Two-thirds of this graph's core-complex interaction evidence comes from experiments where our protein was not the bait, and LEDGF has one bait-supported partner out of thirty-six
also called bait versus prey · interactome provenance · PSIP1 interactome · read queue grading
The database this project draws fusion partners from records whose experiment each interaction came from, and nobody had looked. Two-thirds of the partners turned up incidentally in somebody else's net rather than being deliberately fished for - and for LEDGF, central to this project's main hypothesis, it is thirty-five out of thirty-six.
The same thing, for a scientist
IntAct records an experimental role per interactor. MITAB 2.7 exposes it; MITAB 2.5, which this project's existing interactome analyses use, does not. Grading every partner of the six core-complex proteins by whether **our** protein was ever the bait: | bait | partners | bait-supported | prey-only | other | |---|---:|---:|---:|---:| | **PSIP1 (LEDGF)** | 36 | **1** | 29 | 6 | | KMT2A | 91 | 16 | 70 | 5 | | AFF1 | 35 | 7 | 27 | 1 | | MEN1 | 31 | 7 | 14 | 10 | | AFF4 | 71 | 21 | 47 | 3 | | DOT1L | 26 | 14 | 10 | 2 | | **total pairs** | **290** | **66 (22.8%)** | **197 (67.9%)** | 27 (9.3%) | **And every one of the seven top rows of the curated read queue is prey-only** — BUD13, CAMKV, ILF3, MED19, PIP4K2A, RPS11, YWHAG. The queue ranks by bait count and publication count and records nothing about whose experiment produced the row.
What prey-only does and does not mean
It does not mean the interaction is wrong. A protein detected in someone else's pulldown may be a perfectly real partner. Large-scale interactome projects are how much of molecular biology's interaction data exists and refusing it would leave this graph with almost nothing.
It means the experimental context belongs to another question. The cell line, the tagging, the expression level, the conditions and the controls were chosen to study a different protein. A partner found because KMT2A drifted into an RNA-binding protein's BioID in K562 tells you something, but not that anyone asked what binds KMT2A.
The 9.3% 'other' column is deliberately separate and is not prey-only. Those are rows where IntAct records neutral component, unspecified role or similar — typically mass spectrometry of a purified complex where no single protein was bait. That is directed work on the complex and should not be penalised.
And 'bait-supported' is a floor, not a measure of effort. It means at least one row exists somewhere in which our protein was the bait. A partner with one bait row and a partner with twenty score identically here.
PSIP1 is the number that should worry this project
One bait-supported partner out of thirty-six.
LEDGF/PSIP1 is not peripheral here. It is the chromatin tether in the menin-KMT2A-LEDGF module, it is the subject of a standing ledger item about an unperformed removal experiment, and the graph reasons about its interactions repeatedly. Almost none of that interaction evidence comes from an experiment that set out to ask what LEDGF binds.
By contrast DOT1L is the healthiest at 14 of 26 (54%), which fits a protein that has been the direct object of a drug-development programme — people have fished with it deliberately.
The spread itself is the useful output. It is a per-protein measure of how much directed interaction work exists, and it tracks how much commercial and academic attention each protein has had rather than how important it is to this disease.
What follows for the read queue and for the proposed experiment
The read queue needs a provenance column, not a smaller list. Seven of seven top rows being prey-only does not mean they should be dropped — it means a beat writing from them should say, in the claim, that our protein was not the bait and name the system the row came from. That is rule 73 applied to interaction data.
And it sharpens the case for the proximity-labelling experiment this graph proposed this cycle. A TurboID of KMT2A::AFF1 in an index-fusion line would generate bait-supported evidence, in the right cell type, for a protein whose current interactome is 82% other people's by-products. On this measurement that is not one more dataset — it is the first directed one at scale.
A caution against the obvious over-correction. This project has twice discovered a provenance problem and briefly over-read it. The right response is to record the role, not to purge the prey-only rows — which are two-thirds of everything the graph knows about these proteins' partners.
per bait — the numbers behind the chart
{
"AFF1": {
"partners": 35,
"bait_supported": 7,
"prey_only": 27,
"neither": 1
},
"AFF4": {
"partners": 71,
"bait_supported": 21,
"prey_only": 47,
"neither": 3
},
"DOT1L": {
"partners": 26,
"bait_supported": 14,
"prey_only": 10,
"neither": 2
},
"KMT2A": {
"partners": 91,
"bait_supported": 16,
"prey_only": 70,
"neither": 5
},
"MEN1": {
"partners": 31,
"bait_supported": 7,
"prey_only": 14,
"neither": 10
},
"PSIP1": {
"partners": 36,
"bait_supported": 1,
"prey_only": 29,
"neither": 6
}
}totals — the numbers behind the chart
{
"pairs": 290,
"bait_supported": 66,
"prey_only": 197,
"neither": 27
}read queue graded
{
"BUD13": [
"KMT2A:prey",
"MEN1:prey"
],
"ILF3": [
"KMT2A:prey",
"PSIP1:prey"
],
"RPS11": [
"KMT2A:prey",
"MEN1:prey"
],
"YWHAG": [
"KMT2A:prey"
],
"CAMKV": [
"AFF1:prey",
"AFF4:prey"
],
"PIP4K2A": [
"AFF1:prey",
"AFF4:prey"
],
"MED19": [
"AFF4:prey"
]
}- Do not purge prey-only rows on the strength of this. They are two-thirds of everything this graph knows about these proteins' partners, many are real, and this project has twice over-read a provenance discovery before. The fix is a recorded role, not a smaller graph.
- And do not read 22.8% as a quality score for the interactome. It measures whose experiment produced each row, which correlates with how much commercial attention a protein has had - DOT1L, with a drug programme behind it, scores best - rather than with how well established any individual interaction is.
- Would a TurboID of KMT2A::AFF1 in an index-fusion line change these numbers materially, or would it add one bait-supported source against 197 prey-only rows?
- Why is PSIP1 at 1 of 36? Is that neglect, or has LEDGF been studied by structural and genetic methods that generate no IntAct bait rows?
- Do the other fourteen read-queue entries grade the same way? Only the seven the ledger names were checked.
- E1strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.analysis
Across IntAct records for the six core-complex proteins KMT2A, MEN1, PSIP1, AFF1, AFF4 and DOT1L, 290 partner-bait pairs resolve to an experimental role, of which 66 (22.8%) include at least one interaction where the core-complex protein was the bait, 197 (67.9%) rest only on rows where it was prey, and 27 (9.3%) carry neither role; PSIP1 is the extreme at 1 bait-supported partner of 36 and DOT1L the healthiest at 14 of 26.
interactome-bait-versus-prey: experimental-role grading of core-complex interaction partners from IntAct MITAB 2.7 · Oncofusion knowledge graph (internal analysis) 2026What this evidence cannot show
Run 2026-09-20 against IntAct via PSICQUIC. **Roles are curator-assigned and incompletely populated** - the 9.3% carrying neither role are mostly purified-complex mass spectrometry, which is directed work and is deliberately counted separately rather than as prey. **Counted as partner-bait PAIRS, not unique proteins**, so a protein partnering three baits contributes three. **Bait-supported is a floor**: one bait row scores the same as twenty. IntAct curates what was published and pools screens of very different quality; this grades PROVENANCE and says nothing about whether an interaction is real.
Evidence that argues against this (1)
- A low bait-supported fraction partly measures the era and economics of interaction proteomics rather than neglect of a protein. Since roughly 2015 the field has been dominated by large-scale affinity and proximity atlases in a few workhorse cell lines, which generate enormous numbers of prey detections for every protein in the proteome. A protein studied intensively by other methods - structural biology, genetics - would still score poorly here. Khoroshkin M et al. 2024
- E2strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.database
All seven of the highest-ranked entries on this project's curated core-complex interactome read queue - BUD13, CAMKV, ILF3, MED19, PIP4K2A, RPS11 and YWHAG - are supported only by interactions in which the core-complex protein was prey rather than bait.
IntAct molecular interaction database via PSICQUIC, MITAB 2.7 records for KMT2A, MEN1, PSIP1, AFF1, AFF4 and DOT1L with experimental roles · EMBL-EBI IntAct 2026What this evidence cannot show
The queue was built from the same IntAct data, so this is a re-grading of a known list rather than an independent check of it. Seven entries were graded because they are the ones the ledger names; the remaining fourteen queue rows were not individually graded here. Accession-level matching means a queue entry recorded under a different isoform accession could be missed.
Evidence that argues against this (1)
- Several of these are plausible partners on other grounds and should not be dismissed. BUD13 and ILF3 are RNA-associated proteins and the index fusion sits in an elongation complex with real RNA biology; MED19 is a Mediator subunit and Mediator contact is independently expected. Prey-only provenance lowers the grade of the evidence, not the prior on the biology. EMBL-EBI IntAct 2026
QUANTIFIES ITS POINT ACROSS THE WHOLE COMPLEX: 67.9% of partner-bait pairs are prey-only, and all seven top read-queue rows are.
ONE BAIT-SUPPORTED PARTNER OUT OF THIRTY-SIX. Almost no interaction evidence for LEDGF comes from an experiment asking what LEDGF binds.
THE HEALTHIEST OF THE SIX at 14 of 26 bait-supported, which fits a protein with a drug programme behind it.
- related to →KMT2A / MLL1 — the gene that breaks
16 OF 91 PARTNERS BAIT-SUPPORTED: most of what the graph records as KMT2A's interactome is other people's by-products.
THE SAME FAILURE IN A DIFFERENT LAYER: interaction rows carry no indication of whose experiment or which cells, and rule 73 applies to them too.
THE COMPLEMENT OF ITS FINDING: all 56 bait-supported partners are written, so the 136 absent ones are entirely prey-only.
A worked instance: six partners of a core protein, all somebody else's proteome-scale experiment, all one pulldown
not the grading, the denominator: that analysis counted 290 partner pairs through PSICQUIC where the REST record holds 449, and PSIP1's 36 was drawn from 43.4% of its record