start / graph / ana-prey-recovery-tracks-size-in-proximity-labelling-and-not-in-ap-ms-and-the-tail-was-scored-on-the-wrong-side-of-the-pair

AnalysisstrongQ136-interactome-tail-size-and-abundance

Prey recovery in the two BioID depositions tracks SIZE and not abundance - and the interactome tail was being scored on the wrong side of the pair, because every one of the 22 partners is the BAIT and every core protein is the PREY

also called the size-and-abundance adjudication of the interactome tail · IM-28767 and IM-30059 prey-recovery regression · the orientation correction

Two big screens are the only evidence for 22 proteins this project lists as partners of KMT2A and its complex. We asked whether those screens pick a protein up because it genuinely sticks to the bait, or just because it is big and common. Size matters a little, commonness does not matter consistently at all - and the bigger discovery was that we had been measuring the wrong protein of each pair.

The same thing, for a scientist

Pre-registered (`analyses/interactome-tail-size-and-abundance-adjudication/PREREGISTRATION.md`) before any row was fetched: regress the number of distinct baits recovering a prey on that prey's length and cellular abundance, over IMEx **IM-30059** (Khoroshkin 2024, 50 RBP BioID2 baits, K562, 19,403 rows) and **IM-28767** (Goos 2022, 109 TF baits, HEK293 Flp-In, 8,331 rows) [E1][E2]. **Result on the coefficients: log10 length is positive and consistent in both proximity arms (negative-binomial standardised beta +0.289 and +0.205), abundance FLIPS SIGN (+0.165 in K562, -0.314 in HEK293), and McFadden pseudo-R2 is 0.054 and 0.076 - so the pre-registered M1 (both terms >= 0.20, R2 >= 0.10) fails and so does M2, giving the pre-registered NEITHER branch** [E3]. **The sharpest coefficient result is a contrast within one publication:** in IM-28767's own AP-MS arm, run by the same lab on the same 109 baits in the same cells, the length coefficient collapses to **+0.030 (p = 0.27, not significant)** while abundance turns **+0.239** - so the size effect is a property of **proximity labelling**, not of protein interaction [E3]. **And the finding that reorganises the beat is orientation:** in both depositions **none of the six core proteins is a bait** and **22 of 22 partners that appear at all are baits**, so the size/abundance confound applies to KMT2A, MEN1, PSIP1, AFF1, AFF4 and DOT1L **as preys**, and every partner of a given core protein in a given deposition inherits the same prey-side statistic and cannot be ranked against the others by it [E4].

Go deeper

The pre-registered decision, and which branch fired

Fixed before the first fetch: M1 (size and abundance explain recovery) required both standardised coefficients >= 0.20 and McFadden pseudo-R2 >= 0.10 in both proximity arms. M2 (they do not) required pseudo-R2 < 0.05 in both. Anything else was the declared NEITHER branch, whose stated consequence was: issue no class verdict from the regression, fall back to a matched empirical null.

Observed 3:

arm	n preys	b log10 length	b log10 abundance	McFadden R2	Pearson dispersion
IM-30059 BioID2, K562	2,627	+0.289	+0.165	0.054	4.95
IM-28767 BioID, HEK293	933	+0.205	-0.314	0.076	13.24
IM-28767 AP-MS, HEK293	631	+0.030 (n.s.)	+0.239	0.026	-

NEITHER fired, on two counts at once: neither arm clears the R2 floor, and abundance is positive in one deposition and negative in the other. Size is the only term that behaves the same way twice.

The orientation correction, which is the main finding and which the pre-registration got wrong

The pre-registered outcome variable was the partner's own bait count. Reading the first output, ZC3H11A came back at 32 of 50 baits and U2AF2 at 19 - impossible for proteins this project records as one-bait partners. The resolution is that core-complex-interactome-gap's n_baits counts how many of the six core proteins a partner touches, which is a different quantity from a deposition's bait panel.

But checking it surfaced the real problem. Parsed from the rows themselves, with PSI-MI experimental roles MI:0496/MI:0498 4:

IM-30059: none of KMT2A, MEN1, PSIP1, AFF1, AFF4, DOT1L is a bait. 13 of the 22 partners ARE baits - DDX52, DDX55, EFTUD2, EXOSC5, FASTKD2, HNRNPA1, RPS11, SF3B4, U2AF2, UTP3, ZC3H11A, ZC3H8, ZNF800.
IM-28767: same - no core protein is a bait; 9 of the 22 are - ELF1, EN1, FEV, KLF12, KLF3, NFIX, NFYC, NHLH1, TBXT.

So for 22 of 22 the pair exists because the partner was tagged and the core protein turned up in its neighbourhood. The confound the beat was commissioned to test applies to KMT2A, not to DDX52. And it has a consequence that no amount of better regression fixes: all thirteen IM-30059 partners of KMT2A share one prey-side number (25 of 50), so prey recovery cannot order them against each other at all.

This is the same failure family as exp-no-proximity-labelling-has-ever-been-done-on-a-kmt2a-fusion and ana-q104-every-row-in-the-read-queue-is-prey-side: this project's core proteins have never been baits in a modern proximity experiment, and every quantitative instrument built on these depositions inherits that.

My own model was over-confident by a factor of five, and the check that caught it

A binomial refit first reported KMT2A's 25-of-50 recovery in IM-30059 as 2.06x predicted, p = 7.4e-05 - which contradicted ana-kmt2a-proximity-promiscuity-is-explained-by-size..., whose length-matched decile null put KMT2A at the 90.7th percentile (p ~ 0.09, EXPLAINED). Two instruments, one deposition, one 25, opposite verdicts.

Before writing that as biology I tested the assumption under it. Pearson dispersion is 4.95 in IM-30059 and 13.24 in IM-28767 5. Bait counts in a proximity map are heavily overdispersed and a Poisson likelihood assumes they are not, so its tail probabilities are far too small. KMT2A's upper-tail p for the same 25:

model	p
Poisson, as first reported	9.8e-04
quasi-Poisson, dispersion-scaled	0.052
negative binomial (alpha = 0.50)	0.110

The check now passes and the prior analysis stands. KMT2A at 25 against a predicted 12.5 is p = 0.110 - not distinguishable from its size and abundance, in agreement with the 90.7th percentile it appeared to refute. ana-kmt2a-proximity-promiscuity-is-explained-by-size... survives its own generalisation, and now survives it with the abundance term that analysis explicitly declared missing as its U2.

The pre-registered empirical arbiter DECLINED for KMT2A and that is reported, not hidden: matching preys to KMT2A on both length (+/-25%) and abundance (+/-0.5 log10) yields a pool of 14, and length-only yields 30 - both under the pre-registered floor of 40. The prior analysis reached n = 268 by using a whole length decile, a far wider bin. For KMT2A the assumption-free test cannot run, so the negative binomial is the best available statement and it is model-based.

What the core proteins look like as preys, which is the number that actually adjudicates the tail

IM-30059, 50 RBP BioID2 baits in K562 45:

prey	baits	of panel	predicted (negbin)	ratio	matched pool
KMT2A	25	50%	12.5	2.0x	pool n=14, DECLINED
PSIP1	9	18%	7.8	1.2x	n=229, 66th pct
AFF1	4	8%	7.6	0.5x	n=160, 46th pct
AFF4	4	8%	7.9	0.5x	n=192, 39th pct
DOT1L	4	8%	7.9	0.5x	n=83, 48th pct
MEN1	3	6%	6.2	0.5x	n=213, 30th pct

IM-28767, 102 TF BioID baits in HEK293 Flp-In: KMT2A 11 of 102 (predicted 12.2, ratio 0.9, p = 0.42 - exactly what size and abundance predict), DOT1L 2 of 102 (predicted 11.5, ratio 0.17 by the model; 21.4th percentile of a matched pool of 42 whose median is 6, by the empirical test) 5.

Two readings of that table are worth keeping apart. (i) KMT2A is the only core protein above its own prediction anywhere, and only in the RBP/K562 map. (ii) The other five are all at or below prediction, so "our proteins are sticky" is false as a general statement. MEN1 in particular is recovered by 3 of 50 at half its predicted rate and the 30th percentile of its size- and abundance-matched peers - menin is a hard protein to reach by proximity labelling, which is a fact about the instrument that any future menin BioID needs to plan around.

The pair-level statistic, and why nothing here is significant

Since prey recovery cannot rank partners, the pre-registration's NEITHER fallback was re-specified for the corrected orientation: a null that holds both margins of the bipartite bait x prey graph - each bait's reach and each prey's recoverability - and asks how often that specific pair appears. Implemented as exact weighted sampling without replacement (exponential-race), 20,000 draws, seed 20260923, with the analytic Chung-Lu value printed beside every simulated one as a control 6.

62 pairs were tested (49 in IM-30059, 13 in IM-28767). The two smallest are GATA3 -> DOT1L p = 0.032 and TBXT -> DOT1L p = 0.039. A Bonferroni threshold over 62 tests is 0.0008. Nothing in either deposition is significant after correction, and this entry does not claim otherwise. The p-values are used as an ordering of an otherwise unorderable tail - which is precisely what ana-the-interactome-tail-is-orderable-after-all-by-cost-to-complete-and-not-by-evidence-strength said was missing - and never as a test.

Applying the pre-registered thresholds (EXPLAINED at p >= 0.20, PROMOTED at p < 0.05) to the 22: 12 EXPLAINED, 9 NOT RESOLVED, 1 PROMOTED 6.

EXPLAINED, declined as a class with a measurement: DDX52, EFTUD2, KLF3, U2AF2, NFIX, KLF12, EXOSC5, ZNF800, HNRNPA1, FEV, SF3B4, ZC3H11A.
NOT RESOLVED: UTP3, RPS11, NFYC, FASTKD2, EN1, NHLH1, ELF1, DDX55, ZC3H8.
PROMOTED: TBXT.

An EXPLAINED verdict does not say the proteins do not interact. It says this deposition carries no information about that pair beyond what the two degrees already imply.

Two of the nine were already adjudicated and a gap list could not see it

Before writing anything, every candidate symbol was grepped across all 2,536 entries and the hits read. UTP3 has a full declined entry of its own - interface-psip1-utp3-declined-psip1-is-prey-to-nine-of-fifty-rbp-baits - and RPS11 carries a ribosomal-class rejection recorded in the gap file itself. Both are in the gap list's all_missing because own_entry matches gene-*/protein-* ids and cannot see an interface-*-declined-* node.

That is the standing trap - a gap list cannot see what was decided - firing again, at 2 of 22 (9%). It cost nothing here because the grep was run first, and it is recorded so the next beat does not have to rediscover the rate.

What was dropped, named

Every filter's drop count and a sample of the dropped identifiers are in results.json. Totals: IM-30059, zero rows dropped at parse; IM-28767, 10 rows dropped for a prey with no resolvable gene symbol, plus 1,631 rows of other detection methods (the AP-MS and two-hybrid arms, fitted separately rather than discarded) and 1 unresolved symbol. In the join to covariates, IM-30059 lost 54 of 2,681 preys (no UniProt length or no PaxDb abundance) and IM-28767's BioID arm lost 13 of 946. The dropped symbols are listed; they are dominated by protein-group leading ids of the form RL10_RL10L_HUMAN, which this graph already has a concept entry about.

Abundance source: a deviation from the pre-registration, recorded not hidden. The pre-registration named HPA rna_celline. That file returns HTTP 404 as of 2026-09-23, the HPA download index now lists only proteinatlas.tsv.zip and proteinatlas.json.gz, neither carries per-cell-line nTPM, and search_download.php has no per-cell-line column specifier - all verified by fetch 7. Substituted: PaxDb v6.1 9606-WHOLE_ORGANISM-integrated (protein ppm, 99% coverage, 19,483 symbols), with the host-line-matched 9606-K562_Geiger_2012 and 9606-HEK293_Geiger_2012 as a sensitivity fit. The substitution moves toward a stronger covariate - protein copies rather than transcript - which makes EXPLAINED easier and PROMOTED harder, i.e. conservative for the promotions this beat acted on.

Facts and figures

script	analyses/interactome-tail-size-and-abundance-adjudication/run.py
preregistration	analyses/interactome-tail-size-and-abundance-adjudication/PREREGISTRATION.md
status	run
preregistered verdict code	NEITHER
orientation	22 of 22 partners are BAITS; 0 of 6 core proteins is a bait in either deposition
result summary	Size predicts proximity recovery consistently (standardised negative-binomial beta +0.29 and +0.21) and predicts AP-MS recovery not at all (+0.03, n.s.) in the same publication; abundance flips sign between depositions; pseudo-R2 is 0.05-0.08 so the pre-registered class verdict fails both ways. The finding that reorganises the beat is that all 22 partners are baits and all six core proteins are preys, so prey recovery scores the wrong protein and cannot rank the tail. A both-margins pair null declines 12 of the 22 as a class and promotes 1.

supporting scripts

[
  "analyses/interactome-tail-size-and-abundance-adjudication/orientation.py",
  "analyses/interactome-tail-size-and-abundance-adjudication/pair_null.py",
  "analyses/interactome-tail-size-and-abundance-adjudication/overdispersion.py",
  "analyses/interactome-tail-size-and-abundance-adjudication/adjudicate.py"
]

inputs

[
  "IntAct PSICQUIC tab27 pubid:39251607 (19,403 rows, endpoint count 19,403)",
  "IntAct PSICQUIC tab27 pubid:35140242 (8,331 rows, endpoint count 8,331)",
  "UniProt reviewed human proteome stream: accession, gene_primary, length (20,190 symbols)",
  "PaxDb v6.1 9606-WHOLE_ORGANISM-integrated protein ppm (19,483 symbols, coverage 99%)",
  "PaxDb v6.1 9606-K562_Geiger_2012 and 9606-HEK293_Geiger_2012 (sensitivity)"
]

negative binomial coefficients

{
  "IM-30059_bioid2_k562": {
    "n": 2627,
    "beta_log10_length": 0.289,
    "beta_log10_abundance": 0.165,
    "pearson_dispersion": 4.95,
    "alpha": 0.5
  },
  "IM-28767_bioid_hek293": {
    "n": 933,
    "beta_log10_length": 0.205,
    "beta_log10_abundance": -0.314,
    "pearson_dispersion": 13.24,
    "alpha": 1
  }
}

poisson coefficients apms arm for contrast

{
  "IM-28767_apms_hek293": {
    "n": 631,
    "beta_log10_length": 0.0303,
    "p_length": 0.27,
    "beta_log10_abundance": 0.2392,
    "mcfadden_pseudo_r2": 0.0261
  }
}

core proteins as prey

{
  "IM-30059_of_50_baits": {
    "KMT2A": 25,
    "PSIP1": 9,
    "AFF1": 4,
    "AFF4": 4,
    "DOT1L": 4,
    "MEN1": 3
  },
  "IM-28767_of_102_baits": {
    "KMT2A": 11,
    "DOT1L": 2
  }
}

adjudication of the 22

{
  "EXPLAINED": [
    "DDX52",
    "EFTUD2",
    "KLF3",
    "U2AF2",
    "NFIX",
    "KLF12",
    "EXOSC5",
    "ZNF800",
    "HNRNPA1",
    "FEV",
    "SF3B4",
    "ZC3H11A"
  ],
  "NOT_RESOLVED": [
    "UTP3",
    "RPS11",
    "NFYC",
    "FASTKD2",
    "EN1",
    "NHLH1",
    "ELF1",
    "DDX55",
    "ZC3H8"
  ],
  "PROMOTED": [
    "TBXT"
  ]
}

already adjudicated before this beat

{
  "UTP3": "interface-psip1-utp3-declined-psip1-is-prey-to-nine-of-fifty-rbp-baits",
  "RPS11": "ribosomal-class rejection, cycle 57 Q104, recorded in core-complex-interactome-gap results.json"
}

How this could bite us back

This analysis declines 12 partners as a class. If any of the twelve later turns out to be a real and druggable contact, the decline will have delayed it - and the decline rests on an ordering, not on a significance test, because nothing here survives multiplicity correction.
The size coefficient is a reason to distrust proximity evidence for LARGE targets specifically. KMT2A at 3,969 residues is the largest protein in this project's core, so the instrument's main bias points directly at its main subject.

Still open

Why does abundance predict recovery POSITIVELY in the K562 RBP map (+0.17) and NEGATIVELY in the HEK293 TF map (-0.31)? Two competing readings: the TF panel labelled a chromatin-restricted subvolume in which abundant cytoplasmic proteins are actively depleted, or the two studies filtered their bona-fide lists differently before deposition. The second is cheap to test - compare the deposited pair count against each paper's stated raw hit count - and nobody has.
Size predicts BioID recovery and does not predict AP-MS recovery in the same publication, same baits, same cells. Is that a lysine-count effect (a bigger protein presents more biotinylatable residues in the labelling radius) or a residence-time effect? Counting accessible lysines per prey and refitting with lysine count in place of length would separate them in an afternoon.
MEN1 is recovered by 3 of 50 RBP baits at half its predicted rate and the 30th percentile of its size- and abundance-matched peers. Is menin genuinely hard to reach by proximity labelling - buried in a compact complex, or lysine-poor on its accessible surface - and if so, what does that imply for the design of the menin-bait proximity experiment this graph keeps asking for?
Every quantitative statement in this beat is prey-side, because no core protein has ever been a proximity bait. What would the same regression say if run on a deposition where KMT2A::AFF1 IS the bait? That experiment does not exist, which is the point.

Every source behind this entry (7)

E1established confidence 5 of 5: Multiple independent lines of evidence; treated as settled in the field.paper
Goos et al. identified 6,703 BioID and 1,536 AP-MS protein-protein interactions for 109 human transcription factors in HEK293 Flp-In cells, deposited as IMEx IM-28767; IntAct's PSICQUIC endpoint serves 8,331 rows for PMID 35140242, of which 946 prey-resolving rows are MI:1314 BioID, 641 are MI:0004 affinity chromatography and 33 are MI:0007 anti-tag co-immunoprecipitation.
Human transcription factor protein interaction networks. · Göös H, Kinnunen M, Salokas K, Tan Z, Liu X, Yadav L, Zhang Q, Wei GH, Varjosalo M · Nat Commun 2022
PMID 35140242 DOI
What this evidence cannot show
Abstract read; full text not read. The row-level counts are this beat's own parse of the PSICQUIC tab27 stream, paginated to exhaustion and checked against the endpoint's ?format=count (8,331 fetched, 8,331 declared). Every bait is an over-expressed tagged construct in embryonic kidney cells. The 109 baits in the paper resolve to 102 distinct bait symbols in the BioID arm of the deposition, which this beat did not reconcile.
Evidence that argues against this (1)
- PSICQUIC is a floor relative to IntAct REST (rule 93: 161 vs 376 rows for Q03164), so every denominator here is a lower bound. The coefficients are estimated on what PSICQUIC serves; whether REST's extra rows would change them was not tested. source 2026
E2established confidence 5 of 5: Multiple independent lines of evidence; treated as settled in the field.paper
Khoroshkin et al. fused BioID2 to 50 human RNA-binding proteins in K562 cells and deposited the curated pair list as IMEx IM-30059; IntAct's PSICQUIC endpoint serves 19,403 rows for PMID 39251607, all MI:1314, resolving to 50 baits and 2,681 human preys, which matches the 19,403 rows two prior analyses in this repository recorded.
Systematic identification of post-transcriptional regulatory modules. · Khoroshkin M, Buyan A, Dodel M, Navickas A, Yu J, Trejo F, et al. · Nat Commun 2024
PMID 39251607 DOI
What this evidence cannot show
Abstract read; full text not read. The authors themselves state that RBP over-expression, BioID2 fusion and the use of a single isoform per protein may bias the data. K562 is a BCR::ABL1 chronic myeloid leukaemia line with no KMT2A rearrangement, so this is a leukaemia cell but not the right one.
Evidence that argues against this (1)
- The bait panel is 50 RNA-binding proteins. A prey that genuinely resides in the RNA-processing compartment is recovered widely for a real biological reason that this analysis cannot separate from complex membership. Pre-registered as U4 and not controlled. Go CD et al. 2021
E3strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.dataset
Regressing a prey's distinct-bait count on standardised log10 protein length and standardised log10 PaxDb abundance gives, under a negative binomial, beta_length +0.289 and beta_abundance +0.165 in IM-30059's BioID2 arm (n=2,627) and beta_length +0.205 and beta_abundance -0.314 in IM-28767's BioID arm (n=933); in IM-28767's own AP-MS arm on the same baits and cells the length coefficient falls to +0.030 (p=0.27) while abundance is +0.239. McFadden pseudo-R2 is 0.054, 0.076 and 0.026 respectively.
analyses/interactome-tail-size-and-abundance-adjudication/ - PREREGISTRATION.md, run.py, orientation.py, overdispersion.py, results.json (this repository) 2026
What this evidence cannot show
Pre-registered before any fetch. Covariates are protein length from the UniProt reviewed human proteome and PaxDb v6.1 whole-organism integrated protein ppm - NOT host-line matched, and not RNA. A nuclear-residence term was requested by the commissioning memo and is NOT in the model (declared as U5): omitting it means length and abundance absorb any correlation with nuclear residence, so the model OVER-explains, which makes an EXPLAINED call generous and a PROMOTED call conservative. The AP-MS contrast is a point-estimate comparison; its standard error is inflated by unmodelled overdispersion like the others.
Evidence that argues against this (1)
- A pseudo-R2 of 0.05-0.08 means size and abundance together explain under a tenth of the variation in bait recovery. That cuts both ways: it is why the class cannot be declined by the regression, and it is equally a reason not to read the positive length coefficient as a large effect. Per standard deviation of log10 length the incidence rate ratio is only about 1.33. source 2026
E4established confidence 5 of 5: Multiple independent lines of evidence; treated as settled in the field.database
Parsed from PSI-MI experimental roles MI:0496 (bait) and MI:0498 (prey) in the MITAB 2.7 rows, none of KMT2A, MEN1, PSIP1, AFF1, AFF4 or DOT1L is a bait in either IM-30059 or IM-28767, while 13 of the 22 partners under adjudication are baits in IM-30059 and 9 are baits in IM-28767.
IntAct molecular interaction database (EMBL-EBI), PSICQUIC MITAB 2.7 for pubid:39251607 and pubid:35140242, fields 19 and 20 (experimental roles), retrieved 2026-09-23 2026
What this evidence cannot show
Roles are the curator's assignment, not a measurement. Both depositions are MI:1060 spoke-expanded, so no binary pair was ever measured: the bait-prey direction is exactly the information spoke expansion preserves and it is the only direction information these rows carry.
Evidence that argues against this (1)
- This graph has previously reported that PSICQUIC tab27 'does not label the bait reliably per row' and counted both orientations instead. That statement is wrong for these two depositions - every row here carries MI:0496/MI:0498 - but it may hold for other depositions, and a parser that assumes roles are always present will silently drop rows where they are not. This one counts and names what it drops. source 2026
E5strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.dataset
Pearson dispersion of the fitted Poisson models is 4.95 in IM-30059 and 13.24 in IM-28767, so Poisson tail probabilities are anti-conservative; KMT2A's 25-of-50 recovery in IM-30059 has upper-tail p = 9.8e-04 under Poisson, 0.052 under a dispersion-scaled quasi-Poisson and 0.110 under a negative binomial with alpha = 0.50, against a predicted 12.5 baits. The empirical length-and-abundance-matched pool for KMT2A contains only 14 preys and is declined under the pre-registered floor of 40.
analyses/interactome-tail-size-and-abundance-adjudication/overdispersion.py and overdispersion.json (this repository) 2026
What this evidence cannot show
The negative binomial's alpha was profiled over a fixed grid by likelihood, not estimated jointly. The empirical matched pool - the only assumption-free test - DECLINES for KMT2A because the pool is too small, so the p = 0.110 that reconciles this analysis with the prior one is itself model-based. A different dispersion model could move it.
Evidence that argues against this (1)
- Agreement between two instruments after one of them is corrected is weaker evidence than agreement before. The Poisson fit disagreed with the prior decile null and the negative binomial agrees with it; the correction was made because the assumption was demonstrably violated (dispersion 4.95), not because it produced agreement, but a reader should hold the reconciliation more loosely than the raw counts. source 2026
E6moderate confidence 3 of 5: One strong study, or several smaller consistent ones.dataset
Under a null holding both margins of the bipartite bait-by-prey graph (exact weighted sampling without replacement, 20,000 draws, seed 20260923), 62 bait-to-core-protein pairs were scored across the two depositions; the two smallest p-values are GATA3-DOT1L at 0.032 and TBXT-DOT1L at 0.039, and applying the pre-registered thresholds to the 22 partners under adjudication gives 12 EXPLAINED, 9 NOT RESOLVED and 1 PROMOTED.
analyses/interactome-tail-size-and-abundance-adjudication/pair_null.py and pair_null.json (this repository) 2026
What this evidence cannot show
A Bonferroni threshold over 62 tests is 0.0008; NOTHING is significant after correction and this is used as an ordering, not a test. The null also assumes preys are drawn independently given their degrees, which is false when baits belong to the same complex and share preys - that makes the null anti-conservative in the same direction as the overdispersion problem above. The analytic Chung-Lu value is printed beside every simulated p and tracks it closely, which checks the simulation but not the assumption.
Evidence that argues against this (1)
- A both-margins null penalises exactly the baits that labelled widely, so a genuinely promiscuous protein that is nonetheless a real partner is scored as uninformative. The ordering this produces is an ordering of INFORMATION CONTENT in this deposition, not of biological plausibility, and a partner ranked EXPLAINED here may still be the right one. source 2026
E7established confidence 5 of 5: Multiple independent lines of evidence; treated as settled in the field.database
The Human Protein Atlas per-cell-line RNA download named in the pre-registration no longer exists: https://www.proteinatlas.org/download/rna_celline.tsv.zip and the spelling rna_cellline.tsv.zip both return HTTP 404, the download index lists only proteinatlas.tsv.zip and proteinatlas.json.gz, the JSON carries only an 'RNA cell line specific nTPM' summary field rather than per-line values, and the search_download.php column table contains no per-cell-line specifier.
Human Protein Atlas v25.1 downloadable data and programmatic access pages, fetched 2026-09-23 2026
What this evidence cannot show
Checked by fetch on one day. HPA reorganises its downloads between releases and the file may exist under a path this beat did not try; four spellings and the API column table were tried and are named in run.py.
Evidence that argues against this (1)
- PaxDb's host-line-matched Geiger 2012 datasets ARE available and were used as a sensitivity fit, but they cover only about 3,400 (K562) and 4,300 (HEK293) proteins by spectral counting, and what a shallow proteome misses is not random - it misses the rare proteins, which are exactly the ones the abundance term is supposed to discriminate. That is why the near-complete integrated set is primary. source 2026

Connected to

supports →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
the generalisation reproduces the special case once overdispersion is modelled: KMT2A 25 of 50 vs 12.5 predicted, negative-binomial p = 0.110
tests →Wild-type KMT2A is one of the most promiscuously labelled proteins in the proximity record - 25 of 50 baits - and menin is one of the least, at 3
re-derives the 25 of 50 and adds the abundance term the source analysis declared missing
depends on →The 50-RBP BioID2 proximity map in K562 - the first modern proximity dataset this graph can reach, and AFF1 is in it as prey
IM-30059, 19,403 rows, 50 baits, 2,681 human preys
depends on →A 109-transcription-factor BioID network puts KMT2A next to MYC, YY1 and the ETS and KLF families - and AF10 next to ERG and KLF5
IM-28767, 8,331 rows; independently re-derives the eleven TF baits recovering KMT2A and the two recovering DOT1L
supports →PLANNED: run the size-matched promiscuity correction for PSIP1 in the K562 BioID2 map, the way it was run for KMT2A - fully specified here so the next cycle executes rather than redesigns it
PSIP1 at 9 of 50 is 1.2x predicted and the 66th percentile of a 229-member length- and abundance-matched pool - not promiscuous beyond its size
supports →A BioID hit is a neighbour, not a complex member - and the two methods are known to disagree
size predicts BioID recovery (+0.21 to +0.29) and not AP-MS recovery (+0.03, n.s.) in the same publication
related to →KMT2A / MLL1 — the gene that breaks
KMT2A is the only core protein recovered above its size-and-abundance prediction, and only in the RBP/K562 map
related to →MEN1 / menin — the adaptor the fusion cannot work without
menin is recovered by 3 of 50 RBP baits, half its predicted rate and the 30th percentile of its matched peers - hard to reach by proximity labelling
supports →PSIP1-UTP3 is one spoke-expanded BioID2 proximity row in K562 with UTP3 as bait and PSIP1 as prey at miscore 0.27; PSIP1 is recovered by 9 of the 50 baits in that screen and only 2 of the 9 are nucleolar, which is the panel's own base rate - declined
UTP3 was already declined; a gap list cannot see an interface-*-declined-* entry, measured here at 2 of 22
related to →ELF2 / NERF - an ETS factor that works with RUNX1 on a B-cell kinase promoter, and one of eleven TF baits recovering KMT2A
answers this entry's leading open question - the eleven TF rows are one observation, not eleven, and ELF2's own row is EXPLAINED-adjacent at p = 0.141
← supportsTBXT / brachyury - the notochord master regulator, and one of only two of 102 transcription-factor baits whose neighbourhood contained DOT1L
the only partner of 22 to clear the pre-registered PROMOTED threshold, p = 0.039 of 62 pairs
← supportsZC3H8 / Fliz1 - the last little-elongation-complex subunit without an entry, discovered in fetal liver haematopoietic progenitors, and a selective BioID2 bait that reaches KMT2A in K562
a relatively selective bait (152 of 2,681 preys, 40th of 50) whose KMT2A row is still NOT RESOLVED at p = 0.185
← supportsELF1 - a direct transcriptional target of MLL fusion proteins and HOXA9 that is ALSO a BioID neighbour of KMT2A, and that is focally deleted in paediatric AML without a KMT2A rearrangement
ranked NOT RESOLVED; the case for the entry comes from outside the deposition, which is the beat's point about the tail
← supportsNFYC / NF-YC - a histone-fold transcription factor whose partner subunit carries an H2B-like ubiquitination, and the most selective of the eleven TF baits that reached KMT2A
the narrowest of the eleven TF baits (49 preys vs a panel median of 58.5) and the lowest p, in an ordering that is monotonic in bait width
← supportsDDX55 - a large-subunit rRNA helicase that is nucleolar by imaging and nucleoplasmic by its own founding paper, and the narrowest RBP bait to reach KMT2A after DDX52
one of the three narrowest RBP baits reaching KMT2A, all three NOT RESOLVED
← supportsNHLH1 / HEN1 / NSCL1 - a TAL1-family bHLH factor found by cross-hybridising with the T-ALL oncogene, whose knockout mouse is normal and whose only imaging call puts it in the GOLGI
third of the eleven by p and third by bait narrowness - the ordering is the ordering of labelling breadth
← refutesKMT2A-EN1 declined: one spoke-expanded BioID row from an ectopic neural homeobox construct in a kidney cell, and the only screen that ever looked found EN1 unexpressed in AML lines
declined on biology AGAINST its own ranking (p = 0.123, second of eleven) - a case where the pair statistic and the evidence disagree
supports →KMT2A's proximity promiscuity is BAIT-CLASS-SPECIFIC: it is recovered at twice its size-and-abundance prediction by RNA-binding-protein baits in a leukaemia cell and at exactly its prediction by transcription-factor baits in a kidney cell
the 2.0x versus 0.9x contrast between the RBP and TF panels is this analysis's output
← testsPLANNED: refit prey recovery on accessible lysine count instead of protein length, which separates the two mechanisms that could produce the size effect
the discriminating refit for the size coefficient measured by the parent analysis