Oncofusion

start / graph / hyp-q160-the-bait-prey-grade-discards-the-bait-free-biophysics-that-is-the-best-evidence-in-the-tail

Hypothesismoderate confidence 3 of 5: One strong study, or several smaller consistent ones.q160-the-prey-only-interactome

The bait/prey grade this project uses to triage the interactome systematically discards its strongest evidence, because the strongest methods have no bait

also called bait-free evidence hypothesis · prey-only grading artefact

We sort protein partners by whether our protein was the hook in the experiment, and treat the rest as low quality. But the best experiments - measuring two purified proteins binding directly - have no hook at all, so they land in the reject pile.

The same thing, for a scientist

**Statement.** The grade `bait-supported` versus `prey-only` is a proxy for *did somebody set out to study our protein*, and it is computed from IntAct's `experimentalRole` field. **Symmetric methods - isothermal titration calorimetry, NMR titration, gel filtration, crosslink mass spectrometry, BRET, FRET-type assays - assign no bait role to either partner.** A filter asking *was our protein the bait* scores that empty column as not-bait, and the partner is graded prey-only and treated as low-grade. **Measured: 24 of the 212 human prey-only partners (11.3%) have NO literal prey role in any record** [E1] - `AKAP8` among them, graded prey-only on a gel-filtration co-elution from a dedicated Roeder-lab MLL-complex paper. **The damage is wider than those 24, because a partner can carry bait-free records alongside prey ones and the grade sees only the prey**: `H3C1`, the top-ranked partner in the uncovered tail, has **11 bait-free records** (six ITC, two NMR, one x-ray diffraction, all typed `direct interaction`) and **two prey records** (pull down), so it is not in the 24 and its biophysics is invisible to the grade all the same [E2]. **A second, distinct mechanism does the same damage without an empty role column:** a single publication can produce bait records for one partner and prey records for another, from the same figure set, and the grade splits it - PMID 25751424 gives SIRT1 four bait co-IP rows and CLOCK one prey two-hybrid row [E3]. **Prior probability 0.75** that the grade is materially miscalibrated in this direction; the mechanism is verified, and what is uncertain is its size.

Go deeper
Why this is not the same as rule 92

Rule 92 found the opposite error in a sibling instrument: a screen-like classifier flagged focused complex purifications as screens, misclassifying five of nine at the top of the ranking. Its lesson was hand-read the top.

This is a different failure and hand-reading the top is exactly how it was found. The bait/prey grade is not misclassifying focused studies as screens; it is discarding entire methods. Rule 92's fix - read the titles - would not have caught it, because the titles of the affected papers are fine. What catches it is reading the experimentalRole column and noticing it is empty for the best assays.

And rule 92's own conclusion partly rests on the grade being sound: it concluded the filter is sound and the coverage gap is real, on hand-verified labels over the bait-supported set. That conclusion is about which publications are focused. It says nothing about partners the grade never let into the bait-supported set in the first place.

What would kill this hypothesis

The falsifiers below are scoreable and two of them are computable from data already on disk. The honest position is that the mechanism is established and the magnitude is not: 11.3% is a minority, the bait/prey grade is right about most of the population, and a reader could reasonably conclude that two good examples out of 24 does not justify changing a triage rule that works for 188.

A registry row points the opposite way, and it is the strongest objection to this hypothesis

data/symbol-collisions.json row 'IntAct curates enzyme-substrate assays as interactions, so a partner census counts substrates as partners' (added 2026-09-22) prescribes excluding enzymatic-assay and structural-fragment evidence from the partner census 4. That is the exact class this hypothesis says is the tail's best evidence.

The row's decisive example is the one this hypothesis leaned on: all 11 KMT2A-H3C1 ITC, NMR and X-ray rows concern the PHD3-bromodomain cassette, which EVERY KMT2A fusion deletes. Beautifully measured, and irrelevant to the index fusion.

So the two positions are not symmetric and the resolution is not 'split the difference'. What the H3C1 case shows is that method quality and fusion relevance are independent axes, and that a re-grade on method class alone would promote exactly the rows the registry wants removed. AKAP8 survives both tests - gel filtration, and the contact is with DPY30 in a complex the fusion works alongside rather than with a deleted cassette - which is why it is the hypothesis's only clean example.

The honest restatement, narrower than the title: the bait/prey grade discards bait-free evidence indiscriminately, and a method-class grade would admit it indiscriminately. Neither is a substitute for asking whether the measured region is in the part of the protein the fusion keeps - which is a third column neither grading has.

Facts and figures
statementThe bait-supported / prey-only grade discards evidence classes rather than weak evidence, because symmetric biophysical methods assign no bait role, and it therefore systematically under-rates the best-evidenced partners in the interactome tail.
origindata-pattern
prior probability0.75
predictions
[
  "Reading the remaining 23 never-prey partners will find at least two more with a focused, hypothesis-driven source paper.",
  "Re-grading the prey-only population on METHOD class rather than on role will move H3C1 and AKAP8 out of the discard bin and leave the BRET and spoke-expanded rows in it.",
  "Other single publications that split across the bait/prey line will be found, and in most of them the grade will agree with evidence strength - so the split mechanism will prove less harmful than the bait-free mechanism."
]
experiments
[
  "exp-q160-re-grade-the-interactome-on-method-class-instead-of-bait-role"
]
analyses
[
  "ana-q160-the-prey-only-interactome-is-36-percent-written-not-zero"
]
falsifiers
[
  "Reading all 24 bait-free prey-only partners finds that 22 or more are BRET screen rows, crosslink-MS artefacts or substrate rows, with at most AKAP8 and H3C1 having a focused source. Scoreable on the census file with no new data: if fewer than 4 of 24 have a hypothesis-driven source paper whose title or abstract names one of the six baits, the class is not enriched for good evidence and the hypothesis is dead.",
  "Re-grading on method class (biophysical/direct vs AP-MS vs proximity vs display) reorders the top 20 of the uncovered tail by fewer than 5 positions. If the two gradings agree, the role column was never doing the damage.",
  "An audit of publications that supply both a bait record and a prey record for our six baits finds that in every case the bait-side partner also has strictly more independent evidence than the prey-side partner. That would show the grade is a lagging indicator of strength rather than an independent and sometimes wrong signal.",
  "The registry row 'IntAct curates enzyme-substrate assays as interactions...' shows that all 11 KMT2A-H3C1 structural rows map to the PHD3-bromodomain cassette every fusion deletes. SCOREABLE EXTENSION OF THAT: if, across the 24 never-prey partners, the majority of bait-free records whose feature ranges are recorded map to KMT2A regions LOST in KMT2A::AFF1 (i.e. C-terminal to the common breakpoint region, or within PHD1-3/BRD/FYRN/TAD/SET), then bait-free evidence is correctly discarded for this project's purpose and this hypothesis is REFUTED regardless of how good the measurements are."
]
measurements
[
  {
    "quantity": "fraction",
    "subject": "human prey-only partners with no literal prey role in any record",
    "value": 0.113,
    "units": null,
    "n": 212,
    "universe": "IntAct REST, six core baits plus two KMT2A taspase1 chains, 2026-09-24, human only",
    "null": null,
    "note": "24 of 212. Methods: crosslink MS, BRET, molecular sieving, FPS, methyltransferase assay."
  },
  {
    "quantity": "count",
    "subject": "bait-free prey-only partners read in full by this beat",
    "value": 1,
    "units": "partners",
    "n": 24,
    "universe": "the 24 never-prey partners; only AKAP8 has been read in full. H3C1 was also read but is NOT one of the 24.",
    "null": null,
    "note": "23 of the 24 unread. This is the hypothesis's main weakness and its cheapest falsifier."
  }
]
How this could bite us back
  • Loosening a triage filter because two examples survived scrutiny is how a corpus fills with weak entries. The honest use of this hypothesis is to add ONE column to the triage - method class - not to promote the bait-free 24 as a group.
Still open
  1. What do the other 23 never-prey partners look like when read? One of one read so far has a focused source, which is the least reliable sample size there is.
  2. Should this project's interactome triage grade on METHOD class rather than on experimental role? The method field is present in every IntAct record and is not currently used for triage anywhere.
  3. Does the same defect affect the BAIT-supported side - are there partners graded bait-supported whose only bait record is a proteome-scale atlas where every protein is a bait? Rule 92 measured this at 15 of 56 and the answer is yes, which means both sides of the grade have a known failure mode and neither is corrected.
Every source behind this entry (4)
  1. E1established confidence 5 of 5: Multiple independent lines of evidence; treated as settled in the field.dataset

    Of the 212 human prey-only partners of the six core proteins in the 2026-09-24 IntAct REST pull, 188 have at least one record where our protein's experimental role is literally 'prey'. The other 24 have none: 15 'neutral component', 8 'fluorescence accept' (BRET), and one 'unspecified role' (AKAP8, molecular sieving). Breaking the 'neutral component' group down: 13 crosslink MS, one methyltransferase assay, one FPS. These are symmetric assays in which neither partner is a bait.

    What this evidence cannot show

    11.3% is a minority and this entry says so. The claim is that the grade is uninformative for those 24, not that the grade is wrong for the population. Eight of the 24 are BRET rows that q128 declined individually on other grounds, so bait-free does not imply good.

    Evidence that argues against this (1)
    • The bait/prey grade was introduced to answer a specific question - has anybody deliberately fished with our protein - and for 188 of 212 partners it answers it correctly. A rule that is right 88.7% of the time and whose failures are identifiable by one extra column is a working rule with a documented exception, not a broken one. this project 2026
  2. E2established confidence 5 of 5: Multiple independent lines of evidence; treated as settled in the field.paper

    The top-ranked uncovered prey-only partner by focused-publication support is H3C1 (P68431), whose 17 records comprise 11 with our protein as 'neutral component' from PMID 20541251 (six ITC, two NMR, one x-ray diffraction, all typed 'direct interaction') and six with our protein as 'prey' (two pull down from the same PMID, four BioID from PMIDs 29568061 and 25281560) - so H3C1 carries bait-free biophysics AND prey records, and is therefore NOT one of the 24. The second-ranked is AKAP8 (O43823), supported by gel-filtration co-elution from PMID 23995757, which IS one of the 24. PMID 23995757 is Jiang, Lu, Shimada, Dou, Tang and Roeder, Nat Struct Mol Biol 2013, titled 'Regulation of transcription by the MLL2 complex and MLL complex-associated AKAP95', which additionally shows endogenous DPY30 co-IP recovering AKAP95, direct binding of purified AKAP95 residues 1-100 to purified DPY30, and enhancement of MLL2-complex H3K4 methylation on recombinant chromatin by recombinant AKAP95.

    Regulation of transcription by the MLL2 complex and MLL complex-associated AKAP95. · Jiang H, Lu X, Shimada M, Dou Y, Tang Z, Roeder RG · Nat Struct Mol Biol 2013
    PMID 23995757DOI
    What this evidence cannot show

    H3C1 is a histone-cluster gene standing in for the H3 tail SUBSTRATE and this corpus already records that it must not become gene-h3c1, so only AKAP8 is a partner-level example. CORRECTED before shipping: this entry first said both were members of the 24-partner bait-free class. H3C1 is not - it has two pull-down prey records alongside its eleven bait-free ones. It illustrates a related but distinct problem, that the grade reads any prey record and never sees the biophysics beside it.

    Evidence that argues against this (1)
    • Two examples, one of which is a substrate rather than a partner, is thin. The remaining 22 bait-free partners have not been read, and if most turn out to be BRET screen rows and crosslink artefacts then the class is not enriched for good evidence and this hypothesis loses most of its force. this project 2026
  3. E3established confidence 5 of 5: Multiple independent lines of evidence; treated as settled in the field.database

    PMID 25751424 yields seven IntAct records involving KMT2A: four anti-tag co-IPs with KMT2A as bait and human SIRT1 as prey (miscore 0.51), one two-hybrid with SIRT1 as bait (0.51), one mouse Sirt1 co-IP with KMT2A as bait (0.40), and one two-hybrid with CLOCK as bait and KMT2A as prey (0.37). SIRT1 is therefore graded bait-supported and CLOCK prey-only from the same publication.

    What this evidence cannot show

    In this particular case the grade and the evidence strength happen to agree - SIRT1 has four co-IP rows and CLOCK has one two-hybrid - so this is a demonstration of the mechanism rather than of harm. The AKAP8 case is where the mechanism causes harm.

    Evidence that argues against this (1)
    • If the grade and the evidence agree whenever a single paper splits, then the split mechanism is harmless and only the bait-free mechanism matters. That has not been checked across other split publications and the census file has the material to check it. this project 2026
  4. E4established confidence 5 of 5: Multiple independent lines of evidence; treated as settled in the field.dataset

    data/symbol-collisions.json carries the row 'IntAct curates enzyme-substrate assays as interactions, so a partner census counts substrates as partners', added 2026-09-22, which prescribes filtering the partner census by detection method to EXCLUDE enzymatic-assay and structural-fragment evidence, and records that all 11 KMT2A-H3C1 ITC/NMR/X-ray rows concern the PHD3-bromodomain cassette that every KMT2A fusion deletes.

    What this evidence cannot show

    Project-internal registry; the PHD3-bromodomain mapping of the 11 rows was not independently re-checked against IntAct feature ranges here. This beat's independent count of 11 direct-interaction H3C1 records agrees with the registry's count.

    Evidence that argues against this (1)
    • This row is the strongest single objection to this hypothesis and is recorded as such rather than argued away. If structural-fragment evidence is routinely about deleted regions, then bait-free evidence is not merely mis-graded but correctly discarded for this project's purpose, and the hypothesis is close to dead on its own. this project 2026
Connected to