Oncofusion

start / graph / ana-the-corpus-predicted-the-partner-heuristic-would-fail-and-q142-measured-it-failing

Analysisstrong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.

This corpus argued from the literature that intersecting fusion peaks with partner peaks cannot attribute occupancy, and eighty cycles later a beat measured that partner antibody sitting at the negative control in primary patient samples - two independent derivations from the same paper, with no edge between them and one wrong premise

also called the attribution audit predicted q142 · prediction confirmed by measurement · heuristic 2 fails, argued then measured · AFF1 at 1.03-1.45 is the predicted null

Years of this project's notes argued, from a 2010 paper, that you cannot prove where a fusion protein sits by checking where its partner sits — because the partner is normally at those same places anyway. A recent analysis measured exactly that in real patients' cells and found the partner antibody performing no better than a control that was meant to mean nothing. The prediction and the measurement were made independently and neither knew about the other.

The same thing, for a scientist

**Statement.** `ana-attribution-audit-chromatin-occupancy-kmt2a-aff1` rules on the two heuristics the field uses to attribute chromatin occupancy to a KMT2A fusion: | heuristic | verdict | |---|---| | MLL-N **minus** MLL-C subtraction | **survives qualitatively** | | intersect MLL-N peaks **with partner peaks** | **FAILS** | **Its reason is a prediction.** Yokoyama et al. (**PMID 20153263**, *Cancer Cell* 17:198–212, 2010 — verified) show AEP is recruited to MLL-target chromatin for **physiologic** transcription, so **wild-type AF4 and wild-type MLL co-occupy MLL target genes by design**. The intersection is therefore *expected under the null*, and an MLL-N-plus-AF4-C peak is consistent with the fusion **and** with ordinary wild-type co-occupancy at the same promoter. **Cycle 124's q142 beat measured that null.** In primary infant KMT2A::AFF1 patient samples, enrichment at N/C-defined fusion promoters over wild-type promoters: | readout | ONCO/WT | |---|---:| | ENL | **1.90–2.78** | | DOT1L | 1.22–1.65 | | **AFF1** (mouse mAb) | **1.03–1.45** | | **AFF1** (rabbit pAb) | **0.56–0.76** | | H3K4me3 — *negative control* | **0.97–1.06** | **The partner antibody sits on the negative control.** That is the argued prediction, measured, in primary patient material rather than a line. 🔴 **Neither derivation knew about the other.** q142's hypothesis entry **cites the same Yokoyama paper** and **does not cite the attribution audit**; there are **zero edges** between them. This entry supplies the link. 🟠 **And it corrects q142's framing.** Its hypothesis is titled *"…and this corpus has been treating it as such"*. **The corpus's own attribution audit explicitly did not** — it ruled the partner-intersection heuristic invalid and named the paper that invalidates it. **q142's measurement stands; its claim about the corpus does not**, and it is the kind of claim an agent cannot check without a search of what the corpus already decided.

Go deeper
Why a confirmed prediction is worth more here than either half

The attribution audit's verdict was reasoning from one 2010 paper. It could have been wrong in the ordinary way arguments are wrong — the physiologic co-occupancy might have been quantitatively negligible in leukaemic cells, leaving the intersection usable in practice.

q142's measurement could equally have been an artefact of one deposit — two patients and a cell line, one protocol, no spike-in, and an AFF1 result that differs by antibody (1.03–1.45 mouse versus 0.56–0.76 rabbit) more than either differs from 1.

🔵 Together they are much stronger than apart, because they fail in different ways: an argument from mechanism and a measurement in primary material, agreeing on a number the argument predicted to be ~1. This is the only instance I can find in this corpus of a reasoning-derived methodological verdict later confirmed by an independent measurement inside the project.

🔴 What it does NOT establish (U2): that published AFF1 ChIP in general is uninformative. q142 is one deposit, and the audit's own survivor — MLL-N minus MLL-C — remains the best reagent-only discriminator the field has.

Can the corpus be re-graded on this? Mostly not, and that is measurable

If readout identity decides whether an occupancy number is a result, then re-grading past claims requires knowing which readout each used. It is usually not recorded in the claim.

Of the 50 evidence blocks that assert index-fusion occupancy in a single claim string:

n%
name an assay918.0
name a targeted reagent612.0
name both36.0
name neither3876.0

🟠 This number is weak and is reported as weak. n = 50, and the selector is over-inclusive — it caught claims about breakpoint junctions recovered from neonatal blood spots, which are not occupancy claims at all. It bounds nothing. What it does show is that the provenance needed to apply q142's lesson is not in the claim text, so any re-grading is a reading job rather than a query.

🔴 My first attempt at this census was broken and the failure is the more useful half. I classified entries by which readout they mention, and AFF1 matched 1,253 of 1,253 — 100% — because the index fusion is named KMT2A::AFF1. A keyword cannot separate the partner-as-antibody from the partner-as-half-of-the-fusion's-name, which is the same measuring-an-index-instead-of-the-thing failure this project has now logged six times.

Facts and figures
questiondid this corpus predict the partner-attribution failure that q142 measured?
answeryes, from PMID 20153263, and neither derivation cited the other
correction to q142s framingits hypothesis is titled '...and this corpus has been treating it as such'; the corpus's attribution audit explicitly ruled the partner-intersection heuristic INVALID and named the paper that invalidates it, so the measurement stands but the claim about the corpus does not
my first census was brokenclassifying entries by readout name returned AFF1 in 1,253 of 1,253, because the index fusion is NAMED KMT2A::AFF1; a keyword cannot separate the partner-as-antibody from the partner-as-half-of-the-fusion-name
partner attribution is not categoricalENL is also a partner protein and discriminates at 1.90-2.78; the audit's argument applies to physiologically co-recruited AEP components, so the effect is partner-specific rather than general
absence classmeasured-not-curated
bite backThis entry presents a qualitative expectation surviving a measurement as a confirmed prediction. The audit never stated what enrichment its null implies, so the prediction could not have failed quantitatively, and a reader should weight it as 'not contradicted' rather than 'confirmed'.
attribution audit verdicts
{
  "MLL-N minus MLL-C": "survives qualitatively",
  "intersect MLL-N with partner peaks": "fails - refuted by PMID 20153263"
}
q142 measurement onco over wt
{
  "ENL": "1.90-2.78",
  "DOT1L": "1.22-1.65",
  "AFF1_mouse_mAb": "1.03-1.45",
  "AFF1_rabbit_pAb": "0.56-0.76",
  "H3K4me3_negative_control": "0.97-1.06"
}
the link was missing
{
  "q142_cites_the_audit": false,
  "edges_between_them": 0,
  "q142_cites_PMID_20153263": true
}
readout provenance in claim text
{
  "index_fusion_occupancy_claims": 50,
  "names_assay": 9,
  "names_targeted_reagent": 6,
  "names_both": 3,
  "names_neither": 38,
  "confidence": "weak - n=50, over-inclusive selector, citations not followed"
}
Still open
  1. Follow the citations on the 38 occupancy claims that name no readout. Each carries a source; resolving a sample would settle whether provenance is genuinely lost or one hop away, and the 6% figure invites the pessimistic reading without earning it.
  2. Amend the q142 hypothesis title, which generalises from AFF1 to partner proteins while its own data show ENL discriminating at 1.90-2.78. The effect is partner-specific and the entry currently overreaches.
  3. Search the corpus before writing a dispatch brief, not after. Both halves of this finding exist because an agent rederived a verdict the corpus had already reached from the same paper; the deposit and citation indexes built in cycles 120-123 were designed for exactly this and were not used.
  4. Test the surviving heuristic the same way. MLL-N minus MLL-C is the audit's qualitative survivor and q142's N/C ratio is that heuristic in quantitative form, so the deposit can say how well it actually separates - which nobody has asked.
Every source behind this entry (2)
  1. E1strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.paper

    The committed entry ana-attribution-audit-chromatin-occupancy-kmt2a-aff1 records two field heuristics for attributing chromatin occupancy to a KMT2A fusion and rules on each: MLL-N minus MLL-C subtraction survives qualitatively, while intersecting MLL-N peaks with partner peaks fails, refuted by PMID 20153263 on the grounds that AEP is physiologically recruited to MLL targets so wild-type AF4 and wild-type MLL co-occupy MLL target genes by design and the intersection is expected under the null. PMID 20153263 resolves to Yokoyama A et al., senior author Cleary ML, A higher-order complex containing AF4 and ENL family proteins with P-TEFb facilitates oncogenic and physiologic MLL-dependent transcription, Cancer Cell 17:198-212, 2010, DOI 10.1016/j.ccr.2009.12.040. The cycle-124 q142 beat independently measured, in primary infant KMT2A::AFF1 patient samples, enrichment at N/C-defined fusion promoters over wild-type promoters of 1.03 to 1.45 for AFF1 with a mouse monoclonal antibody and 0.56 to 0.76 with a rabbit polyclonal, against 1.90 to 2.78 for ENL and 0.97 to 1.06 for an H3K4me3 negative control. The q142 hypothesis entry cites PMID 20153263 and does not cite the attribution audit, and no edge in this graph connects them.

    PMID 20153263DOI
    What this evidence cannot show

    THE TWO DERIVATIONS ARE NOT FULLY INDEPENDENT: both rest on the same 2010 paper, which q142 cites directly. What is independent is the EVIDENCE TYPE - an argument from mechanism against a measurement in primary material - not the underlying source. q142's MEASUREMENT IS ONE DEPOSIT: two patients, one of whom supplied two samples across a lineage switch, plus a cell line, one protocol, no spike-in, and its AFF1 result differs more between two antibodies (1.03-1.45 versus 0.56-0.76) than either differs from 1, so the point estimate is reagent-dependent even though the conclusion is not. NOTHING HERE SHOWS THAT PUBLISHED AFF1 ChIP IS UNINFORMATIVE IN GENERAL, and the audit's surviving heuristic - MLL-N minus MLL-C - remains the best reagent-only discriminator available. I did not re-derive q142's numbers from the deposit; they are quoted from its committed entry, which I have read but not recomputed.

    Evidence that argues against this (3)
    • Calling this a confirmed prediction flatters both entries. The attribution audit did not predict a NUMBER - it said the intersection is 'expected under the null' without stating what enrichment the null implies - so any measured value near 1 can be read as agreement after the fact. A prediction that cannot fail quantitatively is weaker evidence than this entry's framing suggests, and the honest version is that a qualitative expectation was not contradicted.
    • The more deflating reading of 'neither knew about the other' is that this project does not read its own corpus before dispatching, which it has now demonstrated across several cycles - five of the last eight briefs carried a false premise. An agent independently rederiving a settled verdict is duplicated effort dressed up as convergence, and the same tokens could have extended the audit rather than repeating it.
    • ENL at 1.90-2.78 is itself a partner protein, and it discriminates. That undercuts any general statement that partner occupancy cannot attribute: the audit's argument applies to AEP components that are physiologically co-recruited, and the data show the effect is partner-specific rather than categorical. An entry that generalises from AFF1 to 'partner proteins' would be overreaching, and the q142 hypothesis title does exactly that.
  2. E2weak confidence 2 of 5: Preliminary, small, indirect, or not yet peer reviewed.analysis

    Walking all 12,338 evidence blocks in the committed corpus for claim strings that name the index fusion and contain occupancy language yields 50 blocks. Of these, 9 name an assay such as ChIP-seq, CUT&RUN or CUT&Tag, 6 name a targeted reagent in a context adjacent to antibody or immunoprecipitation language, 3 name both and 38 name neither. An earlier version of this census classified whole entries by which readout name appeared anywhere in them and returned AFF1 in 1,253 of 1,253 entries, because the index fusion is named KMT2A::AFF1.

    What this evidence cannot show

    n IS 50 AND THE SELECTOR IS OVER-INCLUSIVE - it matched claims about breakpoint junctions recovered from neonatal blood spots, which are not occupancy claims - so these percentages bound nothing and the confidence on this item is weak deliberately. NAMING A READOUT IS NOT USING ONE and the converse is likelier still: a claim that omits its reagent may rest on the best available one, and several of the 38 do. THIS COUNTS CLAIM STRINGS ONLY - the citation attached to a block frequently identifies the assay even when the claim text does not, and that was not followed through, so the true recoverability of readout provenance is HIGHER than 6% by an unmeasured amount. This is a statement about what is queryable, not about what is knowable.

    Evidence that argues against this (1)
    • Reporting 6% without following the citations is the weakest thing in this entry. Each of those 50 blocks carries a citation, and resolving even a sample of them would establish whether the provenance is genuinely lost or merely one hop away - a question this analysis raises and then declines to answer, while presenting a number that invites the pessimistic reading.
Connected to