DepMap co-dependency cannot rank the 97 unwritten interactome partners either, pan-cancer OR restricted to leukaemia lines: mean r 0.008 rising to 0.023 under restriction while the leukaemia-matched null rises with it, giving 1.08x the chance-expected count. Recomputing the null on the restricted lines is the only thing that stopped this being reported as a discovery, and my first pass criterion called a below-chance result a PASS

The first version of U1 read: "PASSES U1 — 4 of 93 ranked partners exceed the 95th percentile of random pairs (0.1046)." That is arithmetically true and substantively backwards. The 95th percentile is defined so that 5% of draws exceed it. With 93 partners, chance predicts 4.65. Four is below that. The criterion compared a count to a threshold and never to the count the threshold implies, so a result at 0.86× chance was reported as evidence of signal.

This is the seventh criterion in this project sound in form and useless in substance, after a reagent that could not distinguish its subject, a population never stated, a measure saturated at its sample size, a control not independent of the treatment, an or that let a dead clause decide, and a clause written per-cohort and evaluated per-pool. The new variant: a threshold test with no baseline count.

The criterion now requires three things together — ≥2× the chance-expected count above the null p95, an observed spread wider than the null, and a top correlation clearing the null p99 — and it reports which of the three failed. On this data all three fail.

What caught it was writing the null down before running. The informativeness clause did its job even though the pass rule inside it was wrong: the null's own numbers were printed next to the observed ones, and the comparison was visible on the page.

The null is drawn from the same 101-gene pool, not the genome — and that makes it too hard, not too easy. Random pairs are sampled from the 97 partners plus 6 baits, a pool that contains SRSF4/SRSF5/SRSF6, YWHAB/YWHAE/YWHAH and several ribosomal proteins. Paralogues and complex co-members are genuinely co-dependent, which is why the null's maximum reaches 0.5834 while no real partner-bait pair exceeds 0.155. So the null tail is inflated and the U1 bar is conservative against my own conclusion. The absolute argument above is what makes the negative safe: mean 0.008 is near zero on any reading. A genome-wide null would be the correct instrument and is not what was run.

Correlation is computed against the bait that NAMED the partner, not against the fusion. KMT2A::AFF1 has no DepMap profile — a fusion is not a gene in that matrix — so this measures co-dependency with wild-type baits. A partner specific to the fusion's neomorphic complex would be invisible here, and that is the population the PI's priority is actually about.

RL3_RL3L_HUMAN does not resolve in HGNC, confirming cycle 91's suspicion that it is a parser artefact sitting inside the instrument's 217 denominator. TCTE1 does not resolve either — the resolver deliberately refuses to accept an alias that HGNC maps to a different current symbol, which is how a plausible id lands on the wrong gene. So at least 2 of the 97 are not genes as written, and the 217 denominator is soft by at least that much.

This entry's own counter-evidence said the pan-cancer negative may be an artefact of not conditioning, and named the leukaemia-restricted version as the correct follow-up. Run before shipping, per the cycle-90 rule.

Restricting to the 227 of 1,208 lines (18.8%) classified leukaemia/lymphoid, everything gets bigger:

The observed mean rises 2.9× and the maximum 1.8×. That is what a discovery looks like — and it is entirely noise. The null rises with it, because a correlation over ~227 lines is simply less stable than one over 1,208. Scored properly: 5 of 93 above the leukaemia null's 95th percentile against 4.7 expected by chance — 1.08×. The observed spread is still narrower than the null. The top value clears the null 99th percentile by 0.0026, and with 93 partners chance predicts 0.93 such values.

Had I reused the pan-cancer null — the obvious shortcut, since it was already computed — ZC3H11A at 0.2754 would have sat far above its 0.1793 99th percentile and I would have reported a leukaemia-specific co-dependency. The V1 clause that forced the null to be recomputed on the same restricted lines is the only thing standing between this entry and a false positive. It was written down before the pull, for exactly this reason, and it is the single most transferable thing in this analysis: when you restrict a population, the null must be restricted the same way, or the restriction manufactures the signal.

OGT — flagged in cycle 91 as the one absent partner whose absence looked surprising — comes in at r = 0.138 with KMT2A in leukaemia lines, inside the noise and common-essential. That does not make it uninteresting; it makes DepMap silent about it.

This entry calls its population "the 97 unwritten interactome partners". The count is right and the word is misleading. They lack an entry of their own — true, and that is the PI's granularity rule — but the instrument's absent flag was computed from id, title and aliases only, so it could not tell "has no entry" from "nobody here has ever discussed it".

Re-run with the whole node scanned: of 217 partners 80 are discussed somewhere without an entry of their own, and only 34 have never been discussed at all. So a substantial share of the 97 ranked here already have reasoning about them in the corpus, sometimes an explicit adjudication.

This does not change the result. The co-dependency ranking failed on its own terms, pan-cancer and leukaemia-restricted alike, and that is a statement about DepMap correlations, not about corpus coverage. What it changes is the framing of why the ranking was wanted — the remaining work is smaller and partly already decided, so 'no principled order for 97 unwritten partners' overstates both the number and their unwrittenness.

This entry's caveat said the null was drawn from the same 101-gene pool, that paralogues and complex co-members inflated its tail, and that "a genome-wide null would be the correct instrument and is not what was run." Run: 400 genes sampled at random from HGNC's 19,254 approved protein-coding symbols with an Ensembl id, excluding this analysis's own 99 genes, 376 with DepMap profiles, 2,960 random pairs.

G2 — the negative stands. 4 of 93 above the genome-wide p95 against 4.7 expected — 0.86×, the same figure as before — observed spread still narrower, top correlation still short of the null p99.

And the caveat was half right in a way worth separating. Its observation was correct: the pool's extreme tail was contaminated, and removing paralogues halves the maximum from 0.5834 to 0.2986. Its inference was wrong. I argued the inflated tail made the bar too high and therefore made the negative conservative — but p95 moved the wrong way (0.1046 → 0.1087) and p99 barely moved (0.1793 → 0.1754). The decision thresholds are set by the body of the distribution; a couple of paralogue pairs move the maximum and touch nothing that any verdict depends on.

The transferable point: a contaminated maximum is not a contaminated percentile. I reasoned from the most visible number in the null — its extreme value — to the thresholds, and the thresholds do not care. The right check was to compare percentiles, which costs one line and was not done until now.

One thing the genome-wide null does confirm: excluding common-essential pairs changes almost nothing (p95 0.1087 → 0.1098), so essentiality was not driving the null either. Both contamination worries were real in kind and negligible in size.