Oncofusion

start / graph / hyp-atlas-only-spokes-are-predicted-by-prey-abundance-not-by-the-bait

Hypothesisspeculative confidence 1 of 5: Our own inference or analogy. Not shown to be true.Q75-dot1l-partnersdisputed

Atlas-only interactome partners are predicted by how abundant the prey is, not by which bait was pulled - and if so this graph's interactome denominator is inflated

also called abundance predicts the atlas residue · the interactome denominator is inflated

TESTED AND IT DID NOT HOLD. The idea was that the weakest partners found by huge screens are mostly just abundant proteins that stick to everything, which would have made the list still to write up shorter than it looks - but across all 31,944 results of the screen, how common a protein is barely affects whether it lands in the weak group.

The same thing, for a scientist

**OUTCOME, cycle 43: NOT SUPPORTED in the form that matters.** See `data.tested` and `E_k43_tested`. Abundance moves B- membership by an odds ratio of 1.04 per SD; what predicts A+ is the prey:bait ratio. The DOT1L spike is decisive without a model - MLLT6 is A+ at 169 copies while EBNA1BP2 is B- at 1.7 million. The statement below is retained unaltered as the hypothesis as written. **Statement.** Among the low-confidence spokes of proteome-scale AP-MS atlases, membership is predicted by the prey's cellular abundance and by nothing about the bait. The DOT1L pulldown from PMID 26496610 is consistent with this - its most abundant prey, EBNA1BP2 at 1.7 million copies per HeLa cell, has the highest interaction stoichiometry and the lowest enrichment in the experiment, and the reciprocal GFP-IBTK pulldown in the same study returned the same ribosomal and nucleolar background class under a completely unrelated bait [E1][E2]. **Consequence if true.** The 54-partner denominator behind the graph's 68.5% interactome-coverage figure includes an atlas-only residue of 15 partners whose membership says nothing about the six core proteins, and the figure should be reported twice - once over all bait-supported partners and once over partners with at least one focused source. **This is a hypothesis and not a finding**: the supporting observations are two pulldowns, and the test that would settle it is a regression over all 1,330 baits in the atlas, which is specified here and has not been run. **It is also not a licence to prune** - AF9, a structurally validated DOT1L partner, is more substoichiometric than any of the six [E1], so abundance filtering would remove true partners if applied naively.

Go deeper
What would kill it

The regression is the whole test and it can fail in three informative ways.

  • If prey abundance explains little of the variance in low-class spoke membership, the hypothesis is simply wrong and the atlas residue has to be read partner by partner, which is expensive and is then the honest answer.
  • If abundance predicts membership equally well for the A+ class, then abundance is a property of what mass spectrometry can see at all rather than of contamination, and the hypothesis explains nothing about confidence.
  • If the DOT1L pulldown's residue is not enriched for high-abundance preys relative to a random bait's residue, then the specific case that motivated this is unremarkable and the generalisation has no anchor.
Facts and figures
statementAmong low-confidence-class spokes of proteome-scale AP-MS atlases, membership is predicted by prey cellular abundance and not by the identity of the bait; consequently the atlas-only residue of this graph's interactome denominator carries little information about the six core proteins.
origindata-pattern
prior probability0.55
predictions
[
  "Across all 1,330 baits in PMID 26496610, prey cellular abundance will predict membership in the B/B- classes substantially better than it predicts membership in the A+ class.",
  "The nucleolar and ribosomal preys of the DOT1L pulldown will appear in a large fraction of unrelated baits' low-class prey lists.",
  "Partners supported only by atlases will be systematically less likely to be replicated by any later focused study than partners with a focused source, measured over the 56 bait-supported partners of the six core proteins.",
  "The prediction that would make this interesting rather than obvious: interaction stoichiometry will NOT separate low-class from high-class spokes, because AF9 already violates that ordering."
]
analyses
[
  "ana-the-six-new-dot1l-partners-are-nine-spokes-of-one-pulldown"
]
experiments
[
  "exp-abundance-versus-confidence-regression-over-the-hein-atlas"
]
falsifiers
[
  "Prey abundance explains little variance in low-class spoke membership across the 1,330 baits - the hypothesis is wrong.",
  "Abundance predicts A+ membership as well as it predicts B- membership - abundance is about detectability, not contamination, and the hypothesis explains nothing about confidence.",
  "The DOT1L pulldown's low-class residue is not enriched for high-abundance preys relative to a random bait's residue - the motivating case is unremarkable.",
  "A focused study replicates any one of the six DOT1L spokes - that partner is real and the residue is not background."
]
tested
{
  "when": "2026-09-21, cycle 43",
  "by": "analyses/hein-abundance-versus-confidence/",
  "outcome": "NOT SUPPORTED in the form that matters",
  "prediction_1": "absolute claim FAILS (abundance odds 1.04 per SD on B-); comparative ordering holds only because abundance predicts A+ negatively (-0.288)",
  "prediction_2": "CONTRADICTED - promiscuity across bait cell lines has a negative coefficient on B- membership (-0.134)",
  "prediction_3": "UNTESTED - independent replication of atlas-only partners is still open",
  "practical_consequence": "the proposed triage of the 15 atlas-only partners by copy number does not exist; they must be read or left standing with their provenance recorded"
}
How this could bite us back
  • The direct risk of this hypothesis is that it licenses pruning, and pruning is how the cycle-34 figure closed a line of work for seven cycles. Nothing here should remove a partner from the denominator; it should split the reporting.
  • A regression that confirms it would hand the graph a tempting automated filter, and AF9 is the standing proof that such a filter would discard a structurally validated partner.
Still open
  1. If the regression holds, should the ledger report interactome coverage as two numbers - all bait-supported partners, and partners with a focused source - or as one number with a stated source-type split?
  2. Does the same abundance signature explain the atlas-only residue for the other five core baits, or is DOT1L's residue unusual?
Every source behind this entry (4)
  1. E1established confidence 5 of 5: Multiple independent lines of evidence; treated as settled in the field.dataset

    In the GFP-DOT1L pulldown of PMID 26496610, the prey with the highest cellular abundance (EBNA1BP2, 1,732,944 copies per HeLa cell, 466-fold over the bait) has the highest interaction stoichiometry (0.442) and the lowest enrichment (1.056 SD) of any prey, while MLLT3/AF9 - a structurally validated DOT1L partner graded A+ - has the lowest interaction stoichiometry in the experiment at 0.0058.

    What this evidence cannot show

    Nine preys from one pulldown. This is the observation that generated the hypothesis, so it cannot also be its test - the regression named in data.analyses is the test, and it has not been run. Downloaded 2026-09-21 from the Elsevier static CDN with magic bytes checked.

    Evidence that argues against this (1)
    • The AF9 row is a direct counter-example to the naive version of this hypothesis: a true, structurally proven partner sits at the very bottom of the stoichiometry ranking. Any filter derived from this idea must be checked against AF9 or it will remove real biology. Cell 163(3), supplementary material 2015
  2. E2moderate confidence 3 of 5: One strong study, or several smaller consistent ones.dataset

    The reciprocal GFP-IBTK pulldown in the same study, with a cytoplasmic bait unrelated to chromatin, returned a prey list dominated by the same abundant ribosomal and nucleolar class (seven ribosomal proteins plus DDX21, NAT10, TSR1, WDR74, POP7 and RPP25L), which is the background signature this hypothesis predicts under any bait.

    What this evidence cannot show

    One comparison bait. IBTK has reported functions in translation, so its ribosomal prey list may be specific rather than background, which would remove this observation's force entirely. That ambiguity is precisely why the regression over all baits is the test rather than any pair of anecdotes.

    Evidence that argues against this (1)
    • If IBTK genuinely is a translational regulator, then the shared ribosomal signature is two specific results rather than one background class, and this evidence item supports the opposite conclusion. UniProt Consortium 2026
  3. E3moderate confidence 3 of 5: One strong study, or several smaller consistent ones.analysis

    The graph's current interactome coverage figure of 68.5% is computed over 54 bait-supported partner accessions with no weighting by source type, and 15 of the 56 bait-supported partners across the six baits rest purely on proteome-scale atlases.

    What this evidence cannot show

    The 15/56 atlas-only figure comes from a separate coordinator analysis run during this beat and is quoted, not re-derived. The 68.5% figure is deliberately generous by construction - it counts a partner as present if accession, node id OR symbol matches - so splitting it by source type would change the number in both directions at once.

    Evidence that argues against this (1)
    • Reporting coverage over a reduced denominator is exactly how the cycle-34 figure came to suppress seven cycles of work. A split figure must keep the full denominator visible alongside it, or this hypothesis will have bought a cleaner number at the cost of the same failure. Oncofusion knowledge graph (internal analysis) 2026
  4. E_k43_testedstrong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.analysis

    TESTED AND NOT SUPPORTED, one cycle after being written, over all 31,944 rows joined to the HeLa proteome at 97.3%. Prediction 1 said prey abundance would predict B/B- membership substantially better than A+ membership. **The comparative ordering holds and the absolute claim fails, and the difference matters**: standardised log-odds are +0.041 for B- (odds 1.04 per SD, i.e. nothing) against -0.288 for A+. Abundance is 'better' at predicting B- only because it predicts A+ NEGATIVELY, not because it predicts B- at all. Unadjusted, the B- share FALLS across abundance deciles, 28.6% to 24.0%. Prediction 2 is separately contradicted: promiscuity across bait cell lines has a NEGATIVE coefficient on B- membership (-0.134), so preys appearing under many unrelated baits are LESS likely to be low-class, not more. What A+ actually tracks is the log10 prey:bait ratio (+0.884, odds 2.42 per SD) - stoichiometry, not abundance.

    What this evidence cannot show

    One atlas, one cell background, one pipeline. This tests the hypothesis where it was stated - the Hein atlas - and does not establish a general property of AP-MS. Confidence class was constructed partly from stoichiometry, so 'stoichiometry predicts class' is partly tautological. The load-bearing result is the negative about abundance. The adjusted and unadjusted abundance coefficients disagree in SIGN. The marginal is preferred here because the adjustment conditions on promiscuity and prey:bait ratio, which are themselves abundance-related, so the model partials out the proposed pathway. Either way the effect is ~1.04 per SD. The hypothesis's own third prediction - that atlas-only partners replicate less often in independent datasets - is UNTESTED here and remains open. It is the one prediction that would still rescue the practical conclusion.

    Evidence that argues against this (1)
    • A defender could say prediction 1 was comparative and the comparison did come out in the predicted direction, so the hypothesis survives in weak form. That reading is available and is why the outcome is recorded as 'absolute claim fails, comparative ordering holds trivially' rather than a flat refutation. What the hypothesis cannot survive is its practical purpose: it was proposed so the 15 atlas-only partners could be triaged by copy number, and an odds ratio of 1.04 per SD cannot triage anything. Hein MY et al. 2015
Connected to