The four AFF4 two-hybrid partners are three rows of one screen, not three experiments - and in that same screen GOLGA2 has 249 partners while AFF4 has five
also called AP2B1 · GOLGA2 · MTUS2 · TRAF2 · HI-II-14 · IM-23318 · Rolland 2014
Four proteins were listed as AFF4 partners with three database rows each, which reads like three separate experiments. All twelve rows come from one yeast two-hybrid screen, and three of the four proteins stick to hundreds of other things in that very screen.
The same thing, for a scientist
Four of the seventeen unwritten bait-supported partners named in the 2026-09-21 recount - **AP2B1, GOLGA2, MTUS2 and TRAF2** - were dispatched with "3 IntAct rows" each, which reads as replication. **It is not.** All twelve rows carry one publication (**PMID 25416956**, Rolland et al. 2014) and one IMEx record (**IM-23318**), and the three rows per pair are three PSI-MI method annotations of one CCSB pipeline: `two hybrid prey pooling approach` (primary screen), `two hybrid array` (retest) and `validated two hybrid` (pairwise validation) [E1]. **One screen, three assay records.** All twelve are in host *Saccharomyces cerevisiae* (taxid:4932) - the "human bait-supported" grading counts an interaction assayed in yeast [E1]. Two further measurements, both taken from the same dataset so that the comparator is the assay itself: **AFF4 has exactly five partners in that screen - AP2B1, GOLGA2, MTUS2, SIAH1 and TRAF2 - and three of those five preys are hubs in the same screen**: GOLGA2 **249** distinct partners, MTUS2 **153**, TRAF2 **97**, against AP2B1's **10** and AFF4's own **5** [E2]. A prey found with 249 different baits in one screen is reported because of a property of the prey, not of the bait. **And none of the four pairs is recovered in HuRI (Luck et al. 2020), the successor map from the same laboratory over the same ORFeome space**: IntAct holds 246 GOLGA2, 204 TRAF2, 76 RPRM and 16 PBX2 rows from that publication and **zero AFF4 rows of any kind** [E3]. So the four pairs are unreplicated, and the non-replication is in the one screen best positioned to reproduce them. **Consequences for this beat:** AP2B1 is written at `weak`; GOLGA2, MTUS2 and TRAF2 are written at `speculative` with the hub degree in the first line of each entry. **Consequence for the graph:** a MITAB row count is an assay-annotation count, and any future interactome measurement that ranks partners by rows is ranking curation granularity.
The row count, and why it reads as replication when it is not
PSI-MI gives distinct identifiers to distinct assay protocols, so one interaction detected by a screen with a retest and a validation step generates three MITAB lines. Pulled pairwise from PSICQUIC on 2026-09-21:
| pair | rows | interaction ACs | methods |
|---|---|---|---|
| AFF4-AP2B1 | 3 | EBI-10627110, EBI-10321207, EBI-10473943 | validated two hybrid; two hybrid prey pooling approach; two hybrid array |
| AFF4-GOLGA2 | 3 | EBI-10542896, EBI-10321217, EBI-10473953 | same three |
| AFF4-MTUS2 | 3 | EBI-10627130, EBI-10321239, EBI-10473975 | same three |
| AFF4-TRAF2 | 3 | EBI-10627120, EBI-10321229, EBI-10473965 | same three |
Every row: pubmed:25416956, imex:IM-23318, taxid:4932(yeasx), MI-score 0.56, no expansion.
Note also the role asymmetry: in the two hybrid array and prey pooling rows AFF4 is bait and the partner is prey; in the validated two hybrid row both are neutral component. So the bait/prey role that made these partners "bait-supported" comes from two of the three rows, and in Y2H the bait/prey distinction is the DB/AD fusion orientation - a construct choice, not a statement that anyone set out to study AFF4.
This is the concrete instance of brief rule 90's general point. A count with no persisted members goes stale invisibly; a count whose unit is not what the reader assumes is worse, because it does not need to go stale to mislead.
The internal comparator, which is the strongest thing here
The usual objection to calling a two-hybrid hit an artefact is that the sceptic has no null. Here the null is inside the experiment: the same assay, the same screen, the same reagents, different proteins.
| protein | distinct partners in Rolland 2014 |
|---|---|
| GOLGA2 | 249 |
| MTUS2 | 153 |
| TRAF2 | 97 |
| AP2B1 | 10 |
| AFF4 | 5 |
Rolland et al. report roughly 14,000 interactions in the whole map 4. GOLGA2 alone accounts for about 1.8% of them.
The reading is asymmetric and that asymmetry is the point. A high prey degree is strong evidence that the pair carries little information about the bait. A low prey degree - AP2B1's 10 - removes the commonest artefact explanation and establishes nothing positive; a protein can be a specific contaminant of one bait class and still have few partners.
I have not computed what fraction of Rolland 2014's proteins exceed 97 partners, so "three of four are hubs" is a description of these four and not a test against the screen's degree distribution. The comparison that carries the weight is the within-pair one - 249 against 5 - because both numbers come from the same assay and the ratio does not depend on a denominator I did not compute.
The non-replication, and the limit of what it shows
Luck et al. 2020 (HuRI, PMID 32296183) is the successor screen from the same laboratory, larger and over an expanded ORFeome. IntAct's holdings of it contain no AFF4 row at all - not the four pairs, not any other partner - while GOLGA2 (246), TRAF2 (204), RPRM (76) and PBX2 (16) are all present 3.
What this cannot distinguish: whether AFF4's four 2014 hits were tested in HuRI and failed, or whether the AFF4 ORF was absent from the HuRI search space. IntAct records what was found, not what was attempted, and I did not retrieve HuRI's screened-space table. The honest statement is "not recovered in HuRI; whether AFF4 was screened there is not resolvable from IntAct", and that is what the four entries say.
Even at that strength it is informative, because the alternative - that AFF4 was dropped from the search space - would itself mean the pairs have had no opportunity to replicate in six years.
What I decline to conclude
Three of the four preys are long coiled-coil or rod-like proteins - GOLGA2 is the GM130 golgin, MTUS2's alternative name is cardiac zipper protein, TRAF2 carries a coiled-coil - and the fourth, AP2B1, is an alpha-solenoid adaptin and the one that is not a hub. That correlation is suggestive and I am not asserting it, because with n = 4 and no base rate for coiled-coil content among Rolland 2014's hubs it is exactly the shape brief rule 84 warns about: a pattern that feels like an observation and is a statistic with no null.
It is recorded as an open question with the test attached.
| script | none - PSICQUIC pairwise and per-interactor queries, recorded in the evidence URLs |
|---|---|
| status | run |
| result summary | Twelve MITAB rows across four AFF4 pairs reduce to one publication and one screen. Prey degree within that screen: GOLGA2 249, MTUS2 153, TRAF2 97, AP2B1 10, against AFF4's 5. No AFF4 row exists in IntAct's holdings of HuRI (Luck 2020). |
| rows per pair | 3 |
| distinct publications per pair | 1 |
| distinct experiments per pair | 1 |
| assay host | Saccharomyces cerevisiae (taxid:4932) |
| intact miscore | 0.56 |
| generalisable finding | A MITAB row count is a count of PSI-MI assay annotations, not of experiments or of publications. Any partner ranking built on row counts ranks curation granularity. Count distinct publications, and read the expansion column before calling a row a binary interaction. |
inputs
[ "IntAct PSICQUIC MITAB 2.7 pairwise queries, nine partners x six core baits, retrieved 2026-09-21", "IntAct PSICQUIC per-interactor queries for Q9UHB7, P63010, Q08379, Q5JR59, Q12933, partitioned by source publication" ]
aff4 partners in rolland 2014
[ "AP2B1", "GOLGA2", "MTUS2", "SIAH1", "TRAF2" ]
already written of those five
[ "gene-siah1" ]
prey degree within pmid 25416956 — the numbers behind the chart
{
"GOLGA2": 249,
"MTUS2": 153,
"TRAF2": 97,
"AP2B1": 10,
"AFF4": 5
}huri luck 2020 rows in intact
{
"AFF4": 0,
"GOLGA2": 246,
"TRAF2": 204,
"RPRM": 76,
"PBX2": 16,
"note": "absence for AFF4 cannot distinguish 'tested and failed' from 'not in the search space'"
}confidence assigned downstream
{
"gene-ap2b1": "weak",
"gene-golga2": "speculative",
"gene-mtus2": "speculative",
"gene-traf2": "speculative"
}- Downgrading these four is itself a claim. If AP2B1 or TRAF2 turns out to contact the SEC in a haematopoietic cell, this entry will have told a future reader to ignore it, and the prey-degree argument is exactly the kind of prior that suppresses real hits in a sticky-protein family.
- Nothing here touches an intervention. The risk it creates is editorial: a reader who takes 'three rows is one screen' as a reason to discount every atlas-derived row will discount the BioPlex evidence in this beat too, where the row structure is different and the reasoning does not transfer.
- Three of the four preys are long coiled-coil or rod-like proteins and the fourth, the only non-hub, is not. Is coiled-coil content predictive of prey degree in Rolland 2014? The test is one join of the screen's degree table against a coiled-coil predictor over all its proteins, and until it is run the observation is a pattern with no null.
- What fraction of Rolland 2014's interactors exceed 97 partners? Without that denominator, 'three of four are hubs' describes these four rather than testing them, and the denominator is one pass over the same MITAB dump already retrieved.
- Was the AFF4 ORF in HuRI's search space? The HuRI screened-space table answers it in one lookup and converts a bounded non-replication into either a real negative or a real absence of testing.
- Does the same row-count-as-experiment-count error inflate other partner rankings in this project? The interactome recount and the bait-versus-prey analysis both operate on MITAB rows, and neither deduplicates to publications.
- E1established confidence 5 of 5: Multiple independent lines of evidence; treated as settled in the field.database
All twelve IntAct rows joining AFF4 to AP2B1, GOLGA2, MTUS2 and TRAF2 come from PMID 25416956 / IMEx IM-23318, in host Saccharomyces cerevisiae, as three PSI-MI method annotations per pair: two hybrid prey pooling approach, two hybrid array, and validated two hybrid.
IntAct molecular interaction database, PSICQUIC MITAB 2.7, pairwise queries id:Q9UHB7 AND id:{P63010,Q08379,Q5JR59,Q12933}, retrieved 2026-09-21 · IntAct (EMBL-EBI) 2026What this evidence cannot show
PSICQUIC returns IntAct's holdings only; another IMEx partner could hold additional rows. Three assay annotations of one pipeline is the standard CCSB workflow and is not a defect in the curation - the defect is reading the row count as an experiment count, which is a reader error, and one this beat's dispatch brief made.
Evidence that argues against this (1)
- A retest and a pairwise validation are genuine additional measurements within the pipeline, not mere re-annotations of a single observation. Treating the three rows as one experiment slightly understates the work done, in the opposite direction from treating them as three. Rolland T et al. 2014
- E2strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.database
Within PMID 25416956, AFF4 has 5 distinct partners (AP2B1, GOLGA2, MTUS2, SIAH1, TRAF2) while GOLGA2 has 249, MTUS2 153, TRAF2 97 and AP2B1 10.
IntAct molecular interaction database, PSICQUIC MITAB 2.7, per-interactor queries with publication-level partner deduplication, retrieved 2026-09-21 · IntAct (EMBL-EBI) 2026What this evidence cannot show
Degrees computed from IntAct's import of the publication, deduplicated to distinct UniProt accessions per publication, not from the paper's own supplementary tables; IntAct's import may be incomplete or may merge isoform rows. maxResults was set to 8000 and no interactor approached that ceiling, so no truncation occurred. These are degrees within one dataset and are not the proteins' degrees in the literature.
Evidence that argues against this (1)
- A high-degree protein is not automatically a false positive. GOLGA2 is a golgin scaffold and TRAF2 a signalling adaptor, and both genuinely have large interactomes. Degree flags a pair as uninformative about the bait; it does not prove the pair is wrong. Huttlin EL et al. 2021
- E3moderate confidence 3 of 5: One strong study, or several smaller consistent ones.database
IntAct holds no AFF4 interaction row from PMID 32296183 (HuRI, Luck et al. 2020) while holding 246 GOLGA2, 204 TRAF2, 76 RPRM and 16 PBX2 rows from the same publication, so none of the four 2014 AFF4 two-hybrid pairs is recovered in the successor map.
IntAct molecular interaction database, PSICQUIC MITAB 2.7, per-interactor queries partitioned by source publication, retrieved 2026-09-21 · IntAct (EMBL-EBI) 2026What this evidence cannot show
This is `moderate` and not `strong` because IntAct records what a screen found, not what it attempted. Whether AFF4's ORF was in HuRI's search space and failed to score, or was absent from it, cannot be resolved from IntAct; the HuRI screened-space table was not retrieved. Either reading leaves the four pairs unreplicated, which is the claim the four entries make.
Evidence that argues against this (1)
- Yeast two-hybrid has a well-known sampling problem and any single screen recovers only a fraction of true binary interactions, so failure to recover a pair in a second screen is weak evidence against it. That is precisely why the load-bearing observation here is the prey degree, not the non-replication. Rolland T et al. 2014
- E4established confidence 5 of 5: Multiple independent lines of evidence; treated as settled in the field.paper
Rolland et al. 2014 describe a systematic map of approximately 14,000 high-quality human binary protein-protein interactions obtained by yeast two-hybrid, presented as a reference interactome rather than as a study of any particular protein.
A proteome-scale map of the human interactome network. · Rolland T, Taşan M, Charloteaux B, et al., Vidal M · Cell 2014What this evidence cannot show
Read as abstract, not full text, and new to this graph - not previously cited here. The authors' own framing is that the map is deliberately unbiased and systematic, which is a strength for network-level conclusions and precisely why an individual pair from it carries little weight about either protein.
Evidence that argues against this (1)
- The authors argue their systematic map is of equal or better quality than the literature-curated interactome and about 30% larger at equal quality, so dismissing its pairs as low-grade is not what the paper supports. The correct discount is for the absence of any focused follow-up on this particular pair, not for the assay class. Huttlin EL et al. 2017
AFF4's five two-hybrid partners in Rolland 2014, and the prey degrees that make four of them unusable
- related to →SIAH1 — the E3 ligase that degrades AFF1, and that the reciprocal AF4::KMT2A fusion escapes
the fifth AFF4 two-hybrid partner from the same screen, and the only one already written
- supports →A BioID hit is a neighbour, not a complex member - and the two methods are known to disagree
same family of argument: the assay class determines what a row can assert
the non-hub control that makes the degree argument informative rather than circular
the highest-degree prey of the four
second-highest-degree prey of the four
third hub prey of the four