BRD9 survives the null; CEBPA sits at the median of random genes
also called lineage consistency null · BRD9 validation
Last cycle we found that some genes only look important to this leukaemia because of which cell types were tested. This puts the survivors through a proper statistical check against seven hundred random genes. The menin group comes out as the most extreme in the genome, one new gene holds up, and the two rejected ones score no better than chance.
The same thing, for a scientist
Cycle 41 split the KMT2A-rearranged DepMap panel by lineage and required a dependency in both halves, but recorded that surviving a split is not a significance test and named BRD9 as needing permutation and line-normalisation. This runs them. The statistic is **max(z_myeloid, z_lymphoid)** — the WEAKER half caps the score, so a one-sided lineage effect cannot rank however extreme its strong half. Per-line offsets come from the random panel only. Against 647 random protein-coding genes: **MEN1 and PSIP1 at the 0.00th percentile, KMT2A 0.15th, BRD9 0.62nd.** The rejected candidates behave as the statistic predicts they should: **CEBPA 48.53rd** (the median of random genes) and **SPI1 95.52nd**.
The result
Line-normalised. Percentile against 647 random genes run through the identical pipeline; lower is more extreme.
| gene | z myeloid | z B-lineage | consistency | percentile |
|---|---|---|---|---|
| MEN1 | -2.12 | -2.157 | -2.12 | 0.0 |
| PSIP1 | -2.698 | -1.754 | -1.754 | 0.0 |
| KMT2A | -1.693 | -1.24 | -1.24 | 0.155 |
| BRD9 | -1.047 | -1.532 | -1.047 | 0.618 |
| MEIS1 | -0.974 | -1.515 | -0.974 | 0.773 |
| MYB | -1.175 | -0.54 | -0.54 | 3.091 |
| DOT1L | -0.446 | -0.77 | -0.446 | 4.637 |
| EP300 | -0.4 | -0.707 | -0.4 | 5.255 |
| ACTB | -0.83 | 0.174 | 0.174 | 43.586 |
| CEBPA | -1.815 | 0.225 | 0.225 | 48.532 |
| SPI1 | -1.57 | 0.988 | 0.988 | 95.518 |
BRD9 is validated. It was the open blocker: a lineage-independent dependency with no significance test behind it. At the 0.62nd percentile after removing each line's global sensitivity offset, it stands.
The controls behave. ACTB sits at the 43.6th percentile, which is correct — a pan-essential gene is required everywhere and therefore is NOT differentially depleted in the index arm. A statistic that ranked ACTB highly would be measuring essentiality rather than selectivity.
CEBPA and SPI1 fail independently of the argument that rejected them. Cycle 41 rejected them by inspecting the split; here they are scored by a null that knows nothing about that reasoning, and CEBPA lands at the median while SPI1 lands in the opposite tail. SPI1 at the 95.5th percentile is LESS lineage-consistently depleted than a typical random gene.
Why the weaker half caps the score
The statistic is deliberately pessimistic. Taking max(z_myeloid, z_lymphoid) means a gene scores only as well as its worst half, so the myeloid-identity pattern — one half extreme, one half at zero — cannot rank no matter how extreme the strong half is. CEBPA's myeloid half is -1.81, which would look impressive alone; its B-lineage half is +0.23 and that is what it scores.
This is the property the cycle-41 finding needed expressing as a single number so a null could be built on it. It also means a real dependency whose second half is merely underpowered will be under-ranked, which is the cost of the design and is why MEIS1 at the 0.77th percentile should not be read as weaker than BRD9 on this evidence alone.
| script | analyses/lineage-consistent-dependency-null/run.py |
|---|---|
| status | executed |
| run date | 2026-09-19 |
| statistic | max(z_myeloid, z_lymphoid) vs other-leukaemia arm, line-normalised |
| n index myeloid | 7 |
| n index lymphoid | 4 |
| n other leukaemia | 43 |
| n null genes | 647 |
| mean offset index | -0.1516 |
| mean offset other | -0.1473 |
| headline | BRD9 validated at the 0.62nd percentile; MEN1 and PSIP1 at 0.00; CEBPA at the median of random genes (48.5) and SPI1 in the opposite tail (95.5). |
| caveat | The lymphoid half is small, so its mean is unstable and the statistic is capped by the noisier of the two halves by construction - that is the point, but it also means a real dependency with one weakly-powered half will be under-ranked. Lineage is assigned from the DepMap disease string, which is a label rather than an expression signature, and KMT2A-rearranged disease includes genuinely ambiguous-lineage cases. The percentile is against random protein-coding genes, most of which are not expressed in leukaemia, so it tests distinctiveness against the genome and not against plausible rival targets. Chronos measures knockout, not interface blockade. |
percentiles — the numbers behind the chart
{
"MEN1": 0,
"PSIP1": 0,
"KMT2A": 0.155,
"BRD9": 0.618,
"MEIS1": 0.773,
"MYB": 3.091,
"DOT1L": 4.637,
"EP300": 5.255,
"ACTB": 43.586,
"CEBPA": 48.532,
"SPI1": 95.518
}z norm — the numbers behind the chart
{
"MEN1": {
"myeloid": -2.12,
"b_lineage": -2.157
},
"PSIP1": {
"myeloid": -2.698,
"b_lineage": -1.754
},
"KMT2A": {
"myeloid": -1.693,
"b_lineage": -1.24
},
"BRD9": {
"myeloid": -1.047,
"b_lineage": -1.532
},
"MEIS1": {
"myeloid": -0.974,
"b_lineage": -1.515
},
"MYB": {
"myeloid": -1.175,
"b_lineage": -0.54
},
"DOT1L": {
"myeloid": -0.446,
"b_lineage": -0.77
},
"EP300": {
"myeloid": -0.4,
"b_lineage": -0.707
},
"ACTB": {
"myeloid": -0.83,
"b_lineage": 0.174
},
"CEBPA": {
"myeloid": -1.815,
"b_lineage": 0.225
},
"SPI1": {
"myeloid": -1.57,
"b_lineage": 0.988
}
}- Every number here is capped by a four-line B-lineage half. The genome-wide percentile makes it look like a well-powered result and it is not; what it is, is a well-specified one.
- The analysis was written to check a claim the coordinator made last cycle and it confirmed it. That is the third consecutive coordinator analysis to conclude the menin axis is the specific thing, and each was designed after seeing the result it was checking.
- The null is random protein-coding genes, most not expressed in leukaemia. Re-running against an EXPRESSED-gene null would shift every percentile and is the stricter test; does BRD9 survive it?
- MEN1 and PSIP1 both land at 0.00, meaning no random gene of 647 matched them. How many genes genome-wide share that property, and does an unbiased scan on this statistic find anything the hand-picked panel missed?
- Only 4 of the 16 accession-resolved KMT2A-rearranged DepMap lines are B-lineage. Is there any public CRISPR dataset with more KMT2A-rearranged B-ALL lines, given that infant KMT2A::AFF1 B-ALL is the disease this project is actually about?
- E1moderate confidence 3 of 5: One strong study, or several smaller consistent ones.analysis
Scoring each gene by max(z_myeloid, z_lymphoid) against the other-leukaemia arm, on line-normalised DepMap gene effects with per-line offsets taken from the random panel only, and comparing to 647 random protein-coding genes run through the identical pipeline: MEN1 and PSIP1 fall at the 0.00th percentile, KMT2A at 0.15, BRD9 at 0.62, MEIS1 0.77, MYB 3.09, DOT1L 4.64 and EP300 5.25. The pan-essential control ACTB sits at 43.59, and the two candidates rejected by the cycle-41 lineage split score at or beyond the middle of the null: CEBPA 48.53 and SPI1 95.52. BRD9's lineage-independent dependency therefore survives line-normalisation and a random-gene null.
Lineage-consistency null test for KMT2A dependencies: line-normalised, random-gene percentile · Oncofusion knowledge graph (internal analysis); DepMap Chronos via Open Targets 2026What this evidence cannot show
Run by the coordinator on 2026-09-19 to discharge a blocker recorded in ana-lineage-confound-myeloid-tf-dependencies. The index arm was resolved from Cellosaurus by DepMap accession (16 lines with a curated KMT2A gene fusion) and split by the DepMap disease string, giving 7 myeloid and only **4 B-lineage** lines - a four-line half whose mean is unstable, and the statistic is capped by it by construction. Lineage from a disease label is not an expression signature, and KMT2A-rearranged disease includes genuinely ambiguous-lineage cases. The null is random protein-coding genes, most not expressed in leukaemia, so the percentile measures distinctiveness against the genome rather than against plausible rival targets. RPL13A could not be scored. Chronos measures knockout, not interface blockade.
Evidence that argues against this (3)
- A percentile against random protein-coding genes flatters any real dependency, because most random genes are not expressed in leukaemia at all and cannot be depleted in any arm. The honest comparator would be a null of genes EXPRESSED in these lines, which would shift every percentile up and could move BRD9 materially. The 0.62 should be read as 'extreme against the genome', not as a p-value against a plausible alternative. Oncofusion knowledge graph (internal analysis); DepMap Chronos via Open Targets 2026
- The B-lineage half is four lines, and two of the field's standard KMT2A::AFF1 models are among the most heavily passaged lines in use. The statistic is capped by that half by design, so every number in the table is effectively a four-line measurement wearing a genome-wide percentile. CEBPA's +0.23 and SPI1's +0.99 - the observations that make them artefacts - rest on those same four lines. Oncofusion knowledge graph (internal analysis); DepMap Chronos via Open Targets 2026
- Line normalisation can remove real signal as well as artefact. The offset subtracted is each line's mean across the random panel, and if KMT2A-rearranged lines are genuinely more dependency-sensitive for biological reasons, subtracting that removes part of what was being measured. The earlier normalisation analysis found the offset difference between arms to be small (-0.0036), which bounds this concern without eliminating it. Oncofusion knowledge graph (internal analysis); DepMap Chronos via Open Targets 2026
Discharges the blocker that entry recorded: BRD9 survives line-normalisation and a random-gene null at the 0.62nd percentile.
BRD9's lineage-independent dependency validated at the 0.62nd percentile of 647 random genes.
MEN1 falls at the 0.00th percentile on a statistic capped by its weaker lineage half.
LEDGF/PSIP1 falls at the 0.00th percentile alongside MEN1.
Third of three tests separating the menin axis from a confound: programme, lineage, and now a random-gene null.
SPI1 fails the lineage-consistency null independently of the argument that rejected it.
CEBPA fails the lineage-consistency null independently of the argument that rejected it.
Discharges both open questions it recorded: the leukaemia-relevant null does not move any candidate materially, and an unbiased sweep found nothing beating MEN1.
- ← related toLineage-split DepMap: neither USP2 nor MAML2 is a dependency in KMT2A-rearranged lines, and USP22 is
same null framework, extended to USP2, MAML2 and USP8