CORRECTED SAME CYCLE: the replication test was UNINFORMATIVE, not negative. Eleven of the twelve proximity panels do not sample the nucleus - they are centrosome, RTK, kinase, ion-channel, CFTR and mitochondrial baits - so KMT2A's 0th percentile in three of them measures panel composition, not promiscuity. In the one transcription-factor panel KMT2A is at the 71.7th and menin is absent, which is DIRECTIONALLY CONSISTENT with the K562 result at n=1
also called the replication test was uninformative not negative · eleven of twelve panels do not sample the nucleus · my own counter-evidence was right
Two earlier cycles argued that one protein is picked up by proximity experiments far less than expected, and this set out to check whether that holds in every other such experiment available. The first answer was that it does not hold - but looking at what those experiments were actually about showed that eleven of the twelve were mapping things outside the cell nucleus, where neither protein belongs, so they cannot speak to the question at all. The one experiment that did look at nuclear proteins points the same way as the original, which means the question is still open rather than settled either way.
The same thing, for a scientist
**CORRECTED WITHIN THE SAME CYCLE — read the first section before the rest.** The verdict below was **does not generalise**. Looking up what the twelve publications actually study shows **eleven of them do not sample the nucleus** (centrosome–cilium, RTK, kinase, CFTR, Kir2.1, mitochondrial ClpP, Hippo, phosphatases, a method paper, a colorectal screen). The three 0th-percentile results for KMT2A come from **centrosome, RTK and kinase panels**. **In the one transcription-factor panel KMT2A is at the 71.7th percentile and menin is absent — the same shape as K562.** The corrected verdict is **UNINFORMATIVE, not negative**: cycles 94 and 95 are neither confirmed nor refuted. --- **Statement.** Cycle 94 found menin reached by **3 of 50** BioID2 baits in the K562 map against KMT2A's **25** and read it as menin being selectively localised. Cycle 95 survived an abundance attack and identified the **head-to-head as the strongest form of the claim**: the two proteins sit at essentially the same K562 abundance (30.2 vs 28.6 nTPM) while differing 25-to-3. Both cycles rest on **one experiment**, and cycle 95 named replication as the test that would make *selectively localised* a property of menin. **Run across every proximity publication reachable.** Rows grouped by publication, then **each publication re-fetched individually by `pubid`** so that per-publication data is complete — twelve publications, **all twelve fetched completely** (reported count equals fetched count for every one). | | scoreable in | result | |---|---:|---| | **MEN1** | **1 of 12** | 0.0th percentile in PMID 29568061; **ABSENT from the other 11** | | **KMT2A** | 5 of 12 | **0.0, 0.0, 0.0**, 61.5, 71.7 | **`W4` — my own control on the control — FAILS.** I declared before running: *"if KMT2A is NOT consistently high across publications, then the K562 contrast was a property of that panel and menin's side of it cannot be trusted either."* **KMT2A is at the 0th percentile of its length-matched peers in three of the five publications where it can be scored** — fewer baits than every size-matched protein in those maps. It is above the 60th in two. **KMT2A is not a promiscuous prey in general; the K562 map is the outlier.** **And menin cannot be scored at all almost anywhere.** It is absent from **11 of 12** publications. By the pre-registered `W1` clause, absence is **not** scored as zero — a protein missing from a map may be maximally under-reached or may simply not have been captured by a shallow or differently-targeted panel. So the replication attempt yields **one** independent data point (0.0th percentile, which is maximal under-labelling) and eleven uninformative absences. **Verdict: the claim is NOT established, and cycles 94 and 95 must be narrowed to the K562 map wherever they are quoted.** What survives is narrower and still worth something: in the one other map where menin appears, it is at the bottom of its size-matched peers; in the K562 map the same-abundance head-to-head remains a fact about that experiment.
CORRECTION, same cycle: I pooled panels that cannot address the question, and my own counter-evidence said so before I checked
This entry first concluded W4 failed and the contrast does not generalise, because KMT2A sits at the 0th percentile of length-matched peers in three of five scoreable publications. Its own counter-evidence flagged the risk: "KMT2A at the 0th percentile in three maps... may indicate those maps are cytoplasmic or organelle-focused rather than that KMT2A is non-promiscuous. I did not check what any of the twelve publications were actually studying, which is one lookup per PMID."
One lookup per PMID. The counter-evidence was right.
| PMID | what the panel maps | baits | KMT2A |
|---|---|---|---|
| 26638075 | centrosome–cilium interface | 59 | 0.0 |
| 35384245 | receptor tyrosine kinases | 53 | 0.0 |
| 32707033 | kinase interaction network | 322 | 0.0 |
| 36012204 | CFTR | 2 | absent |
| 32541000 | Kir2.1 ion channel | 1 | absent |
| 26058080 | mitochondrial protease ClpP | 1 | absent |
| 24255178 | Hippo pathway | 27 | absent |
| 28330616 | protein phosphatases | 61 | absent |
| 29568061 | MAC-tag method paper | 37 | absent |
| 30108113 | colorectal point mutations | 2 | absent |
| 36931259 | 14-3-3 proteins | 133 | 61.5 |
| 35140242 | human transcription factors | 115 | 71.7 |
The three zeroes are centrosome, RTK and kinase panels. A nuclear chromatin scaffold being reached by fewer baits than its size-matched peers in a cilium map is not evidence that it is non-promiscuous — it is evidence that the panel does not go where it lives. W4 did not fail; I applied it without conditioning on whether the panel samples the nucleus.
And the one panel that does is directionally consistent with the K562 result. In Human transcription factor protein interaction networks (115 baits) KMT2A is at the 71.7th percentile and menin is absent — high KMT2A, unreached menin, which is the shape cycles 94 and 95 described. At n = 1 that is not replication either.
So the corrected verdict is UNINFORMATIVE, not negative. Eleven of twelve panels cannot address the question. Cycles 94 and 95 are neither confirmed nor refuted by this test, and the narrowing applied to the prior entry should be read as "not yet replicated" rather than "failed replication".
This is the third time in four cycles that the fix was to condition on the right stratum — length-matching, then abundance-matching, now panel composition — and the first time I failed to apply my own lesson before publishing a verdict.
The instrument truncated silently, and the publication list is therefore incomplete
The global query detmethod:"MI:1314" AND species:9606 reports 64,342 rows and PSICQUIC stopped returning them at 36,500 — 57%. The pagination ends without an error; a short page simply arrives and the loop ends normally. My first run scored a verdict on that truncated pull and returned R1 — IT REPLICATES, which was wrong three ways at once: the data was 57% of the record, the majority test read len(low) > len(scored)/2 with scored = 1 so a majority of one passed, and W4 was reported as a note rather than gating the verdict.
Fixing it required not a bigger page but a different unit. Per-publication queries by pubid do not truncate — all twelve returned their full reported row count. The publication list, however, was enumerated from the truncated global pull, so there are proximity publications this analysis never saw. That is a real bound on the result and the reason the verdict is not established rather than refuted.
This corroborates an instrument trap Q129 filed the same day from the other direction: findInteractions/pubid:<PMID> returns a clean totalElements: 0 on IntAct REST while the bare PMID as free text returns the real count, because pubid: is PSICQUIC MIQL and not REST syntax. Two silent false answers from the same database in one cycle, one a false zero and one a false completion. Neither raised an error.
| question | does menin's proximity under-labelling replicate beyond the K562 50-bait map? |
|---|---|
| verdict | NOT ESTABLISHED. Narrow cycles 94 and 95 to the K562 map wherever quoted. |
| publications enumerated | 12 |
| publications fetched completely | 12 |
| W4 control on the control | FAILS. I declared that if KMT2A is not consistently high the K562 contrast was a property of that panel. It is at the 0th percentile in 3 of 5 scoreable publications. |
| W1 absence not scored | declared before running; the 11 menin absences carry no weight in either direction, which as the counter-evidence notes also discards the observations most consistent with the hypothesis |
| corroborates q129 the same day | Q129 filed that findInteractions/pubid:<PMID> returns a clean totalElements:0 on IntAct REST while the bare PMID as free text returns the real count. Two silent false answers from one database in one cycle - a false zero and a false completion - neither raising an error. |
| what survives | in the one other map where menin appears it is at the bottom of its size-matched peers; in the K562 map the same-abundance head-to-head remains a fact about that experiment |
| absence class | measured-not-analysed |
| absence class note | All twelve depositions were public and queryable throughout; what had never been analysed is whether the K562 observation appears in any of them. |
MEN1
{
"scoreable_in": 1,
"absent_from": 11,
"the_one_score": {
"pmid": "29568061",
"percentile": 0,
"n_baits_in_study": 37
}
}KMT2A
{
"scoreable_in": 5,
"percentiles": {
"26638075": 0,
"32707033": 0,
"35140242": 71.7,
"35384245": 0,
"36931259": 61.5
}
}my first run returned the opposite verdict and was wrong three ways
{
"truncated_data": "36,500 of 64,342 rows, pagination ended without an error",
"majority_of_one": "len(low) > len(scored)/2 with scored=1 passed - the eighth criterion in this project sound in form and wrong in substance",
"W4_not_gating": "the control on the control was printed as a note instead of blocking the verdict"
}psicquic truncation
{
"reported": 64342,
"returned": 36500,
"pct": 56.7,
"fix": "per-publication pubid queries do not truncate; all 12 returned their full reported count",
"residual": "the publication LIST was enumerated from the truncated pull, so unseen publications exist"
}CORRECTION same cycle
{
"what_changed": "verdict moves from 'does not generalise' to UNINFORMATIVE",
"why": "11 of 12 panels do not sample the nucleus - centrosome-cilium, RTK, kinases, CFTR, Kir2.1, ClpP, Hippo, phosphatases, a method paper and a colorectal mutation screen",
"the_three_zeroes_are": {
"26638075": "centrosome-cilium",
"32707033": "kinase network",
"35384245": "receptor tyrosine kinases"
},
"the_one_nuclear_panel": {
"pmid": "35140242",
"subject": "human transcription factor interaction networks",
"baits": 115,
"KMT2A_pct": 71.7,
"MEN1": "absent",
"reading": "high KMT2A, unreached menin - directionally CONSISTENT with the K562 result, at n=1"
},
"W4_did_not_fail": "I applied it without conditioning on whether the panel samples the nucleus",
"status_of_cycles_94_95": "neither confirmed nor refuted; the narrowing on the prior entry should read 'not yet replicated' rather than 'failed replication'",
"source": "analyses/does-menin-under-labelling-replicate/publication_subjects.json, PubMed esummary 2026-09-23"
}- ANSWERED IN THIS ENTRY, and it changed the verdict: eleven of twelve panels do not sample the nucleus. The question that replaces it: are there proximity datasets with NUCLEAR bait panels beyond the two here? That, not the count of all proximity papers, is the denominator this test needed - and the 57% PSICQUIC truncation means the list was never complete anyway.
- Enumerate the proximity publications properly. The list came from a pull that returned 57% of the record, so the denominator of this test is unknown. IntAct REST free-text counts differ from PSICQUIC's field-scoped ones (72,314 vs 64,342), so the right query has to be established before the right list can be.
- Is W1 too strict to let this hypothesis be tested at all? Refusing to score absence is correct for a shallow dataset and discards exactly the observations the hypothesis predicts. A depth-aware version - score absence as zero only where a protein of that abundance would be expected - would be testable rather than merely safe.
- The K562 head-to-head still stands within its experiment. Is there any other map containing BOTH menin and KMT2A? None of the twelve does, which is why no second head-to-head exists, and that absence is itself worth stating to anyone who wants to repeat the comparison.
- E1established confidence 5 of 5: Multiple independent lines of evidence; treated as settled in the field.analysis
Twelve human proximity-labelling publications were enumerated from IntAct and each re-fetched individually by pubid; for all twelve the fetched row count equals the count the API reports. Scoring baits-per-prey against length-matched peers within each publication, MEN1 is scoreable in 1 of 12 (PMID 29568061, 0.0th percentile) and absent from 11; KMT2A is scoreable in 5, at the 0.0th percentile in PMID 26638075, 35384245 and 32707033, the 61.5th in 36931259 and the 71.7th in 35140242. The global query detmethod:"MI:1314" AND species:9606 reports 64,342 rows and PSICQUIC returns 36,500 before pagination ends without error.
analyses/does-menin-under-labelling-replicate/run.py, results.json and per_publication.json, this repository; IntAct PSICQUIC queried 2026-09-23; length join from the UniProt reviewed human proteome · Oncofusion knowledge graph, coordinator · internal analysis, reproducible from the repository 2026What this evidence cannot show
THE PUBLICATION LIST IS INCOMPLETE. It was enumerated from a global pull that returned 57% of the record, so proximity publications exist that this analysis never saw, and the twelve are not a random sample of them. The twelve differ enormously in design - bait panels range from 1 to 322 - so a percentile in a 1-bait study and a 322-bait study are not the same quantity, and no attempt was made to weight or stratify by panel size. ABSENCE IS NOT SCORED, by a clause declared before running, so the eleven absences carry no evidential weight in either direction and the headline rests on five KMT2A observations and one for menin. Length-matching only; no abundance control was applied per publication because the cell lines differ and a K562 abundance table does not transfer.
Evidence that argues against this (3)
- Reporting this as 'does not generalise' may be too strong given what the absences could mean. If menin is genuinely under-reached by proximity labelling, then being ABSENT from 11 of 12 maps is exactly what the hypothesis predicts - and my W1 clause, which forbids scoring absence as zero, systematically discards the observations most consistent with the claim it is testing. The clause is right to prevent a shallow dataset counting as evidence, but it also makes the hypothesis nearly untestable by this route, and I did not notice that when writing it.
- KMT2A at the 0th percentile in three maps is a stronger anomaly than this entry treats it as. A very large, abundant chromatin protein being reached by FEWER baits than every size-matched peer is surprising in the opposite direction, and it may indicate those maps are cytoplasmic or organelle-focused rather than that KMT2A is non-promiscuous. I did not check what any of the twelve publications were actually studying, which is one lookup per PMID and would change how much weight the W4 failure deserves.
- This entry retracts two cycles of my own work and the retraction is cheap for me to make, which should raise rather than lower suspicion about the original claims. The pattern across cycles 94, 95 and now 96 is that I built an increasingly elaborate defence of a single-experiment observation - size-matching, abundance-matching, a head-to-head - and only tested generalisation last. The correct order would have been to ask whether the observation exists anywhere else before spending two cycles controlling confounds within one dataset.
narrows that entry to the K562 map: across twelve other proximity publications KMT2A is at the 0th percentile in three and menin is absent from eleven, so its W4 control failed
menin is absent from eleven of twelve proximity publications and scoreable in one, where it sits at the 0th percentile
- related to →KMT2A / MLL1 — the gene that breaks
KMT2A is not a promiscuous proximity prey in general - 0th percentile in three of five scoreable publications; the K562 map is the outlier