Oncofusion

start / graph / ana-kmt2a-proximity-promiscuity-is-explained-by-size-97th-percentile-becomes-91st-among-proteins-its-own-length

Analysisestablished confidence 5 of 5: Multiple independent lines of evidence; treated as settled in the field.

NOT YET REPLICATED (corrected cycle 97). Within the K562 map KMT2A's 25-of-50 promiscuity is explained by size (97.1st percentile falling to 90.7th among proteins its own length) and menin is under-reached at the 23.5th, surviving an abundance control. The replication attempt was UNINFORMATIVE, not negative: eleven of twelve other proximity panels do not sample the nucleus, and the one that does shows the same shape

also called KMT2A proximity promiscuity is size · 97th percentile becomes 91st when length-matched · the 50-bait census cannot support individual partners

A survey using fifty different bait proteins found the gene this project studies turning up near twenty-five of them, far more than almost any other protein, which could mean it genuinely sits in a shared neighbourhood with RNA-processing machinery. It is one of the largest proteins in the human body, though, and bigger proteins are easier to bump into: comparing it only against proteins of similar size, it is no longer exceptional, and similar-sized proteins in the same survey average nearly twice as many hits as ordinary ones. So the striking number is mostly a fact about its size, and the individual proteins found near it should not each be written up as partners.

The same thing, for a scientist

**SCOPE, corrected cycle 97 — read first.** The replication test across twelve other proximity publications is **UNINFORMATIVE, not negative**: eleven of the twelve panels do not sample the nucleus (centrosome–cilium, RTK, kinase, CFTR, Kir2.1, mitochondrial ClpP, Hippo, phosphatases, a method paper, a colorectal screen), so KMT2A's 0th percentile in three of them measures panel composition. **In the one transcription-factor panel KMT2A is at the 71.7th percentile and menin is absent — the same shape as K562, at n = 1.** So what follows is **not yet replicated** rather than **contradicted**. Cycle 96's harsher narrowing was itself overstated and is withdrawn. --- **Statement.** `ana-kmt2a-proximity-promiscuity-25-of-50` (cycle 51) reports wild-type KMT2A recovered by **25 of 50** BioID2 baits in the K562 map — the **97.6th percentile** of the baits-per-prey distribution — against **MEN1's 3**. It names two readings and deliberately separates neither: *"KMT2A is an abundant, very large chromatin protein that any nuclear bait's labelling radius will reach, or KMT2A genuinely occupies a shared compartment with RNA processing while menin does not."* **Cycle 94 made that an urgent question rather than an interesting one.** 12 of the 19 interactome partners that failed the draftability test are RNA-processing factors, and 7 come from this census. Whether they — or any of KMT2A's proximity partners — deserve individual entries turns entirely on which reading is right. **Reproduced first, then departed from.** IntAct PSICQUIC `pubid:39251607`, paginated to exhaustion: **19,403 rows, exactly the count the source entry recorded**; bait panel inferred at **50**; **KMT2A 25 baits, MEN1 3** — all four figures match. **The 97.6th percentile is the wrong comparison, and this is the whole finding.** It asks whether KMT2A is labelled more than a *typical* protein. **A 3,969-residue protein is not typical.** The question that separates the readings is whether KMT2A is labelled more than proteins **of its own size**: | | | |---|---:| | KMT2A length | **3,969 aa** | | its length decile (9th) | 1,462–5,890 aa, **268 members** | | **median baits, that decile** | **9** | | median baits, all 2,662 preys | 5 | | **KMT2A percentile within its decile** | **90.7** | | KMT2A percentile against all preys | 97.1 | **`P1` — EXPLAINED BY SIZE.** Inside the central 90% of its own length decile. **Large proteins in this map are hit by nearly twice as many baits as ordinary ones**, and once that is held constant KMT2A stops being exceptional. **Consequence for the work order:** this census **cannot support individual proximity partners for KMT2A**, which confirms cycle 94's revision — the RNA-processing block is adjudicated as a class or not at all — and extends it to every partner this map contributes. **MEN1, computed before shipping because this entry's own counter-evidence called its omission odd in a menin-focused project — and it is the more interesting number.** Menin is **610 aa**, in the 5th length decile (585–707 aa, 268 members) where the **median prey is reached by 6 baits**. Menin is reached by **3**, the **23.5th percentile of its size-matched peers**. | | baits | length | decile median | **percentile in its decile** | |---|---:|---:|---:|---:| | KMT2A | 25 | 3,969 aa | 9 | **90.7** | | MEN1 | 3 | 610 aa | 6 | **23.5** | **The two move in opposite directions once size is held constant, and menin's is a positive result rather than a null.** Menin is labelled *less* than proteins of its own size — **not merely small, but selectively localised.** For a project whose central hypothesis is an interface that matters: a protein a 50-bait panel under-reaches is one whose documented interactions are less likely to be proximity artefacts. **The same correction that deflates KMT2A's promiscuity strengthens the case that menin's partners are real.** **CYCLE 95 — the abundance attack this entry named as its own cheapest refutation was run, and menin survives it.** Matched on K562 nTPM from HPA: menin **25.0th** percentile by abundance and **35.4th** matched on length and abundance jointly, both at or below the pre-registered 40th. **And menin and KMT2A sit at essentially the same K562 abundance — 30.2 against 28.6 nTPM — while being reached by 3 and 25 baits**, so abundance is held constant by the biology itself. **Weakening recorded:** jointly matched, menin's peer median is 4 baits against its 3, so it is barely below the middle of its closest peers; the head-to-head with KMT2A is the strongest form of the claim, not the percentile.

Go deeper
What this does not settle, and the margin is not comfortable

90.7 is not far from the 95 that would have flipped the verdict. The pre-registration set the bar before the pull and the result is inside it, but a reader should treat this as "size accounts for most of it" rather than "size accounts for all of it". A different length binning — quintiles, or a continuous regression on log length — could move it either way, and I ran one binning.

Length controls one term of three. The confound the source entry names is abundance × size × nuclear residence, and a length-matched null holds only size constant. So this is evidence against the size reading only. KMT2A could still be labelled more than expected for its abundance, and nothing here touches that. Declared as U2 before running, and it remains the main open route to rescuing the shared-compartment reading.

Baits-per-prey is not a biological quantity. It counts how many of 50 chosen baits reached a prey, so it inherits whatever the panel is biased toward. A chromatin-heavy panel inflates chromatin preys regardless of size, and the length-matched null does not remove that — the peers are size-matched, not function-matched. Declared as U4 and not controlled.

A small discrepancy worth recording: my reproduction puts KMT2A at the 97.1st percentile against all preys where the source entry recorded 97.6. Both rest on 19,403 identical rows, so the difference is in how preys are counted into the denominator — I required at least one panel bait and a UniProt length join, giving 2,662 preys. The direction of this entry's conclusion does not depend on which is right, but the two numbers should not be quoted interchangeably.

CYCLE 95: the abundance attack was run and menin survives it - and the two proteins turn out to sit at the same abundance, which is the cleanest control available

This entry named its own cheapest attack: "if menin's 3 baits is explained by low abundance rather than tight localisation, the positive reading evaporates." Run against K562 nTPM from Human Protein Atlas rna_celline — the same cell line as the map, streamed from the 206 MB archive and the archive deleted. 2,646 of 2,682 preys join to both a length and a K562 abundance.

baitslengthK562 nTPMby lengthby abundanceby length AND abundance
MEN13610 aa30.223.525.035.4
KMT2A253,969 aa28.690.695.689.7

A1 — the under-labelling survives, at or below the pre-registered 40th percentile on both matchings.

And the cleanest fact here was not something I designed for: menin and KMT2A sit at essentially the same abundance in K562 — 30.2 against 28.6 nTPM, the same abundance decile. Two proteins present at the same level in the same cells, one reached by 3 of 50 baits and the other by 25. Abundance cannot explain any part of that difference, because it is held constant by the biology rather than by my stratification.

KMT2A is the control on the control (V4) and it behaves: 90.6 by length, 95.6 by abundance, 89.7 jointly. The matching is not doing anything strange to a protein we already understand.

But the joint matching weakens menin's result and that must be said plainly. Matched on length and abundance together, menin's peer group (n = 113) has a median of 4 baits, and menin has 3. The percentile rises from 23.5 to 35.4. Against its most closely matched peers menin is barely below the middle — it clears the pre-registered bar and it is a thinner result than the length-only comparison suggested. The honest summary is that the abundance attack failed to kill the claim, and the head-to-head against KMT2A is the strongest form of it, not the percentile.

Facts and figures
questionis KMT2A's 25-of-50 proximity promiscuity more than its size predicts?
verdictP1 - EXPLAINED BY SIZE. The census cannot support individual proximity partners for KMT2A.
kmt2a length AA3969
kmt2a percentile within length decile90.7
kmt2a percentile against all preys97.1
source entry reported percentile97.6
why the original percentile was the wrong comparisonit asks whether KMT2A is labelled more than a TYPICAL protein, and a 3,969-residue protein is not typical. Large proteins in this map are reached by a median of 9 baits against 5 overall.
consequence for the work orderconfirms and extends cycle 94's revision - the RNA-processing block, and every other partner this census contributes, is adjudicated as a class or not at all, not promoted partner by partner
absence classmeasured-not-analysed
absence class noteThe 19,403-row deposition has been public and was queried in cycle 51; the source entry records data.script as 'none (database query)', so the distribution behind its percentile was computed and discarded. This re-derives and persists it.
reproduction of the source entry
{
  "rows": 19403,
  "rows_expected": 19403,
  "matches": true,
  "bait_panel": 50,
  "kmt2a_baits": 25,
  "men1_baits": 3
}
length decile
{
  "index": 9,
  "range_aa": [
    1462,
    5890
  ],
  "n_members": 268,
  "median_baits": 9
}
all preys
{
  "n": 2662,
  "median_baits": 5
}
what it does not settle
{
  "margin": "90.7 against a pre-registered bar of 95, on one length binning",
  "U2_abundance": "length controls one term of size x abundance x nuclear residence; the abundance reading is untouched and is the main route to rescuing the shared-compartment interpretation",
  "U4_panel_bias": "baits-per-prey inherits the bait panel's composition; size-matched peers are not function-matched"
}
men1 length matched
{
  "gene": "MEN1",
  "baits": 3,
  "length_aa": 610,
  "decile": 5,
  "decile_range_aa": [
    585,
    707
  ],
  "decile_n": 268,
  "decile_median_baits": 6,
  "percentile_within_decile": 23.5,
  "reading": "menin is labelled LESS than its size predicts - not merely small but selectively localised. A protein a 50-bait panel under-reaches is one whose documented interactions are less likely to be proximity artefacts.",
  "contrast": "the same length correction moves KMT2A to 90.7 and MEN1 to 23.5 - opposite directions relative to their size-matched peers.",
  "output": "analyses/is-kmt2a-proximity-promiscuity-just-size/men1_length_matched.json"
}
abundance matched cycle 95
{
  "verdict": "A1 - the under-labelling survives",
  "abundance_source": "Human Protein Atlas rna_celline, K-562 nTPM; 206 MB archive streamed and deleted",
  "preys_joined_to_length_and_abundance": 2646,
  "MEN1": {
    "baits": 3,
    "length_aa": 610,
    "k562_ntpm": 30.2,
    "pct_by_length": 23.5,
    "pct_by_abundance": 25,
    "pct_by_length_and_abundance": 35.4,
    "joint_stratum_n": 113,
    "joint_stratum_median_baits": 4
  },
  "KMT2A": {
    "baits": 25,
    "length_aa": 3969,
    "k562_ntpm": 28.6,
    "pct_by_length": 90.6,
    "pct_by_abundance": 95.6,
    "pct_by_length_and_abundance": 89.7,
    "joint_stratum_n": 146,
    "joint_stratum_median_baits": 8
  },
  "the_natural_control_i_did_not_design_for": "menin and KMT2A sit at essentially the SAME K562 abundance (30.2 vs 28.6 nTPM, same decile) while being reached by 3 and 25 baits. Abundance is held constant by the biology, not by my stratification, so it cannot explain any part of that difference.",
  "V4_control_on_the_control": "KMT2A is stable across all three matchings (90.6 / 95.6 / 89.7), so the matching is not distorting a protein we already understand",
  "the_honest_weakening": "jointly matched, menin's peer median is 4 baits and menin has 3, moving it from the 23.5th to the 35.4th percentile. Against its most closely matched peers menin is barely below the middle. The claim clears its pre-registered bar and is thinner than the length-only comparison implied; the head-to-head against KMT2A is the strongest form of it.",
  "V3_rna_is_not_protein": "nTPM is a transcript proxy correlating with protein at roughly r=0.4-0.6, so a negative here would have been decisive and this positive is suggestive. The asymmetry was declared before running."
}
NARROWED cycle 96
{
  "what_changed": "the contrast does not generalise beyond the K562 map",
  "KMT2A_across_12_publications": {
    "scoreable_in": 5,
    "at_0th_percentile": 3,
    "above_60th": 2
  },
  "MEN1_across_12_publications": {
    "scoreable_in": 1,
    "absent_from": 11
  },
  "W4_failed": "the control declared in the replication test - if KMT2A is not consistently high the K562 contrast was the panel - failed",
  "scope_of_everything_else_in_this_entry": "the K562 50-bait BioID2 map only",
  "SUPERSEDED_cycle_97": "the cycle-96 narrowing said the contrast does not generalise and W4 failed. Both overstated: 11 of 12 panels do not sample the nucleus, W4 was applied without conditioning on panel composition, and the single transcription-factor panel shows KMT2A 71.7 with menin absent - the same shape. Correct status: NOT YET REPLICATED."
}
Still open
  1. ANSWERED IN THIS ENTRY: menin sits at the 23.5th percentile of its length decile, 3 baits against a size-matched median of 6. The question that replaces it: does menin's under-labelling reproduce in other proximity datasets? If it does, 'selectively localised' is a property of menin rather than of one bait panel.
  2. ANSWERED IN THIS ENTRY (cycle 95): K562 nTPM from HPA gives menin the 25.0th percentile abundance-matched and 35.4th jointly, so the under-labelling survives - and menin and KMT2A sit at the same abundance while differing 3 versus 25 baits. What remains: nTPM is a transcript proxy. A K562 PROTEIN-level abundance dataset would close the last gap, and its absence is why this is suggestive rather than decisive.
  3. Does the verdict survive a different binning? 90.7 against a bar of 95 is close, and I ran deciles once. A continuous regression of baits-per-prey on log length, with KMT2A's residual reported, would not depend on bin edges I chose.
  4. Is length the right scale at all? BioID2 labels within a radius, so accessible surface and residence time are what should scale, not residue count - and KMT2A is substantially disordered. If length overstates expected labelling for disordered proteins, its size-matched peers are the wrong peers.
Every source behind this entry (1)
  1. E1established confidence 5 of 5: Multiple independent lines of evidence; treated as settled in the field.analysis

    IntAct PSICQUIC pubid:39251607 returns 19,403 rows, matching the count recorded by ana-kmt2a-proximity-promiscuity-25-of-50. From them a 50-member bait panel is recovered, with KMT2A reached by 25 baits and MEN1 by 3. Of 2,682 preys with at least one panel bait, 2,662 join to a UniProt reviewed human length. KMT2A is 3,969 aa, placing it in the 9th length decile (1,462-5,890 aa, 268 members), where the median prey is reached by 9 baits against a median of 5 across all preys. KMT2A sits at the 90.7th percentile within that decile and the 97.1st against all preys.

    What this evidence cannot show

    THE MARGIN IS NOT COMFORTABLE: 90.7 against a pre-registered bar of 95, on one length binning that I chose. A continuous regression on log length, or quintiles instead of deciles, could move it. LENGTH CONTROLS ONE TERM OF THREE - the confound named by the source entry is abundance x size x nuclear residence, so this is evidence against the SIZE reading only and leaves the abundance reading untouched. Baits-per-prey inherits the bait panel's composition, and size-matched peers are not function-matched, so a chromatin-heavy panel would inflate chromatin preys at any size. The bait panel was INFERRED from rows recurring at least 50 times rather than read from a declared list, and although it recovers exactly 50 members, a declared panel would be the better instrument. My all-prey percentile is 97.1 where the source entry recorded 97.6 on the same 19,403 rows, so the denominators differ and the two should not be quoted interchangeably.

    Evidence that argues against this (3)
    • A 90.7th percentile is still the top tenth of size-matched peers, and reporting it as 'explained by size' spends the benefit of the doubt in one direction. An equally defensible summary is that KMT2A is labelled somewhat more than its size predicts and the effect is too small to build entries on - which reaches the same practical conclusion by a more honest route. A reader who cares about the mechanism rather than the work order should note the entry chose the framing that supports the decision it was commissioned to inform.
    • The whole test assumes baits-per-prey should scale with length, and that assumption is unexamined. BioID2 labels lysines within a radius, so what should scale is accessible surface and residence time, not residue count - and for a protein that is largely intrinsically disordered, as KMT2A substantially is, length may overstate the expected labelling far more than for a compact 3,969-residue protein. If so, the length-matched peers are the wrong peers and KMT2A's 90.7 is an underestimate of its excess.
    • I noted that leaving MEN1 uncomputed was an odd omission in a menin-focused project. COMPUTED BEFORE SHIPPING: menin is 610 aa, 5th length decile, median 6 baits, and its 3 puts it at the 23.5th percentile of size-matched peers - labelled LESS than its size predicts, a positive statement rather than a null. What this cuts against is the entry's own framing in one respect: if a length-matched null is the right instrument for menin, the same instrument leaves KMT2A in the top tenth of its peers, and 'explained by size' is doing more work for KMT2A than the numbers strictly support.
Connected to