'Four consecutive beats found 40-50% of their genes anti-selective' was the null expectation, not a pattern - and the real test shows the curated set IS enriched for selectivity, p = 0.0016
also called anti-selective base rate · curated set enrichment · DepMap base rate · retraction of the anti-selectivity pattern
Across four rounds of work this project kept noticing that about half the genes it examined were *less* needed in this leukaemia than in others, and treated that as a meaningful signal. About half of all genes are in that group by definition, so the observation carried no information. Testing it properly shows the opposite: the genes this project has chosen to write about really are more selectively required than genes picked at random.
The same thing, for a scientist
**Retraction of a claim this coordinator repeated in four king logs and wrote into the ledger.** Q63 (4/6), Q64 (6/16), Q66 (6/14) and Q69 (7/16) each reported ~40-50% of their gene set sitting in the DepMap 'anti-selective half', and I reported the repetition as a pattern. **'Anti-selective half' means rank above the median of 17,728. Roughly 50% of ALL genes are there by construction.** Measured over the 265 gene entries now carrying a selectivity rank, the graph's own base rate is **45.3%**. Every one of the four beat figures is at or below it. The observation was the null expectation. **Worse, I used it to argue biology.** The cycle-39 log reads: *"the DepMap arm keeps firing in the same direction... if these states were fusion-driven, the genes setting them should be MORE required where the fusion is, not less."* A base-rate figure cannot support that, and the inference is withdrawn. **The test that does carry information gives the opposite sign.** These genes are not a random draw — they are here because somebody judged them relevant to KMT2A::AFF1 — so the question is whether the curated set is enriched for selectivity against the genome: | | | |---|---:| | genes with a rank | 265 | | observed mean rank | **7,960** | | observed median rank | 7,404 | | null expectation | 8,864.5 | | permutation p (20,000 draws) | **0.0016** | **The curated set is significantly, if modestly, enriched.** A shift of roughly 900 ranks in the mean out of 17,728. So the curation carries real signal — and the leg of the lineage-over-fusion argument that rested on anti-selectivity counts does not.
What survives of the lineage-over-fusion conclusion, and what does not
Cycle 39 concluded that the lineage explanation is currently better supported than the fusion one. That conclusion had two legs and only one of them was load-bearing.
The leg that survives is Q69's, and it is independent of DepMap entirely: every cell state this project has documented exists in a mouse with no fusion. HDAC7 blocks B development, MEF2C sets lymphoid fate, KAT5 maintains HSCs, HEXIM1 blocks differentiation in keratinocytes. That is an argument from published loss-of-function genetics and it is untouched by anything here.
The leg that fails is mine. I added "and the DepMap arm keeps firing in the same direction" as corroboration. It was not firing at all — it was reporting a base rate four times. And the enrichment test points weakly the other way, so if anything the DepMap evidence is a mild argument that the curated genes are fusion-relevant.
Net effect on the conclusion: weaker but not overturned. One of two supporting arguments withdrawn, and the remaining one is the stronger of the two. The decisive experiment named there is unchanged and still the right one — a knock-in removing a state-setting modification, scored in normal haematopoiesis and fusion leukaemia in the same animal, as was done for Mef2c S222A/S222A.
Why this took four cycles to notice, and the rule it earns
Each beat reported its own figure honestly and none of them was wrong. 6 of 16 is a true count. The error was entirely in aggregation: I saw the same number recur across independent gene sets and read consistency as evidence, when consistency around a base rate is exactly what a base rate produces.
It is the same failure as cycle 36's sampling artefact, inverted. There, a figure measured on the recent tail was read as a corpus property. Here, a figure that is the corpus property was read as a finding about particular gene sets. Both times the missing step was writing down what the number would be if nothing were going on.
The rule: a proportion is not a result until its null is stated. Before reporting that k of n items fall in a category, write down what fraction of everything falls in that category. If the category is defined as 'above the median', the answer is 50% and the observation carries no information whatever. This project already owns the machinery — six permutation and null analyses exist in analyses/ — and none was applied here, because the figure did not look like the kind of claim that needs one.
| n with rank | 265 |
|---|---|
| anti selective half | 120 |
| base rate PCT | 45.3 |
| observed mean rank | 7960 |
| observed median rank | 7404 |
| null expectation | 8864.5 |
| permutation p | 0.0016 |
| n permutations | 20000 |
| clearly selective PCT | 9.8 |
| deeply anti rank gt 14000 PCT | 30.9 |
beat figures retracted
{
"Q63": "4/6",
"Q64": "6/16",
"Q66": "6/14",
"Q69": "7/16"
}- Do not cite the per-beat anti-selectivity figures as evidence of anything. ~50% of all genes sit in that half by construction and the graph's own base rate is 45.3%.
- Do not read the enrichment result as showing these genes are fusion-driven. It shows the curated set is not a random draw with respect to selectivity, which is partly a fact about how the graph was assembled.
- Do not treat this as overturning the lineage-over-fusion conclusion. It withdraws one of two supporting arguments; the stronger one - that every documented state exists in a mouse with no fusion - is untouched.
- Would an expression-matched or variance-matched null remove the enrichment? Part of it is plausibly 'these are genes people study', and this project already has six null analyses to copy the pattern from.
- How much of the enrichment is circular? Entries in scan-hits/ exist because a DepMap result flagged them, so the curated set is not independent of the measurement.
- Do the other proportions this project reports have stated nulls? Counter-note template rate, shown_for_fusion NO rate, and the fraction of absences that are bounded are all candidates and none carries one.
- E1strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.analysis
Across the 265 gene entries carrying a DepMap KMT2A-selectivity rank, 120 (45.3%) sit above the median rank of 17,728 - the 'anti-selective half' - so the per-beat figures of 40-50% reported by four consecutive beats are at or below the corpus base rate and carry no information about those gene sets.
curated-set-selectivity-enrichment: is this graph's curated gene set enriched for KMT2A-selectivity relative to the genome, and what is the base rate for the anti-selective half? · Oncofusion knowledge graph (internal analysis) 2026What this evidence cannot show
The base rate over the graph's scored genes is 45.3% rather than exactly 50% because the curated set is itself mildly enriched for selectivity, which is E2 - so the two results are not independent of each other. All ranks come from one cached scan whose KMT2A-rearranged arm is 11 lines from nine individuals; the rank is a position within that scan, not a biological constant. Genes with a floor-compressed or positive-in-both-arms profile are included in the count, and Q69 showed at least two such category errors exist.
Evidence that argues against this (1)
- THE BEATS WERE NOT WRONG AND SHOULD NOT BE READ AS HAVING ERRED. Each reported a true count of its own gene set, and two of them - Q69 explicitly - flagged category errors in the metric rather than accepting it. The aggregation into a 'pattern' was the coordinator's, in the king logs and the ledger. A beat reporting its own number without the corpus base rate is doing its job; supplying the base rate is the coordinator's. Oncofusion knowledge graph (internal analysis) 2026
- E2moderate confidence 3 of 5: One strong study, or several smaller consistent ones.analysis
The 265 curated gene entries have a mean DepMap selectivity rank of 7,960 against a null expectation of 8,864.5 for a uniform draw from 17,728 genes, with a permutation p of 0.0016 over 20,000 draws, so this graph's curated gene set is significantly if modestly enriched for KMT2A-selectivity relative to the genome.
curated-set-selectivity-enrichment: is this graph's curated gene set enriched for KMT2A-selectivity relative to the genome, and what is the base rate for the anti-selective half? · Oncofusion knowledge graph (internal analysis) 2026What this evidence cannot show
**A significant p on a modest shift.** The mean moves about 900 ranks out of 17,728 and the median 1,460; with n=265 that is detectable without being large. The permutation null assumes a uniform draw over all scanned genes, which ignores that expressed and essential genes are over-represented in any curated set for reasons unrelated to this fusion - so part of the enrichment is plausibly 'these are real genes people study' rather than 'these are KMT2A-relevant'. A variance-matched or expression-matched null would be the better instrument and this project already has the pattern for one.
Evidence that argues against this (1)
- ENRICHMENT OF A CURATED SET IS PARTLY CIRCULAR AND CANNOT SETTLE THE LINEAGE-VERSUS-FUSION QUESTION. Several of these entries were written BECAUSE a DepMap result flagged them - the scan-hits directory exists for exactly that - so the set is not independent of the measurement it is being tested against. The honest reading is that curation is not random with respect to selectivity, which is a statement about how this graph was built rather than about the biology of the fusion. Oncofusion knowledge graph (internal analysis) 2026
THE COMPANION BASE-RATE LESSON, on the same data. That one found a proportion flipping with the definition of its predictor; this one found a proportion that was the null all along. Both were fixed by writing down what the number would be if nothing were going on.
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.
SAME LESSON FROM THE OTHER DIRECTION. That analysis found a repeated DepMap proportion was the null; this one finds a repeated DepMap concern was answerable from a file already on disk. Both turn on refusing to read a rank as a result without asking what the number would look like if nothing were going on - there, the median; here, what Chronos does when a gene is already deleted.