Oncofusion

start / graph / risk-partner-import-distorts-target-selection

Riskmoderate confidence 3 of 5: One strong study, or several smaller consistent ones.Q-partner-provenance

Bite-back: a target nominated from an imported mechanism inherits a fusion the index disease does not have

also called import risk to target selection · AF9-shaped target list

If a drug target is chosen because of how a mechanism works in MLL-AF9 leukaemia, and the infant fusion does not have the protein part that mechanism runs through, the target can be perfectly real and still be the wrong one for the disease this project is about.

The same thing, for a scientist

The audit's practical consequence. Fifty-four percent of fusion-bearing load-bearing evidence in the sampled beats comes from a non-index fusion [E1], and the two DOT1L-recruitment interfaces that most of the H3K79 mechanism runs through - the AF9/ENL AHD and the AF10 OM-LZ - are **absent from KMT2A::AFF1** [E2][E3]. A target list built from that mechanism is shaped by AF9. **Three concrete failure modes.** 1. **The target exists but the mechanism does not.** An AHD-DOT1L interaction inhibitor is real chemistry against a real interaction [E4]; in t(4;11) cells the fusion occupies that surface for P-TEFb, so the compound would act by a different mechanism than the one it was optimised for, against a biomarker that does not report it. 2. **The biomarker is wrong before the compound is.** Pharmacodynamic endpoints built on H3K79me2 assume a fusion-directed pool that may not exist in the index fusion. 3. **The window is set by normal cells, not by the fusion.** If the H3K79 methylation in t(4;11) cells is endogenous, then removing it removes an activity normal lymphopoiesis uses, and the selectivity argument evaporates. **Mitigation is mechanical:** before nominating a target from a mechanism, check whether the mechanism names an interface, and whether KMT2A::AFF1 has it. AFF1's Pfam complement is three families long [E2]; the check takes a minute.

Go deeper
Severity, graded

High for targets nominated from interface-naming mechanisms in AF9/ENL or AF10 systems - the AHD-DOT1L axis and the DOT1L-AF10 coiled coil. The interface is either present in the index fusion or it is not, and it is not.

Moderate for targets nominated from murine myeloid requirement claims. Requirement transfers better than contact, but the target gene set and the differentiation endpoint do not.

Low for targets on the KMT2A side of the fusion - menin, PAF1c, the CXXC domain - because every translocation retains that moiety. This is the structural reason the menin interface remains this project's best bet even after the audit.

Unknown for cell-line lineage import, because the class could not be counted 1.

Facts and figures
kindcompensation
severityhigh for interface-naming mechanisms imported from AF9/ENL or AF10 systems; moderate for murine myeloid requirement claims; low for targets on the retained KMT2A moiety; unmeasured for cell-line lineage import
mitigationbefore nominating a target from a mechanism, ask whether the mechanism names an interface and whether KMT2A::AFF1 carries it; AFF1's full Pfam complement is PF05110, PF18875, PF18876 and the check is a one-minute UniProt lookup
affects
[
  "drug-pinometostat",
  "drug-revumenib"
]
How this could bite us back
  • Applying the interface test too strictly would reject compounds that work in the index disease by an unintended mechanism, which is a normal and acceptable way for drugs to work.
  • The mitigation is cheap enough that its main failure mode is not being applied, and the beats most likely to skip it are the ones moving fastest toward a nominated target.
Still open
  1. How many targets already nominated in this graph rest on interface-naming mechanisms from AF9/ENL or AF10 systems? The target and vulnerability folders were outside this beat's sample.
  2. Is the reverse risk real - has any target been rejected because a mechanism failed in MLL-AF9 when it would have worked in KMT2A::AFF1? Negative results are not recorded by fusion anywhere this beat could find.
Every source behind this entry (4)
  1. E1moderate confidence 3 of 5: One strong study, or several smaller consistent ones.dataset

    In a stratified random sample of 48 mechanism-bearing entries, 13 of the 24 with fusion-bearing load-bearing evidence (54%) derive from a fusion other than KMT2A::AFF1 and none from a dedicated human KMT2A::AFF1 experiment; a separate cell-line string census over 848 entries finds 80 naming human myeloid lines and 116 naming KMT2A-rearranged B-lymphoid lines, an instrument too coarse to size the lineage class.

    What this evidence cannot show

    Our own counts, 2026-09-19; only the first evidence item of each sampled entry was classified, the sample is stratified by folder rather than proportionally, and n = 48 gives a binomial 95% interval of roughly 33-74% on the 54%. The string census counts a line named in a caveat identically to a line used for a result.

    Evidence that argues against this (1)
    • Twelve of the 13 imports are already flagged in their own entries, so a target-selection process that reads caveats would not be misled; the risk applies to a process that reads summaries, which is how target lists are usually assembled. Oncofusion knowledge graph et al. 2026
  2. E2established confidence 5 of 5: Multiple independent lines of evidence; treated as settled in the field.database

    UniProtKB P51825 (AFF1, 1,210 aa) cross-references exactly three Pfam families - PF05110, PF18875, PF18876 - and not PF17793 (AHD), which is carried by both MLLT3 (P42568) and MLLT1 (Q03111).

    What this evidence cannot show

    Accessions fetched directly on 2026-09-19 with primary gene names and lengths read back rather than resolved by gene_exact query. Pfam absence is absence of a detected match, and AFF1 is annotated as overwhelmingly disordered, where short linear motifs are systematically under-detected.

    Evidence that argues against this (1)
    • AFF1 carries PF18875, the motif that binds the AHD, so a target on the AHD is not irrelevant to the index fusion - it is relevant with the roles reversed, which changes the predicted direction of effect rather than removing it. Leach BI et al. 2013
  3. E3established confidence 5 of 5: Multiple independent lines of evidence; treated as settled in the field.database

    UniProtKB P55197 (MLLT10) annotates residues 703-784 as 'Transactivation domain; required for DOT1L-binding' and 750-778 as a leucine zipper; AFF1 carries no equivalent region.

    What this evidence cannot show

    Fetched directly on 2026-09-19. UniProt regions are curatorial calls from primary literature rather than crystallographic footprints.

    Evidence that argues against this (1)
    • Endogenous AF10 is required for transformation by three fusions that do not contain it, so an AF10-directed target could be effective in the index fusion through the endogenous protein - efficacy without fusion specificity. Deshpande AJ et al. 2014
  4. E4strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.paper

    A small-molecule inhibitor of the AF9/ENL-DOT1L/AF4/AFF4 interactions (IC50 0.9-3.5 uM) reduces SEC assembly and H3K79 methylation, suppresses onco-MLL, DOT1L, HOXA9 and MYC signatures, and has antitumour activity in a mouse MLL-rearranged leukaemia model, and the authors present the interaction as a validated anticancer target.

    Small-molecule inhibitor of AF9/ENL-DOT1L/AF4/AFF4 interactions suppresses malignant gene expression and tumor growth. · Wu F, Nie S, Yao Y, Huo T, Li X, Wu X, Zhao J, Lin YL, Zhang Y, Mo Q, Song Y · Theranostics 2021
    PMID 34373735DOI
    What this evidence cannot show

    Abstract fetched from PubMed and read on 2026-09-19; full text not read. Micromolar potency with a broad transcriptional footprint and no separation-of-function control described; the in vivo model's fusion is not specified in the abstract. Cited here as the worked example of a real target whose mechanism in the index fusion would differ, not as a criticism of the chemistry. **Half resolved.** The in vitro arm is **MOLM-13 at 5 microM** (GSE151660); **the in vivo arm is not deposited and stays unresolved.** Do not rewrite this item as 'the model is MOLM-13'. Resolved 2026-09-19 by the Q18b unstated-systems sweep from the study's deposition, not from its text; bounded as a statement about the deposited data rather than about the whole study.

    Evidence that argues against this (2)
    • If the compound works in t(4;11) cells by detaching the fusion's own P-TEFb route, that is still efficacy, and a target-selection process that rejected it on mechanistic grounds would have discarded an active compound for the wrong reason. Srinivasan RS et al. 2004
    • Resolving one arm and leaving the other open is the case most likely to be mis-summarised later, because the resolved half reads as the answer. MOLM-13 carries KMT2A::MLLT3, so even the resolved arm is an import relative to the index fusion - which is precisely the distortion this entry exists to track, now demonstrated on its own evidence. NCBI Gene Expression Omnibus 2026
Connected to