The atlas heuristic is scoped to six baits instead of to publication scale, so it misses HuRI, OpenCell and five other proteome-wide screens - 49.5% of rows credited with a 'focused' publication have only atlas support, and all five of q178's credible interactors fail
also called q182 · the focused label is half wrong · atlas heuristic defect
To avoid treating a giant automated screen as if it were a dedicated experiment, this project labels publications as 'atlas' or 'focused'. The rule counts how many partners a paper contributes to six specific proteins - so a screen of fifty thousand interactions that happens to touch only a few of those six is filed as a focused study. Checking every borderline paper against PubMed: seven large screens were mislabelled, including two of the largest human interaction maps ever published. Half of all 'focused' support in the list is actually atlas support, and every one of the five partners that looked worth writing up rests on one.
The same thing, for a scientist
> 🔴 **EXTENDED, same cycle: the floor was low. The heuristic catches 20% of atlases, and the focused label is wrong about TWO THIRDS of the time.** This entry originally checked only the 5–14 partner band and recorded 12 atlases and 49.5% as a floor, because 49 PMIDs in the 1–4 band were never fetched and HuRI had been found there by accident. **All 49 have now been fetched from NCBI esummary.** Counting a paper as an atlas only when its own **title** carries an explicit scale marker — deliberately conservative, so this is still a lower bound — adds **13 more**, including *A directed protein interaction network*, *A High-Density Map for Navigating the Human Polycomb Complexome*, *Physical and functional interactome **atlas** of human receptor tyrosine kinases*, ***Proteome-scale** characterisation of motif-based interactome rewiring* and *The solute carrier superfamily interactome*. > > | | atlases | > |---|---:| > | heuristic finds | **5** | > | q182 first pass (5–14 band, all read) | +7 | > | this pass (1–4 band, title marker only) | +13 | > | **total, lower bound** | **25 of 61 PMIDs (41%)** | > > **The heuristic identifies 5 of 25 — it misses 80%.** And of the 111 census rows credited with a focused publication, only **39** keep one: **72 rows, 64.9%, had support that was entirely atlas**, against the 49.5% reported below. --- 🔴 **The rule measures the wrong thing.** q160 defines an atlas as *a publication supporting ≥15 distinct prey-only partners **across the six baits***. That is a property of the publication's **overlap with this bait list**, not of the publication. A genome-scale screen touching few of these six baits is admitted as focused. 🔴 **Seven are missed — a 140% undercount.** Every PMID in the census supporting 5–14 partners was fetched from PubMed and read; **all six are proteome-scale resources**, and HuRI makes seven [E1]: | PMID | paper | scale | partners *here* | |---|---|---|---:| | 32296183 | Luck, *Nature* 2020 — **HuRI** | ~53,000 binary interactions | **1** | | 25609649 | Li, *Mol Syst Biol* 2015 | 214 purifications, 56 TFs + 37 FOX | 5 | | 36931259 | Segal, *Mol Cell* 2023 | all human 14-3-3 paralogs | 6 | | 35512704 | Mo, *Cell* 2022 | 17,792 interactions, 2.17M data points | 10 | | 35140242 | Göös, *Nat Commun* 2022 | 109 TFs, 8,239 interactions | 13 | | 35271311 | Cho, *Science* 2022 — **OpenCell** | proteome-wide endogenous tagging | 13 | | 40205054 | Schaffer, *Nature* 2025 | >5,100 proteins in U2OS | 13 | **HuRI contributes one partner to these baits and is one of the largest interaction maps ever published.** The rule cannot see scale at all. 🔴 **Half the label is wrong.** Of 212 census rows, **111 are credited with ≥1 focused publication. Once the seven atlases are excluded, 56 keep one — so 55 (49.5%) had support that was entirely atlas** [E2]. 🔴 **And it takes q178's shortlist to zero.** All five 'credible' interactors rest on a misclassified atlas — BRAF on Mo 2022, ARFIP1 on HuRI, and CAMKV, EPB41L1 and CD226 all on Schaffer 2025 [E2]. **Not one interactor in the IntAct prey-only census lacks an entry and has genuinely focused support.** The gap q178 reported as closed-but-for-one is closed outright, for a second and independent reason. 🟠 **This was already known once and not generalised.** q160 itself hit this defect and named an entry after it — `interface-psip1-rbpj-declined-a-transcription-factor-tap-atlas-and-the-heuristic-missed-it`. **A single instance was recorded as a case; nobody asked how many more there were.** That is the same shape as the dead duplicate guard found in cycle 158, where the warning existed in one file while the defective instruction stayed in the one people read.
The full count, and why it is still a floor
All 61 distinct PMIDs in the census have now been looked at: the 5 the heuristic flagged, the 6 in the 5–14 band plus HuRI (fetched and read in full), and the 49 in the 1–4 band (fetched, classified on title only).
25 of 61 are atlases on a conservative reading — 41% of the evidence base. The heuristic finds 5.
Still a floor, for a specific reason: the 1–4 band was classified on explicit title markers alone, so a large screen with a modest title is counted as focused. Interaction proteomics of the HMGA chromatin architectural factors and Proteomic profiling of Myc-associated proteins are both unflagged here and may well be large. The conservative rule was chosen so that the number is defensible without adjudicating 49 papers one at a time — not because it is accurate.
What the unflagged set does contain is the corpus's real evidence base, and it is reassuringly on-topic: the MLL/menin SET1-like complex paper, MLL1–MOF, the AF4–ENL P-TEFb higher-order complex, DotCom, the MLL1 PHD3-bromo/CyP33 cassette, the MLL2/AKAP95 paper, the SIRT1–MLL1 circadian deacetylation paper, and the MLL nucleosome-recognition structure. When this census says an interaction has focused support and it is true, the source is usually one of the papers this project already reasons about.
The correct rule, and why it is easy
Scale is a property of the publication. It is available without judgement from the paper's own abstract or from the size of its deposited dataset — HuRI says ~53,000 interactions in its abstract; OpenCell, Göös, Mo, Segal, Li and Schaffer all state their scale in the first few sentences. A workable definition is a publication reporting interactions for hundreds or more baits, or depositing thousands or more interactions, and it does not depend on which baits a downstream analysis happens to care about.
The failure mode of the current rule is one-directional and therefore worse than noise. It can only ever mislabel an atlas as focused, never the reverse, because a small focused paper cannot contribute ≥15 partners. Every error it makes inflates apparent evidence quality.
Do not raise the threshold to fix this. Lowering it from 15 to, say, 5 would catch the six in the band and still miss HuRI at 1. The parameter is not mis-tuned; the variable is wrong.
What this does and does not invalidate
Does not invalidate the interactions. Being in an atlas is not being wrong. A HuRI Y2H hit is real data; it is weaker evidence than a dedicated co-IP with a hypothesis, which is the entire point of the distinction.
Does invalidate a specific inference: that an interactor supported by a 'focused' PMID has dedicated experimental attention. For 49.5% of the rows so credited, it does not.
Strengthens rather than weakens q178's conclusion. That entry argued the IntAct prey gap is closed as a question of consideration; this shows it is also closed as a question of evidence quality. Two independent routes to the same place.
The corrected count of 12 atlases is itself a LOWER bound. Only PMIDs supporting 5–14 partners were systematically fetched. HuRI was found in the 1–4 band by accident, while chasing a single gene, and 50 PMIDs in that band were never checked. Since HuRI — the largest screen in the set — sat there, the 1–4 band is exactly where more atlases would hide. 49.5% is a floor.
| script | analyses/q182-the-atlas-heuristic-misclassifies-every-large-screen/run.py |
|---|---|
| status | executed |
| corrects | ana-q178-the-interactome-gap-is-not-where-the-standing-priority-says-it-is |
| heuristic | a publication supporting >=15 distinct prey-only partners across the six baits |
| n distinct pmids | 61 |
| atlases found by heuristic | 5 |
| atlases verified | 12 |
| atlases missed | 7 |
| undercount PCT | 140 |
| rows credited with focused | 111 |
| rows keeping focused after correction | 56 |
| rows whose focused was all atlas | 55 |
| PCT focused label wrong | 49.5 |
| q178 shortlist surviving | 0 |
| unchecked band | 50 PMIDs supporting 1-4 partners were never fetched; HuRI was found there by accident |
| verdict | the focused/atlas split is scoped to the wrong variable; half the focused label in this census is atlas support, and no unwritten IntAct interactor has focused evidence |
| atlas total lower bound | 25 |
| distinct pmids | 61 |
| heuristic recall | 0.2 |
| atlases missed full | 20 |
| rows keeping focused full | 39 |
| rows entirely atlas full | 72 |
| PCT focused label wrong full | 64.9 |
missed pmids
[ "25609649", "32296183", "35140242", "35271311", "35512704", "36931259", "40205054" ]
measurements — the numbers behind the chart
[
{
"quantity": "count",
"subject": "genome-scale screens misclassified as focused",
"universe": "61 PMIDs in the IntAct prey-only census",
"value": 7
},
{
"quantity": "share",
"subject": "census rows whose 'focused' support was entirely atlas",
"universe": "111 rows credited with >=1 focused publication",
"value": 0.495
},
{
"quantity": "count",
"subject": "q178 credible interactors surviving the corrected definition",
"universe": "5 that passed the original filter",
"value": 0
},
{
"quantity": "share",
"subject": "atlases among the census's distinct PMIDs (lower bound)",
"universe": "61 distinct PMIDs",
"value": 0.41
},
{
"quantity": "share",
"subject": "recall of q160's atlas heuristic",
"universe": "25 verified atlases",
"value": 0.2
},
{
"quantity": "share",
"subject": "census rows whose focused support was entirely atlas (full check)",
"universe": "111 rows credited with >=1 focused publication",
"value": 0.649
}
]- Adjudicate the 36 still-unflagged PMIDs individually rather than on title markers. Two of them - Interaction proteomics of the HMGA chromatin architectural factors, and Proteomic profiling of Myc-associated proteins - are plausibly large screens with modest titles, so 25 and 64.9% both remain floors.
- Replace the heuristic with a publication-scale field rather than a bait-overlap count, and re-run every analysis that consumed focused_pmids. Only q160 and q178 consume it today, so the blast radius is small - but it will grow if the field survives unfixed.
- Decide whether atlas support should disqualify at all. This entry assumes it does, which the corpus has never validated. A benchmarked screen may beat an uncontrolled co-IP, and the focused/atlas split is a proxy for evidence quality rather than a measurement of it.
- Three of the seven - Mo, Segal, Li - have specific biological theses and could defensibly be kept as focused. Settle that call explicitly, because 49.5% moves a long way on it.
- E1strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.journal-article
q160's census spans 61 distinct PMIDs across 212 rows and flags 5 as atlas-like by the rule 'supports >=15 distinct prey-only partners across the six baits'. All six PMIDs supporting 5-14 partners were fetched from PubMed: 40205054 (Schaffer, Nature 2025;642:222-231, multimodal cell maps, >5,100 proteins), 35271311 (Cho, Science 2022;375:eabi6983, OpenCell, proteome-wide), 35140242 (Goos, Nat Commun 2022;13:766, 109 TFs, 6,703 BioID plus 1,536 AP-MS interactions), 35512704 (Mo, Cell 2022;185:1974-1985, 17,792 interactions, 2,172,864 data points), 36931259 (Segal, Mol Cell 2023;83:974-993, all human 14-3-3 paralogs), 25609649 (Li, Mol Syst Biol 2015;11:775, 214 purifications, 2,156 interactions). All six are proteome-scale resources. With HuRI (32296183, Luck, Nature 2020;580:402-408, ~53,000 interactions, contributing 1 partner here) the heuristic misses 7 of 12, a 140% undercount.
What this evidence cannot show
'Atlas' is a judgement about scale, made here from each paper's own abstract and title rather than from a numeric threshold, so it is defensible but not mechanical. Only the 5-14 band was systematically checked; the 50 PMIDs supporting 1-4 partners were not, and HuRI was found there by accident.
Evidence that argues against this (2)
- Three of the seven - Mo 2022, Segal 2023, Li 2015 - are arguably not pure atlases: each has a specific biological thesis (neomorphic PPIs in cancer, a chaperone role for 14-3-3, chromatin-versus-soluble TF complexes) and is not an untargeted all-by-all map in the way HuRI and OpenCell are. Classifying them as atlas is a defensible call about scale but it is a call, and a reviewer could keep them as focused, which would reduce the 49.5% substantially.
- The 140% undercount figure is sensitive to the denominator being 5. Since the 1-4 band is unchecked and is where HuRI was found, both the numerator and the denominator will move, and the percentage should not be quoted as a stable property of the heuristic.
- E2strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.dataset
Of the 212 census rows, 111 are credited by q160 with at least one focused (non-atlas) PMID. Excluding the seven verified atlases, 56 retain at least one and 55 do not, i.e. 49.5% of rows credited with focused support had support that was entirely atlas. Re-scoring q178's five credible interactors under the corrected definition: BRAF's only focused PMID is 35512704, ARFIP1's is 32296183, and CAMKV, EPB41L1 and CD226 all rest on 40205054. All five are atlases, so 0 of 5 survive.
What this evidence cannot show
The 49.5% depends on classifying all seven as atlases; under the narrower reading in E1's counter-evidence, where Mo, Segal and Li stay focused, the figure falls. The recomputation uses q160's own pmids field and inherits any curation error in it.
Evidence that argues against this (2)
- Reducing the shortlist from 5 to 0 removes the only remaining work this line implied, which is a suspiciously tidy outcome for an analysis run by the person who would otherwise owe those entries. The honest statement is that the evidence for all five is atlas-only, not that the five are non-interactors - CD226, CAMKV and EPB41L1 retain IntAct miscores above 0.5 and multiple records.
- Treating atlas support as disqualifying is itself a choice this corpus has not validated. Large screens have measurable false-discovery rates and HuRI in particular was benchmarked against curated sets; a well-controlled atlas hit may be better evidence than a single uncontrolled co-IP in a focused paper. The focused/atlas split is a proxy for evidence quality, not a measurement of it.
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 prior instance of this defect, recorded as a case and never generalised - this measures how many more there are
same shape: a known defect documented in one file while the defective instrument stayed in use