The IntAct interactome gap is closed - all 135 'unwritten' interactors are mentioned and 60% carry an explicit decision - while 95 of the 110 partners unique to the INDEX fusion have never been mentioned at all
also called q178 · the gap is not where the priority says · prey census vs Skucha
**Corrected: the number of leftover partners worth writing is zero, not one.** The single candidate turned out to rest entirely on a yeast assay from a genome-wide screen, which my filter had misread as a focused study. --- The project's standing instruction says roughly a third of known interaction partners have been written up, implying two thirds are waiting. Checking each one: every single supposedly missing partner is already discussed somewhere, and most were examined and explicitly turned down for good reasons. So that gap is essentially closed. But a different list - the partners found only with this project's own fusion, in the one experiment that looked - is almost entirely untouched: 95 of 110 are never mentioned anywhere. The work that remains is the opposite of where it was thought to be.
The same thing, for a scientist
> 🔴 **CORRECTED, same cycle, before deploy: the count is ZERO, not one.** I wrote that **ARFIP1** was "the single credible, undecided, unwritten interactor in the entire prey-only census". I then fetched its one supporting PMID. **32296183 is Luck *et al.*, *Nature* 2020;580:402-408 — HuRI, the all-by-all human binary interactome, ~53,000 interactions.** That is an atlas by construction. It passed my "≥1 focused publication" test only because q160's heuristic defines an atlas as *a publication supporting ≥15 distinct prey-only partners **across these six baits***, and HuRI contributed fewer than 15 *here*. **A genome-scale screen is an atlas whatever it contributed to one bait set.** > > Two further defects in my own filter, both visible in the same row and both missed: > - ARFIP1's three "methods" are `two hybrid array`, `two hybrid prey pooling approach` and `validated two hybrid` — **three flavours of ONE assay.** My `n_methods ≥ 1` criterion read that as methodological support; it is one Y2H experiment. > - `host_cells` is **_Saccharomyces cerevisiae_**. There is **no human-cell observation of ARFIP1 with any of these baits at all**, which no criterion in my filter tested for. > > **So the credible-and-undecided count is 0 of 135, and the IntAct prey-only gap is closed outright.** The substantive conclusion below is strengthened, not weakened — but the number I published an hour earlier was wrong, and it was wrong because I applied an inherited heuristic without checking what it classified. --- 🟢 **The IntAct prey-only gap is effectively closed, and the headline figure misdescribes it.** IntAct's REST census gives **212** clean human prey-only partners across the six baits: **77 with a `gene-` entry, 135 without** — the source of the standing *"roughly a third have an entry"*. Checking all 135 by **content** rather than by id [E1]: - **135 of 135 (100%) are mentioned somewhere in the corpus.** - **81 of 135 (60%) carry an entry whose id explicitly marks a decision** — `declined`, `absence-`, `refused`, `rejected`. - **0 are never mentioned.** `entry: None` in the census answers *"is there a `gene-` entry?"*, not *"has this corpus considered it?"* Those differ by 60 percentage points. 🔵 **And the credible remainder is one gene.** Filtering the 135 on evidence quality — not a histone or ribosomal protein, ≥1 non-atlas publication, IntAct miscore ≥0.40, ≥2 records — keeps **5 of 135** [E2]. Of those five, **four are already explicitly declined** (BRAF, CAMKV, EPB41L1, CD226, each with a named decline entry). **ARFIP1 (P53367, miscore 0.56) is the single credible, undecided, unwritten interactor in the entire prey-only census.** The filter's exclusions, reported rather than hidden: **108** fail miscore <0.40, **105** rest on a single record, **72** are atlas-only, **14** are histone or ribosomal. 🔴 **Meanwhile the real gap sits outside that denominator entirely.** **IntAct has never curated Skucha 2018 (PMID 29777171), the only MLL-fusion AP-MS interactome**, so it is absent from every interactome-coverage figure this project quotes. Its **MLL-AF4** column — and **MLL-AF4 is KMT2A::AFF1** — lists 291 interactors, **110 of which appear in MLL-AF4 and no other fusion** [E3]: | | pan-fusion core (all 7 baits) | AF4-specific (1 bait) | |---|---:|---:| | partners | 42 | 110 | | held by this corpus | **42 (100%)** | **1 (0.9%)** | | never mentioned at all | 0 | **95** | **The corpus has completely written the partners shared by every MLL fusion and skipped the partners unique to its own.** The shared core is the *least* index-specific material in the table, and it is the part that got written. 🟠 **Why this happened is legible in the numbers.** The pan-fusion core is where evidence is strongest and where every review points, so it is what a literature-led beat reaches first. The AF4-specific tail is low-spectral-count, rank-selected material that looks like noise — **and much of it probably is.** But *probably noise* was never the recorded reason; it was never examined. **0.9% is not a judgement, it is an absence of one.**
What the corrected figures are, for quoting
The standing priority says "Q5 interactome — the largest standing gap by design, roughly a third of documented interactors have an entry." Corrected:
- IntAct bait-supported set: 51/51 (100%) —
check-pi-dimensions.py, this cycle. - IntAct prey-only set: 77/212 have a
gene-entry (36.3%), but 212/212 are considered and ≥60% of the remainder carry an explicit recorded decision. The honest statement is considered, not written. - Skucha MLL-AF4 column: 73/291 (25.1%) overall; 54/105 (51.4%) at p<0.05; 25/35 (71.4%) at p<0.01 — and 1/110 (0.9%) of the AF4-specific subset.
The one number that should drive dispatch is 0.9%, and it is not in the standing priority at all.
The measurement this rests on, and where it undercounts
Both halves use a word-boundary content search over every entry — (?<![A-Za-z0-9])SYMBOL(?![A-Za-z0-9]) — so CD226 does not match CD2264 and RPL13 does not match RPL13A. That is the fix for the failure mode this project keeps hitting, where a substring or id search answers a different question from the one asked.
The decision count is a LOWER bound and deliberately so. It requires the id of a mentioning entry to contain declin|refus|absence-|rejected|deferred. A refusal written in prose inside an entry with an ordinary id is invisible to it. This corpus has already recorded exactly that failure — a list of 15 'cleanest gaps' that contained zero real gaps, because refusals hid in prose, table cells and other entries. So the true decided fraction is ≥60%, and 60% should never be quoted as if it were the ceiling.
What the search cannot do is judge quality of consideration. Being mentioned in a table row of a rejected read-queue is 'considered' by this measure and is very thin. The claim here is that the 135 are not overlooked; it is not that each was examined well.
| script | analyses/q178-which-unwritten-interactors-are-actually-credible/{run,decided}.py; analyses/q179-the-af4-specific-interactors-nobody-wrote/run.py |
|---|---|
| status | executed |
| intact denominator | 212 |
| intact with entry | 77 |
| intact without entry | 135 |
| intact mentioned | 135 |
| intact with decision entry | 81 |
| intact never mentioned | 0 |
| intact credible after filter | 5 |
| skucha af4 interactors | 291 |
| skucha af4 specific | 110 |
| skucha af4 specific with entry | 1 |
| skucha af4 specific mentioned | 15 |
| skucha af4 specific never mentioned | 95 |
| skucha pan fusion core held | 42 of 42 |
| verdict | the IntAct prey gap is closed as a question of consideration; the real omission is the AF4-specific tail of the only MLL-fusion AP-MS interactome, which IntAct has never curated |
| intact credible and undecided corrected | 0 |
intact credible and undecided
[]
filter removals — the numbers behind the chart
{
"miscore_below_0.40": 108,
"single_record": 105,
"atlas_only": 72,
"histone_or_ribosomal": 14
}skucha open targets p lt 0 05
[ "RGPD8", "ACADM", "RPA2", "LEF1", "GLUD1", "CEP112" ]
measurements — the numbers behind the chart
[
{
"quantity": "share",
"subject": "IntAct prey-only 'unwritten' interactors mentioned in the corpus",
"universe": "135 with no gene- entry",
"value": 1
},
{
"quantity": "share",
"subject": "IntAct prey-only 'unwritten' interactors with an explicit decision entry",
"universe": "135 with no gene- entry",
"value": 0.6
},
{
"quantity": "count",
"subject": "credible AND undecided unwritten IntAct interactors (corrected)",
"universe": "135 with no gene- entry",
"value": 0
},
{
"quantity": "share",
"subject": "Skucha AF4-SPECIFIC partners held by the corpus",
"universe": "110 partners unique to the MLL-AF4 bait",
"value": 0.009
},
{
"quantity": "share",
"subject": "Skucha pan-fusion core partners held by the corpus",
"universe": "42 partners shared by all 7 baits",
"value": 1
},
{
"quantity": "count",
"subject": "Skucha AF4-specific partners never mentioned in the corpus",
"universe": "110 partners unique to the MLL-AF4 bait",
"value": 95
}
]arfip1 rejected because
{
"only_pmid": "32296183 (HuRI, Luck 2020 Nature 580:402-408) - a proteome-scale atlas",
"methods_are_one_assay": [
"two hybrid array",
"two hybrid prey pooling approach",
"validated two hybrid"
],
"host_cells": "Saccharomyces cerevisiae - no human-cell observation"
}- FIX q160's atlas heuristic. It defines an atlas as a publication supporting >=15 distinct prey-only partners ACROSS THESE SIX BAITS, so a genome-scale screen contributing few partners to this bait set is admitted as 'focused'. HuRI (~53,000 interactions) was admitted that way. Scale should be a property of the publication, not of its overlap with one bait list - and every prior analysis that used `focused_pmids` inherits this defect.
- Add a host-species/host-cell criterion to any future interactor filter. ARFIP1's evidence is entirely Saccharomyces cerevisiae and nothing in the filter noticed; IntAct's host_cells field was present in the row the whole time.
- Re-measure the decided fraction with a PROSE-level decision detector, not an id-level one. 60% is a lower bound and the gap between it and the truth is the size of this corpus's habit of recording refusals inside ordinary entries.
- Decide the AF4-specific tail as a class rather than one gene at a time. If the honest verdict is 'low-spectral-count rank-selected background', that is one analysis entry covering 110 proteins and it is a legitimate outcome - but it must be written, because 0.9% coverage with no recorded reason is indistinguishable from an oversight.
- Does any second AP-MS dataset of a KMT2A fusion exist to check 'AF4-specific' against? The claim currently rests on one experiment's seven baits, and a partner absent from the other six for depth or tagging reasons would be misclassified as index-specific.
- E1strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.dataset
Of the 135 IntAct prey-only interactors that q160's census marks as having no entry, a word-boundary content search over all 3,031 corpus entries finds 135 of 135 (100%) mentioned somewhere, 81 of 135 (60.0%) mentioned by at least one entry whose id contains an explicit decision marker (declin/refus/absence-/rejected/deferred), and 0 never mentioned. Examples of decision entries: interface-menin-cd226-declined-proteome-scale-ap-ms, interface-kmt2a-abt1-declined-bioplex-prey-row, interface-kmt2a-tmco1-declined-an-er-calcium-channel-crosslinked-in-nuclei.
What this evidence cannot show
The decision count requires the marker in the ID of a mentioning entry, so a refusal recorded only in prose is not counted and 60% is a lower bound. The mention count makes no quality judgement: appearing in a rejected read-queue table counts as a mention. Gene symbols that are also common words would over-match, though none of the 135 obviously is.
Evidence that argues against this (2)
- 100% mentioned is partly an artefact of how this corpus works: it writes large enumeration entries that list every row of a screen, so any protein in any curated screen is almost guaranteed a mention. That makes 'mentioned' a weak bar and means the figure demonstrates thoroughness of LISTING rather than of consideration.
- A recorded decision is not necessarily a correct or durable one. This project has overturned its own declines before - the HDAC6 refusal stood for five years of literature and was wrong - so 60% decided should not be read as 60% settled. The declines rest largely on spoke-expansion and cell-line-expression arguments that were themselves made from abstracts.
- E2strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.dataset
Filtering the 135 on: not histone or ribosomal by symbol prefix, at least one focused (non-atlas) PMID, IntAct best_miscore >= 0.40, and at least 2 supporting records, keeps 5 and removes 130. Removal reasons (a gene can fail several): miscore below 0.40 in 108, single record in 105, atlas-only in 72, sticky class in 14. The 5 kept are BRAF (0.60), ARFIP1 (0.56), CAMKV (0.53), EPB41L1 (0.53), CD226 (0.53). Re-checked against decisions, BRAF, CAMKV, EPB41L1 and CD226 all carry explicit decline entries; ARFIP1 does not and is mentioned once. CORRECTED: ARFIP1 also fails, on re-checking its source. Its only PMID, 32296183, is Luck et al., Nature 2020;580:402-408 (HuRI), a proteome-scale all-by-all binary interactome of ~53,000 interactions - an atlas, admitted by q160's heuristic only because it contributed <15 prey-only partners across these particular six baits. Its three IntAct 'methods' are three Y2H variants, and its host_cells value is Saccharomyces cerevisiae, so there is no human-cell observation. The credible-and-undecided count is therefore 0 of 135, not 1.
What this evidence cannot show
The four thresholds were chosen before seeing which genes they select but were not swept; a different miscore cut gives a different shortlist. IntAct's miscore is a heterogeneous confidence metric across methods and is not comparable between an ITC measurement and a BioID row. 'Focused vs atlas' uses q160's heuristic of >=15 distinct prey-only partners across the six baits.
Evidence that argues against this (3)
- The filter that produced this shortlist had three defects, all found by checking one row: it inherited an atlas heuristic scoped to six baits rather than to publication scale; it counted assay sub-variants as independent methods; and it never tested whether an interaction was observed in a human cell. A filter with those properties would admit any genome-scale yeast screen as focused human evidence, and the 5-gene shortlist should be read as an upper bound produced by a lenient instrument.
- Excluding histones and ribosomal proteins as sticky is a prior, not a measurement, and it is applied here to a set where H3C1 has the HIGHEST miscore of all 135 (0.68) with 3 focused PMIDs and 5 methods including ITC, NMR and X-ray. If any histone contact in that list is a real structural interaction, this filter deletes the best-evidenced row in the table.
- Reducing 135 to 5 makes the remaining gap look trivially small, which is a convenient conclusion for a coordinator who would otherwise owe 135 entries. The 130 removed are unranked, not disproven, and a future method with better sensitivity could promote many of them.
- E3strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.journal-article
Skucha 2018's Supplementary Table 5 MLL-AF4 column lists 291 interactors; 110 of them are reported in MLL-AF4 and no other of the seven fusion baits. Of those 110, 1 has its own gene- entry (KMT2D), 15 are mentioned anywhere in the corpus, 3 carry a decision entry, and 95 are never mentioned. By contrast all 42 partners shared by all seven baits are held. Six AF4-specific partners have p<0.05, no entry and no decision: RGPD8 (0.0147), ACADM (0.0208), RPA2 (0.0221), LEF1 (0.0261), GLUD1 (0.0399), CEP112 (0.0436). The column map was verified against the sheet's own labels and the parse asserted at 291 rows with MEN1 present.
What this evidence cannot show
The table's selection is RANK-based, not thresholded - this repo's own skucha analysis records the largest p-value per bait running to 0.175 for MLL-AF4 - so a p of 0.04 here is not a test that passed. Spectral counts for the six open targets are low (2|1 to 10|8). 'AF4-specific' means absent from the other six baits in this one experiment, not specific to KMT2A::AFF1 in biology.
Evidence that argues against this (3)
- Much of the AF4-specific tail is very likely contamination, and the composition says so: ACADM, GLUD1, ECHS1 and PDHA1 are mitochondrial metabolic enzymes, a recognisable AP-MS background signature. If most of the 110 are noise then 0.9% coverage is roughly correct behaviour and this entry's framing overstates the omission - the defect would be that the reason was never recorded, not that the entries are missing.
- The contrast between 42/42 and 1/110 could be driven entirely by evidence strength rather than by any bias toward the pan-fusion core: partners in all seven baits have seven independent observations and much lower p-values, so writing them first is correct prioritisation, not drift. This entry asserts a mechanism for the pattern that it does not test.
- A single experiment defines 'AF4-specific' here. Skucha used one cell system and one set of constructs; absence from the other six baits could reflect expression, tagging or depth differences between the baits rather than biology, and no second AP-MS dataset exists to check it against.
both find an interactome gap figure that measured the wrong thing
CD226 passes the credibility filter and is already declined - the example that exposed the id-vs-content confusion
- related to →KMT2D (MLL2/MLL4) - the gene named in this graph's menin-resistance entry, written here for the first time
the single AF4-specific Skucha partner of 110 that has its own entry
takes q178's credible shortlist from 5 to 0: every one rests on a misclassified atlas, closing the IntAct prey gap for a second independent reason
the gap report itself counts as coverage in the join that measures the gap; 9 AF4-specific symbols are mentioned only by entries written about their absence
the peptide-level follow-through on the AF4-specific tail that q178 relocated the gap to
relocates the gap again and then closes it: the AF4-specific tail is 6 defensible prey, 5 already adjudicated
re-measures its coverage figures on the current corpus (21 mentioned / 89 never, recovering 15 / 95 once the gap report's own mentions are excluded) and adds the four mechanisms that produced the inversion