Is the drug target always in the retained half? An audit of the project's own framing
also called retained half assumption · gain versus loss of function in fusions · framing audit
This project keeps saying the thing to drug is whatever survives the break. A seven-partner audit found that is true for the two commonest partners and false for at least one other, where the oncogenic event is what got thrown away.
The same thing, for a scientist
Audits the assumption that a KMT2A fusion's oncogenic activity resides in retained modules. Across seven partners: AFF1, MLLT1 and EPS15 are addition; MLLT10 is subtraction, proven by in-cis add-back; AFDN, ELL and the partial tandem duplication are both; MLLT3 is not determinable. Subtraction comes in three subtypes and only one is visible to the domain-fate analysis this project had used exclusively.
The assumption, stated plainly
Across this graph the argument runs: KMT2A fusions keep a consistent N-terminal cassette, that cassette includes the menin-binding motif, therefore a drug against the menin interface works regardless of partner. Everything in the therapeutics and interface beats rests on it, and for the fusion this project is built on it is correct.
The question this entry asks is narrower and more uncomfortable: is it a general property of KMT2A fusions, or a fact about a few of them that we have generalised?
Where it holds (revised after the systematic audit)
A seven-partner audit has now replaced the guesswork in this section. The verdicts:
| Partner | verdict |
|---|---|
| AFF1 | addition — the control, and it holds |
| MLLT1 | addition — YEATS is retained in 84.1% and is required 4 |
| EPS15 | addition |
| MLLT10 | subtraction, proven by add-back 3 |
| AFDN | both |
| ELL | both |
| KMT2A-PTD | both |
| MLLT3 | not determinable from current evidence |
So the framing is neither a rule nor a rarity. Roughly half the audited partners carry some subtraction component, and the two commonest lymphoblastic partners, AFF1 and MLLT1, are both pure addition. The menin strategy's core population is exactly the population where the framing is safest, which is reassuring and was not guaranteed.
The audit also found that subtraction is not one mechanism but three: a discarded module, sequestration of the endogenous partner away from its normal place, and allelic silencing. Only the first is visible to domain-fate analysis, which is the only method this project had ever used. Asked of two partners, the sequestration question came back positive for both.
Where it fails: MLLT10 inverts the logic1 deeper
KMT2A::MLLT10 keeps the octapeptide motif and leucine zipper, which recruit DOT1L, and discards the partner's PZP module. That discard is not incidental.
The PZP module of AF10 is a reader that senses unmodified histone H3K27 and couples that recognition to DOT1L-mediated H3K79 methylation. It is required for H3K79 dimethylation, for expression of DOT1L target genes, and for proliferation of DOT1L-addicted leukemic cells 1.
So in intact AF10, PZP is the part that decides where DOT1L acts. The fusion keeps the module that grabs DOT1L and throws away the module that constrains where it is taken. The oncogenic event is therefore not only what was kept; it is substantially what was removed.
That is the same shape as a classical loss of autoinhibition, and it has a direct consequence for target selection: for this partner there is no retained neo-surface to drug, because the relevant change is an absence. You cannot inhibit a missing domain.
Why this is not just a curiosity
MLLT10 is not a rare partner. It is the commonest partner in infant myeloid leukemia at roughly 27% and the fourth commonest in infant lymphoblastic disease. A framing that silently excludes the leading myeloid partner is not a framing about KMT2A fusions; it is a framing about the lymphoblastic subset.
MLLT10 is also the partner whose age-and-breakpoint behaviour runs opposite to the rest, and the one that requires at least three DNA breaks to form because of its transcriptional orientation. It has now diverged from the pack on three separate axes, which starts to look less like noise and more like a genuinely different disease.
What this changes, and what it does not
It does not weaken the menin strategy for KMT2A::AFF1. Menin binds a retained motif; that is measured, structurally resolved, and clinically validated.
It does weaken three inferences this graph has been making:
- That a menin inhibitor is partner-agnostic in effect as well as in binding. It binds the same retained motif in every fusion, but if a partner's oncogenicity runs substantially through a deletion, restoring menin eviction may not restore normal behaviour to the same degree.
- That scanning retained modules is a sufficient strategy for finding new targets. For loss-driven partners the retained modules are the wrong place to look, and nothing in this project's target-selection work has looked anywhere else.
- That the fusion families can be pooled when reasoning about mechanism. On this axis they cannot.
The honest status: the graph has one well-evidenced counterexample and has not systematically checked the others. AFDN, ELL, EPS15 and the partial tandem duplication have not been examined for discarded regulatory modules at all.
Two corrections to this entry, one of which reverses its own hedge
The decisive experiment had already been done. This entry originally said that re-supplying PZP in cis to the fusion "separates gain-by-subtraction from an incidental deletion, and as far as this graph knows nobody has done it." It has been done. Restoring functional PZP into CALM-AF10 prevents transformation in vitro and in vivo and abrogates leukemogenesis, and the authors state that disrupting PZP in the fusion directly leads to transformation while restoring it reverses transformation 3.
So MLLT10 is not a suggestive counterexample. Subtraction is proven for it, and this entry's confidence should have been higher, not lower. The paper was already cited elsewhere in this graph, which is the more embarrassing part: the answer was in the building.
The YEATS generalisation was wrong, and backwards. After writing this entry I briefed an agent on the hypothesis that MLLT1 and MLLT3 might follow the same pattern, since both retain a recruiting domain and appear to lose a YEATS reader. For MLLT1 that is simply false: the YEATS domain is retained in 84.1% of patients and is required — deleting it impairs fusion binding and lowers pro-leukemic gene expression, and point mutations blocking its histone binding reduce stem-cell frequency and lengthen latency 4. MLLT1 is addition, not subtraction.
The graph already contained this. A breakpoint entry recorded the retention, and my brief contradicted it without checking.
| assumption audited | The oncogenic activity of a KMT2A fusion resides in retained modules, so drug targets live there |
|---|---|
| consequence | For a loss-driven partner there is no retained neo-surface to inhibit, so retained-module scanning cannot find a target for it |
holds for
[ "fusion-kmt2a-aff1", "fusion-kmt2a-mllt1 (YEATS retained in 84.1% and required)", "fusion-kmt2a-mllt3 (not determinable)", "fusion-kmt2a-ell (both)" ]
fails for
[ "fusion-kmt2a-mllt10 (proven by add-back, Klein 2021)" ]
not yet checked
[ "fusion-kmt2a-afdn", "KMT2A partial tandem duplication", "EPS15", "SEPTIN family partners" ]
mechanism classes
[
{
"class": "gain by addition",
"description": "retained partner module brings a new activity to KMT2A target loci",
"example": "AFF1 elongation machinery"
},
{
"class": "gain by subtraction",
"description": "a discarded partner module normally constrains the retained one, so its loss releases that activity",
"example": "MLLT10 PZP reader lost, DOT1L recruitment retained"
}
]audit result
{
"subtraction_proven": [
"MLLT10"
],
"both_addition_and_subtraction": [
"KMT2A-PTD",
"AFDN",
"ELL"
],
"addition": [
"MLLT1",
"EPS15",
"AFF1 (control, holds)"
],
"not_determinable": [
"MLLT3"
],
"subtraction_subtypes": [
"discarded module",
"sequestration of the endogenous partner",
"allelic silencing"
],
"method_blind_spot": "only the discarded-module subtype is visible to domain-fate analysis, which is the only method this project had used"
}- Generalising the retained-half framing has already shaped target selection in this project. Every candidate examined so far is a retained module or an interface on one, so a loss-driven mechanism would have been invisible to the whole search.
- Pooling fusion partners in trial analysis could hide a real difference in menin-inhibitor benefit between addition-driven and subtraction-driven partners.
- Selecting patients for a YEATS inhibitor by partner genotype rather than by breakpoint would treat roughly one in six t(11;19) patients whose fusion does not contain the target.
- Sequestration had never been asked about in this project. Asked of two partners it returned positive for both, so how many others sequester their endogenous counterpart and were invisible to domain-fate analysis?
- Allele-specific KMT2A expression in translocation-driven blasts from existing cohorts is the cheapest test of the silencing subtype and needs no new samples.
- MLLT3 excludes YEATS in about 92% of patients as a sequence fact, but causality has never been tested. Does deleting or restoring it change transformation, as it does for MLLT1 in the opposite direction?
- If a YEATS inhibitor is deployed on t(11;19) genotype alone, roughly one patient in six has a fusion lacking the target. What breakpoint assay would stratify them, and does any trial currently do it?
- E1strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.paper
The PZP module of AF10 is a reader that recognises unmodified histone H3K27 and couples that recognition to DOT1L-mediated H3K79 methylation; PZP recognition of H3 is required for H3K79 dimethylation, for expression of DOT1L target genes, and for proliferation of DOT1L-addicted leukemic cells.
The PZP Domain of AF10 Senses Unmodified H3K27 to Regulate DOT1L-Mediated Methylation of H3K79 · Chen S, Yang Z, Wilkinson AW, Deshpande AJ, Sidoli S, Krajewski K, Strahl BD, Garcia BA, Armstrong SA, Patel DJ, Gozani O · Molecular Cell 2015What this evidence cannot show
This paper characterises PZP in intact AF10 and in DOT1L-addicted leukemia cells. It does not itself demonstrate that discarding PZP in a KMT2A::MLLT10 fusion is oncogenic; that inference is ours, and it is the step most in need of a direct test. A different paper was cited to us for this point and could not be verified, so only this source is used.
Evidence that argues against this (1)
- A simpler reading is available: the fusion may be oncogenic purely because the retained octapeptide and leucine zipper recruit DOT1L at KMT2A target loci, with the absence of PZP incidental rather than causal. Distinguishing these requires re-supplying PZP in cis to the fusion, which as far as this graph knows has not been done. Chen S et al. 2015
- E2strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.review
In infant myeloid leukemia MLLT10 is the commonest KMT2A partner at about 27%, and it is among the commonest partners in infant lymphoblastic disease.
Treatment of infant leukemias: challenge and promise · Brown P · Hematology Am Soc Hematol Educ Program 2013What this evidence cannot show
A 2013 educational review; the percentage refers to the subset of infant myeloid cases with an identified partner and comes from referral cohorts.
Evidence that argues against this (1)
- Partner frequencies are detection-limited. Cohorts characterised by targeted RT-PCR panels built around the common partners will report those partners most, cryptic insertions can be missed entirely, and the proportions shift between eras as methods change. The ordering is robust; the percentages are not. MLLT10 is also the partner with a documented orientation problem: because MLLT10 and KMT2A are transcribed in opposite directions, its fusions require complex rearrangement, which makes its cases harder to find and its frequency harder to trust. Blood Cells, Molecules and Diseases 2010
- E3strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.paper
Restoring a functional PZP module in cis to the CALM-AF10 fusion prevents its transforming activity in haematopoietic stem and progenitor cells in vitro and in vivo and abrogates leukemogenesis; disruption of PZP function in the fusion directly leads to transformation, while its inclusion downregulates Hoxa genes and reverses transformation.
The role of the PZP domain of AF10 in acute leukemia driven by AF10 translocations · Klein BJ, Deshpande A, Cox KL, et al. · Nature Communications 2021What this evidence cannot show
Performed on CALM-AF10 rather than KMT2A::MLLT10. Both fusions retain the same AF10 portion and discard PZP, so the read-across is strong but is a read-across. Mouse HSPC and xenograft models.
Evidence that argues against this (1)
- Restoring PZP in cis is an elegant sufficiency experiment in CALM-AF10, which is a DIFFERENT fusion — it does not involve KMT2A at all. Using it in a KMT2A-retained-half audit imports a result about the AF10 partner's requirements into a question about what the KMT2A N-terminus contributes. Oncofusion knowledge graph (internal analysis) 2026
- E4strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.paper
The ENL YEATS domain is retained in 84.1% of MLL-ENL patients and is crucial for leukemogenesis: deleting it impairs fusion binding and lowers pro-leukemic gene expression, and point mutations that block its histone binding decrease stem-cell frequency and lengthen latency.
The ENL YEATS epigenetic reader domain critically links MLL-ENL to leukemic stem cell frequency in t(11;19) leukemia · Hu H, Saha N, Yang Y, et al. · Leukemia 2022What this evidence cannot show
Mouse models of MLL-ENL; the 84.1% retention figure is from patient breakpoint data. Does not exclude a separate contribution from whatever ENL sequence is discarded.
Evidence that argues against this (1)
- Retention in 84.1% of patients is a statement about breakpoint distribution, and the remaining 15.9% are the informative cases: if a sixth of MLL-ENL patients lack the YEATS domain and still have leukaemia, then it is crucial in the models tested and not necessary in every patient. The audit should ask what those cases retain instead. Blood Cells, Molecules and Diseases 2010
this partner inverts the retained-half logic: losing PZP releases DOT1L recruitment from its targeting control
the framing does hold for the menin interface, which binds a retained motif
uses the retained OM-LZ versus discarded PZP composition
every candidate examined so far is a retained module, so a loss-driven mechanism was invisible to the search
supplies the lost-side inventory the retained-half audit had no counterpart for
the mirror of that audit: it warned against reading retained as required, this warns against reading lost as load-bearing
supplies the causal evidence the audit said was missing, and corrects its claim that the in-cis add-back had never been done
second candidate subtraction partner; the verdict is not-determinable, not confirmation
- ← supportsAFF1 as the subtraction control: the refutation is WITHDRAWN - the degron is lost, not retained
the control holds — the retained-half framing is correct for the project's flagship fusion
extends the audit from one counterexample to a three-sub-type taxonomy, and bounds it: the two commonest lymphoblastic partners are still addition
answers the audit's four open questions and corrects one of its factual claims