ASH2L and RBBP5 rank 17,248 and 14,622 of 17,728 for KMT2A-selectivity, which is what this graph's own SET-domain-is-dispensable thread predicts - the audit found agreement where it looked for contradiction
also called dependency claim audit · WRAD anti-selectivity · scan versus claims · SET domain dispensable
This project has argued from mouse genetics that the enzyme part of KMT2A does not matter for the leukaemia, and that the protein complex whose job is to switch that enzyme on is therefore beside the point. A large independent gene-deletion screen agrees: the two complex members turn out to be among the least specifically required genes measured.
The same thing, for a scientist
Backfilling KMT2A-selectivity into 262 gene entries made a question askable for the first time: **do this graph's dependency claims agree with this graph's own genome-wide scan?** **36 disease-scoped requirement claims** were checkable. **16 (44%) are for genes sitting in the anti-selective half** of the scan, and **9 are deep in it (rank > 14,000 of 17,728)**. Reading the deep cases is what matters, and two of the four flagged are **the same claim in two entries** — the Thiel result that menin recruits wild-type MLL and MLL-AF9 to HOX loci — attached to **ASH2L (rank 17,248)** and **RBBP5 (rank 14,622)**. Those are WRAD subunits. **This graph already argues, from mouse genetics, that the SET domain of MLL1 — the catalytic activity WRAD exists to stimulate — is dispensable for maintaining haematopoietic stem cells and for MLL-AF9 leukaemogenesis.** `subtraction-kmt2a-set-domain` puts it bluntly: *"the lost module whose loss demonstrably does not matter"*. **So a genome-wide CRISPR screen and a mouse knock-in agree, from completely different directions, that the WRAD/SET arm is not where this leukaemia's dependency lives.** The audit was built to find claims contradicted by the project's own data and instead found its sharpest case corroborating an existing conclusion. Note RBBP5's pan-essentiality is **−1.105** — it is broadly required, just not *selectively*, which is exactly the distinction the entry's `how_to_read` note exists to enforce.
The four flagged claims, read
| gene | rank /17,728 | difference | pan-essentiality | verdict |
|---|---|---|---|---|
| ASH2L | 17,248 | +0.125 | −0.333 | corroborates the WRAD thread |
| ASF1B | 15,744 | +0.065 | −0.292 | false positive of my detector |
| RELA | 15,425 | +0.059 | −0.406 | genuine unaddressed tension |
| RBBP5 | 14,622 | +0.049 | −1.105 | corroborates the WRAD thread |
ASF1B is my detector's fault, not the entry's. Its claim states the DepMap numbers itself — "−0.183 in 11 KMT2A-rearranged lines against −0.248 in 43 other leukaemia lines" — so the entry engages with the scan directly. My acknowledgement test looked for the word anti-selective or a four-digit rank, and that entry uses neither. Over-reporting.
RELA is the one worth a beat's attention. A strong claim that RELA occupies the promoters of crucial MLL target genes and sustains MLL-dependent leukaemia, for a gene the project's own scan puts at 15,425 with a positive difference. Q64 independently found RELA in the anti-selective half on a different gene set. Nothing in the entry reconciles the two, and the reconciliation may well be the JUND one — a 21-day proliferation screen cannot see a phenotype that needs a stimulus — but it has not been written.
Why 44% is not alarming, and what would be
Sixteen of thirty-six disease-scoped requirement claims are for genes in the anti-selective half. That sounds bad and mostly is not, for a reason this project has already established twice:
A 21-day proliferation screen measures fitness. It is structurally blind to a differentiation-block phenotype, to a chemoresistance phenotype that only appears under drug, to a requirement confined to a subset of lines the KMT2A arm pools away, and to anything needing an intact niche. Q64 made exactly this argument for JUND and was right — JUND's resistance route only opens under menin inhibition, so a dropout screen run without the drug is the wrong instrument.
What would be a real defect is narrower: an entry asserting a KMT2A-selective requirement, citing only screen-type evidence, and not mentioning that the project's own scan disagrees. On this pass that is one entry, gene-rela — and even there the cited evidence is occupancy, not a screen.
The number to watch is not 44% but whether it moves. It is now measurable every cycle, which it was not before today.
My first run reported an empty queue because it was reading my own boilerplate
The first execution of this script reported zero entries in the reading queue, and the reason is worth recording.
The backfill I ran an hour earlier wrote a how_to_read note into 262 entries containing the sentence "A rank near 17,728 is ANTI-selective". My acknowledgement test searched the whole node for the phrase anti-selective. Every entry passed, because I had inserted the phrase into every entry myself.
The test now strips the depmap_kmt2a_selectivity block before looking, and the queue goes from 0 to 4. This is the same failure as everything else found this month — a check satisfied by something other than the thing it was meant to detect — with the distinguishing feature that I manufactured the contamination in the same session, and the clean result it produced was the one I would have preferred.
| n requirement claims | 36 |
|---|---|
| n anti selective half | 16 |
| PCT anti selective half | 44 |
| n deeply anti rank gt 14000 | 9 |
| n flagged after acknowledgement test | 4 |
| first run queue was empty because | the acknowledgement test matched boilerplate the coordinator's own backfill had written into 262 entries an hour earlier |
flagged — the numbers behind the chart
{
"ASH2L": 17248,
"ASF1B": 15744,
"RELA": 15425,
"RBBP5": 14622
}false positives on reading
[ "gene-asf1b" ]
corroborations
[ "gene-ash2l", "gene-rbbp5" ]
genuine unaddressed tension
[ "gene-rela" ]
- Do not read 44% as a defect rate. Most of those claims concern phenotypes a 21-day proliferation screen structurally cannot see, and this project has already accepted that reasoning for JUND on good grounds.
- Do not quote ASH2L or RBBP5's rank without their pan-essentiality. RBBP5 sits at -1.105 and is broadly required; it is unselective, not unimportant, and conflating the two would be a serious misreading in a target-selection context.
- Do not treat the scan agreeing with an entry as independent corroboration. The same numbers were written into 262 entries at once by one join, so an entry agreeing with the scan is usually the same measurement read twice.
- gene-rela is the one genuine tension: a strong occupancy-based requirement claim for a gene the scan puts at 15,425. Does the JUND reconciliation apply - a phenotype the screen cannot see - or is the claim over-scoped to the disease?
- Do the WRAD entries that argue for a non-catalytic structural role need to reconcile with these ranks? The scan is consistent with SET-domain dispensability and silent on structural requirement, and no entry currently says which it supports.
- Should the acknowledgement test be replaced by something semantic? gene-asf1b engages with the scan by quoting its numbers and was still flagged.
- Does the 44% figure move as beats land? It is now measurable every cycle and has no baseline before today.
- E1moderate confidence 3 of 5: One strong study, or several smaller consistent ones.analysis
Of 36 disease-scoped requirement claims in gene entries carrying a KMT2A-selectivity rank, 16 (44%) are for genes in the anti-selective half of the project's own genome-wide scan and 9 rank beyond 14,000 of 17,728; after excluding entries that already engage with the scan, four claims remain flagged.
dependency-claims-vs-own-scan/run.py: every disease-scoped requirement claim checked against the project's own genome-wide DepMap selectivity scan · Oncofusion knowledge graph (internal analysis) 2026What this evidence cannot show
The requirement detector is a regular expression over claim text: it will miss a requirement stated unusually and will catch a sentence that merely reports someone else's requirement claim. **The acknowledgement test over-reports** - gene-asf1b states the DepMap means in its own claim text and was still flagged, because the test looks for the word anti-selective or a four-digit rank. All selectivity values are joined from one cached scan whose KMT2A-rearranged arm is a small number of lines from nine individuals, so the rank is a rank within that scan rather than a biological constant. Because the same block was written into 262 entries at once, agreement between an entry and the scan is not independent corroboration - it is the same numbers read twice.
Evidence that argues against this (1)
- A DISAGREEMENT BETWEEN A CLAIM AND THIS SCAN IS USUALLY NOT AN ERROR, and treating 44% as a defect rate would be wrong. A 21-day proliferation screen cannot see a differentiation block, a drug-conditional phenotype, a subset-restricted requirement, or anything needing a niche. This project already accepted that reasoning for JUND on good grounds. The audit's value is in naming the entries that should say so, not in scoring them. Oncofusion knowledge graph (internal analysis) 2026
- E2moderate confidence 3 of 5: One strong study, or several smaller consistent ones.analysis
ASH2L ranks 17,248 and RBBP5 ranks 14,622 of 17,728 for KMT2A-selectivity, both with positive differences, which independently corroborates this project's existing conclusion from mouse genetics that the SET domain of MLL1 - the catalytic activity the WRAD complex exists to stimulate - is dispensable for maintaining haematopoietic stem cells and for MLL-AF9 leukaemogenesis.
subtraction-kmt2a-set-domain: the SET domain (aa 3829-3945), the lost module whose loss demonstrably does not matter · Oncofusion knowledge graph (internal entry) 2026What this evidence cannot show
**Corroboration is weaker than it looks in one specific way**: WRAD subunits have functions beyond stimulating the MLL1 SET domain, including in other COMPASS complexes, so their anti-selectivity is consistent with the SET-domain argument without being predicted only by it. RBBP5's pan-essentiality of -1.105 means it is broadly required and merely not selectively so - an entry quoting the rank without the pan-essentiality would badly misrepresent it. The two results are also not fully independent of each other in kind: both are loss-of-function measurements, differing in system and scale rather than in logic.
Evidence that argues against this (1)
- THE WRAD THREAD HAS A LIVE OPPOSING ARGUMENT IN THIS GRAPH and the scan does not settle it. Entries on the WRAD interface record that ASH2L and DPY30 matter in KMT2A-rearranged cells through non-catalytic, structural roles, which a SET-domain-dispensability result does not touch. A gene can be anti-selective in a fitness screen and still be required for the fusion's programme, which is the same objection this analysis raises against reading its own 44% as a defect rate. Oncofusion knowledge graph (internal analysis) 2026
THAT AUDIT'S 44% FIGURE NEEDS THE SAME CORRECTION: 44% of disease-scoped requirement claims being for anti-selective genes is close to the 45.3% corpus base rate, so it too is near the null. Its per-entry readings survive; the aggregate does not.
AN INDEPENDENT LINE REACHING THE SAME PLACE: a genome-wide CRISPR screen puts ASH2L at 17,248 and RBBP5 at 14,622 of 17,728 for KMT2A-selectivity, which is what this entry's mouse-genetics argument for SET-domain dispensability predicts.
- related to →RELA (p65) — menin binds it and represses it, and it is required to keep MLL oncoproteins on chromatin
THE ONE GENUINE TENSION THE AUDIT FOUND: a strong claim that RELA sustains MLL-dependent leukaemia, for a gene the project's own scan ranks 15,425 with a positive difference, and nothing in the entry reconciles them.
THE COMPANION AUDIT, ON THE SAME BACKFILLED DATA: that one asked whether paralogue redundancy explains anti-selectivity and found it does not; this one asks what the anti-selective claims are.