Oncofusion

start / graph / hyp-kmt2a-is-over-recovered-by-rna-binding-baits-and-not-by-transcription-factor-baits

Hypothesisspeculative confidence 1 of 5: Our own inference or analogy. Not shown to be true.Q136-interactome-tail-size-and-abundance

KMT2A's proximity promiscuity is BAIT-CLASS-SPECIFIC: it is recovered at twice its size-and-abundance prediction by RNA-binding-protein baits in a leukaemia cell and at exactly its prediction by transcription-factor baits in a kidney cell

also called the RBP-versus-TF asymmetry · is KMT2A promiscuity a compartment claim

KMT2A is found next to half of the RNA-handling proteins tested in a leukaemia cell, but only next to the expected share of transcription factors tested in a kidney cell. If that difference is real, it says KMT2A spends its time in the RNA-processing part of the nucleus - which is a claim about where the protein lives, not about which proteins it binds.

The same thing, for a scientist

**Statement.** KMT2A's high recovery as a proximity prey is not a uniform property of the protein but a property of the **bait class and cell**. Measured in this beat [E1]: in **IM-30059** (50 RNA-binding-protein BioID2 baits, K562) KMT2A is recovered by **25 of 50 baits**, **2.0x** a negative-binomial prediction from its length and abundance (predicted 12.5). In **IM-28767** (102 transcription-factor BioID baits, HEK293 Flp-In) it is recovered by **11 of 102**, **0.9x** its prediction (predicted 12.2, p = 0.42) - **exactly what size and abundance predict and nothing more**. The same covariate model, the same protein, two bait classes, and the ratio differs by more than twofold. **If the asymmetry is real it supports the reading `ana-kmt2a-proximity-promiscuity-25-of-50` named and could not separate** - *"KMT2A genuinely occupies a shared compartment with RNA processing"* - **and it is not a claim the size-explains-it result can refute, because the size term is in the model.** **If it is not real, it is confounded by cell line, by labelling reagent, or by how each study filtered before deposition, and all three confounds are present.**

Go deeper
What the numbers are, and the honest state of the 2.0x
depositionbait classcellKMT2A recoverypredictedratiop
IM-3005950 RNA-binding proteinsK56225 / 50 (50%)12.52.0x0.110 (negative binomial)
IM-28767102 transcription factorsHEK293 Flp-In11 / 102 (11%)12.20.9x0.42

The 2.0x is not statistically significant on its own and this entry does not claim it is. Its Poisson p of 9.8e-04 collapses to 0.110 once the measured Pearson dispersion of 4.95 is modelled, which is what reconciled this beat with ana-kmt2a-proximity-promiscuity-is-explained-by-size... 12. The pre-registered empirical arbiter - preys matched to KMT2A on both length and abundance - declines, with a pool of 14 against a floor of 40.

So the hypothesis rests on a contrast, not on either number alone. A 2.0x that is individually non-significant sitting beside a 0.9x that is squarely on prediction is a pattern worth testing, and the test it needs is a direct one rather than a comparison of two loosely matched studies.

The three confounds, named before anyone builds on this

1. Cell line. K562 versus HEK293 Flp-In. KMT2A's expression, its complex composition and the accessibility of its surface could all differ.

2. Bait panel composition, which is the whole point and also the whole problem. A panel of 50 RNA-binding proteins is a panel of proteins that live where RNA is processed. If KMT2A lives there too, it is recovered widely for a real reason that this design cannot distinguish from complex membership - and this was pre-registered as uninformativeness clause U4 and is not controlled.

3. Deposition filtering. Khoroshkin et al. deposited a curated bona fide pair list; Göös et al. deposited BioID and AP-MS arms with their own thresholds. Different filters produce different denominators, and the ratio of observed to predicted is sensitive to both.

And one that cuts the other way. The other five core proteins are all at or below prediction in the RBP map - MEN1 0.5x, AFF1 0.5x, AFF4 0.5x, DOT1L 0.5x, PSIP1 1.2x 1. So the RBP panel is not simply generous to this project's proteins. Whatever is happening is specific to KMT2A within that panel, which is the observation that makes the hypothesis worth keeping.

Facts and figures
statementKMT2A's recovery as a proximity prey exceeds its size-and-abundance prediction specifically against RNA-binding-protein baits in a leukaemia cell (2.0x) and matches it exactly against transcription-factor baits in a kidney cell (0.9x), which is a claim about where KMT2A resides rather than about what it binds.
origindata-pattern
prior probability0.3
predictions
[
  "In any future proximity dataset with a mixed bait panel in a single cell line, KMT2A's observed-over-predicted recovery ratio will be higher for RNA-processing baits than for chromatin or transcription-factor baits, with the covariate model held fixed.",
  "The same asymmetry will hold for AFF1 and AFF4 if it is a super-elongation-complex property rather than a KMT2A property; it will NOT hold for them if it is specific to KMT2A. Measured here, AFF1 and AFF4 are at 0.5x in the RBP panel, which already argues for the KMT2A-specific reading.",
  "If the asymmetry is a residence effect rather than a complex effect, it should survive removal of the RNA-processing baits that are known KMT2A-complex components - and none of the 50 RBP baits is a known KMT2A-complex component, which makes the prediction immediately checkable.",
  "KMT2A::AFF1, which loses everything C-terminal of roughly residue 1444, should show a DIFFERENT ratio from wild-type KMT2A in a fusion-bait proximity experiment, because the retained N-terminal third carries the CXXC and menin-binding modules and not the SET domain, PHD fingers or bromodomain."
]
analyses
[
  "ana-prey-recovery-tracks-size-in-proximity-labelling-and-not-in-ap-ms-and-the-tail-was-scored-on-the-wrong-side-of-the-pair",
  "ana-lysine-count-versus-length-as-the-proximity-recovery-covariate"
]
experiments
[
  "exp-q129-a-kmt2a-bait-proximity-experiment-in-a-kmt2a-rearranged-leukaemia-line"
]
falsifiers
[
  "SCOREABLE FALSIFIER 1. In a single proximity dataset containing both RNA-processing and chromatin or transcription-factor baits in ONE cell line, KMT2A's observed-over-predicted recovery ratio differs between the two bait classes by LESS THAN 1.5-fold, with length and abundance held in the model. That would show the asymmetry measured here is a between-study artefact of cell line, reagent or filtering rather than a compartment property. The Go et al. 192-bait human cell map (PMID 34079125, HEK293) is the obvious dataset and has never been used for this.",
  "SCOREABLE FALSIFIER 2. KMT2A's 25-of-50 recovery in IM-30059 falls BELOW 1.5x its predicted value when a nuclear-residence covariate is added to the model alongside length and abundance. The commissioning memo asked for three covariates and this beat fitted two; the third is the one most likely to absorb the effect, because a panel of nuclear RBPs will preferentially reach nuclear proteins.",
  "SCOREABLE FALSIFIER 3. AFF1 and AFF4 show the same 2.0x RBP-panel excess as KMT2A in any replication. Measured here they are at 0.5x, so a replication that finds them elevated would mean the effect is a super-elongation-complex property - which would not falsify the compartment claim but would falsify the KMT2A-specific version stated above, and the entry would have to be rewritten around the complex.",
  "SCOREABLE FALSIFIER 4. The 2.0x does not replicate at all in a second RBP-bait proximity dataset in a haematopoietic line, at a ratio below 1.3x. At a negative-binomial p of 0.110 in its only observation, non-replication is the single most likely outcome and this hypothesis should be held loosely until it is attempted."
]
How this could bite us back
  • If KMT2A's proximity neighbourhood is dominated by compartment rather than complex, then any intervention designed from a proximity map targets an address and not a machine, and the partners it nominates will be the ones that happen to live nearby.
  • The hypothesis is most dangerous if it is half-right. A KMT2A that genuinely occupies the RNA-processing compartment would generate a long list of plausible RNA-processing partners, every one of which is a candidate for an entry, and this project has already spent several cycles deciding whether that list deserves entries at all.
Still open
  1. Does a single-cell-line, mixed-bait-class proximity atlas already contain the decisive comparison? Go et al. published 192 baits in HEK293 and this project has never run its own covariate model over that deposition. If KMT2A's ratio differs by bait class WITHIN one study, the hypothesis strengthens sharply; if it does not, falsifier 1 fires.
  2. Is the RBP panel reaching KMT2A because KMT2A is in the RNA-processing compartment, or because KMT2A's transcript-associated pool is what a nuclear RBP bait sees? Those are different claims about the same observation and the graph should not let them merge.
  3. What does a nuclear-residence covariate do to the 2.0x? It is the term this beat declined to fit, for a stated reason - a localisation join with its own failure modes - and it is the term most likely to absorb the effect.
Every source behind this entry (2)
  1. E1moderate confidence 3 of 5: One strong study, or several smaller consistent ones.dataset

    In IM-30059 (50 RBP BioID2 baits, K562) KMT2A is recovered by 25 of 50 baits against a negative-binomial prediction of 12.5 from log10 length and log10 abundance (ratio 2.0, p = 0.110), while MEN1 is 3 of 50 (0.5x), AFF1 4 of 50 (0.5x), AFF4 4 of 50 (0.5x), DOT1L 4 of 50 (0.5x) and PSIP1 9 of 50 (1.2x); in IM-28767 (102 TF BioID baits, HEK293 Flp-In) KMT2A is recovered by 11 of 102 against a prediction of 12.2 (ratio 0.9, p = 0.42) and DOT1L by 2 of 102 against a prediction of 11.5.

    What this evidence cannot show

    Neither ratio is individually significant at 0.05 and the empirical matched-pool test DECLINES for KMT2A in both depositions (pools of 14 and 6 against a pre-registered floor of 40). The two depositions differ in cell line, bait class, labelling reagent and deposition filtering, all uncontrolled. A nuclear-residence covariate was requested by the commissioning memo and is not in the model, so length and abundance absorb any correlation with it.

    Evidence that argues against this (1)
    • The comparison is between two studies, not two arms of one study. The only within-study contrast available - IM-28767's BioID versus AP-MS arms - shows the size coefficient collapsing from +0.21 to +0.03, which demonstrates that this class of coefficient moves substantially with method alone. source 2026
  2. E2strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.dataset

    A length-matched decile null over the same IM-30059 deposition placed KMT2A at the 90.7th percentile of its own length decile (n = 268, median 9 baits), inside the central 90%, and concluded that its 25-of-50 promiscuity is EXPLAINED BY SIZE; the same 25 gives a negative-binomial upper-tail p of 0.110 once the measured overdispersion of 4.95 is modelled, so the two instruments agree.

    What this evidence cannot show

    The earlier analysis used a 268-member length decile; this beat's tighter length-and-abundance matching yields only 14 and declines. The agreement between the two is therefore between a wide-bin empirical test and a model-based test, not between two equivalent instruments.

    Evidence that argues against this (1)
    • This is the strongest single argument AGAINST the hypothesis. The prior analysis's conclusion - that this census cannot support individual proximity partners for KMT2A - survives, and a hypothesis that KMT2A is over-recovered must live with a p of 0.110 in the arm it depends on. source 2026
Connected to