responses, reduced splenic plasma cells and marrow IgG1-producing cells, four-fold fewer LPS-stimulated plasmablasts with a paucity of secreted immunoglobulin heavy chain, and severely reduced Ig kappa, ATF6, BiP, Cyclin B2, OcaB and XBP1.
Transcription elongation factor ELL2 drives Ig secretory-specific mRNA production and the unfolded protein response · Park KS, Bayles I, Szlachta-McGinn A, et al. · Journal of Immunology 2014What this evidence cannot show
Abstract read, not full text. Mouse, CD19-cre B-cell-restricted complete deletion rather than systemic partial inhibition, so a drug effect would be milder. It does not argue against striking ELL2 — it argues against the word used to strike it, which matters because that word travels into how risky a future proposal is judged to be.
Evidence that argues against this (1)
- Insisting on the right word for a mouse phenotype matters less than the entry implies, because that phenotype has repeatedly failed to forecast the clinical liability in either direction. The comparator inside this pathway: menin is a tumour suppressor whose germline loss causes a multi-organ neoplasia syndrome, and blocking its KMT2A interface in 94 patients aged 1.3 to 75 produced an ordinary oncology safety profile - febrile neutropenia, differentiation syndrome, QTc prolongation - alongside 22.8% CR+CRh. The risk of striking ELL2 on the word 'dispensable' is therefore not mainly a vocabulary problem; it is that neither description of a CD19-cre mouse predicts what a partial inhibitor would do in a patient, so the strike and its objection are both arguing past the evidence that would settle it. Issa GC et al. 2025
- E4strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.paper
The instrument doing most of the striking has a demonstrated blind spot: BOD1L1 is called non-essential in all leukemia cell lines by genome-wide DepMap screening, while a 618-sgRNA tiling screen in the KMT2A::AFF1 line MV4-11 finds a strong dependency confined to the N-terminal Shg1 domain, the authors noting that CRISPR perturbation effects depend on sgRNA position.
BOD1L mediates chromatin binding and non-canonical function of H3K4 methyltransferase SETD1A · See PubMed record PMID 38989615 for the full author list (not transcribed here) · Nucleic Acids Research 2024What this evidence cannot show
Full text read (PMC11381347). One gene demonstrates the failure mode exists; it says nothing about how often it occurs, and no systematic audit of this graph's Chronos-based dismissals has been done. Both tiling-screen lines carry FLT3-ITD.
Evidence that argues against this (1)
- The blind spot is real and the instrument's designers have already documented a different one, which matters for how this argument is used. Meyers et al. showed that the dominant systematic error in genome-wide CRISPR essentiality screens is a gene-independent antiproliferative effect of Cas9 cleavage that produces FALSE POSITIVES in copy-number-amplified regions, and built CERES to remove it. So the screen's known failure mode runs in the opposite direction to the one this item highlights: it invents dependencies more readily than it hides them. Both can be true, and a project auditing its Chronos-based dismissals should expect to find inflated dependencies in amplified genes alongside missed ones in long or buffered genes. Meyers RM et al. 2017
- E5weak confidence 2 of 5: Preliminary, small, indirect, or not yet peer reviewed.paper
ASH2L siRNA knockdown in the t(4;11) KMT2A::AFF1 line RS4;11 and in the MLL-rearranged line THP-1 significantly reduces HOXC8 gene and protein expression, and reduces occupancy of RBBP5, WDR5, MLL and BPTF at regions of the HOXC8 promoter by ChIP.
ASH2L-Promoted HOXC8 Gene Expression Plays a Role in Mixed Lineage Leukemia-Rearranged Acute Leukemia · Wu YJ, Tang Y, Li ZF, Li Z, Zhao Y, Wu ZJ, Su Q · OncoTargets and Therapy 2020What this evidence cannot show
siRNA only, two cell lines, one target gene (HOXC8, not HOXA9/MEIS1), no growth, apoptosis or differentiation readout, no comparison against a KMT2A-wild-type line, and no separation of the fusion from wild-type MLL1. The paper's own cell-line labelling is inconsistent ('RS4:11' / 'RSH:11'). Cited to establish that the experiment has been attempted in a KMT2A::AFF1 background, not that the requirement is established.
Evidence that argues against this (1)
- The attribution problem in this experiment is worse than 'no separation of the fusion from wild-type MLL1', because there is a third enzyme. Chen et al. found endogenous MLL1 dispensable for MLL-rearranged leukaemia in independent conditional models, while Mll2 deletion alone significantly impaired survival of MLL-AF9-transformed cells, with MLL1 and MLL2 regulating distinct pathways. ASH2L is a WRAD subunit shared across the KMT2 family, so an ASH2L knockdown that reduces HOXC8 and lowers RBBP5, WDR5 and MLL occupancy at that promoter cannot distinguish the fusion, wild-type MLL1 or MLL2 - and the genetics says the third is the likeliest. That makes the item weaker evidence for ASH2L's relevance to KMT2A::AFF1 than its placement here implies, and it is also a reason the strike may have been right for a reason nobody stated. Chen Y et al. 2017
- refutes →ASH2L — the WRAD subunit that has DPY30's essentiality profile and none of its co-dependency signal
2026-09-19: 'nobody has tested RBBP5/ASH2L in KMT2A::AFF1' withdrawn - ASH2L siRNA in RS4;11 lowers HOXC8 (PMID 32021283)
RBBP5 occupancy at the HOXC8 promoter falls on ASH2L knockdown in a KMT2A::AFF1 line (PMID 32021283)
same BOD1L1/BOD1L alias false alarm; the paper does contain BOD1L
the mitigation that entry called for, now executable rather than described
the audit's own registry read missed the whyStopped field; identity checking certifies the weakest property a citation has
states the bite-back the audit's decisions require: a wrong strike is a silent, non-self-correcting false negative
supplies the load-bearing-reason and reopens-if recording rule the register now carries
the one strike with a material cost: a published biomarker-selected MS023 + FLT3-inhibitor combination in MLL-r ALL is declined, not refuted
- related to →ELL2 — a SEC subunit that almost no cell needs, next to a paralogue almost every cell needs
the cost here is the word 'dispensable', which would license treating ELL2 blockade as low-risk against a published humoral-failure phenotype
the CCNT1 error's real cost was propagation: the alvocidib mortality argument reaching the CDK9 programme
the instrument doing most of the striking has enumerable blind spots, which is why the false negative is systematic rather than random
Restates the 2026-09-19 correction precisely: the ASH2L/RBBP5 strike's premise is narrowed, not overturned - tested once, at one target gene, never at the level of leukaemia maintenance.
- ← related toThe graph's universal 'every KMT2A fusion loses the SET domain' is a silent, propagating error class
A companion error class: that entry covers wrong strikes, this one covers false universals, and both are invisible to every checker in the repository because every citation involved is correct.