Running the HOXA programme is not enough to need menin
also called programme versus rearrangement · three-arm menin dependency
Two leukaemia cell lines switch on the same set of genes as our disease but by different routes. One of them needs menin even more than our own lines do; the other does not need it at all. So needing menin is not simply a consequence of switching those genes on.
The same thing, for a scientist
Every DepMap analysis in this project contrasts KMT2A-rearranged lines against "other leukaemia", which cannot separate the REARRANGEMENT from the HOXA/MEIS1 PROGRAMME. Splitting three ways instead — index, pathway-equivalent, and a cleaned negative arm — dissociates them. **OCI-AML3 (NPM1c) is more menin-dependent than the KMT2A-rearranged index arm (-2.98 SD vs -2.42 SD below the negative arm) and more LEDGF/PSIP1-dependent (-2.80 vs -2.49). U-937 (PICALM-MLLT10) is neither: +0.33 SD on MEN1 and +0.49 SD on PSIP1, i.e. indistinguishable from an average comparator line** — while still being DOT1L-dependent (-0.54 SD). The HOXA/MEIS1 programme is therefore not sufficient for menin dependence.
The dissociation
Standard deviations below the negative arm's mean. More negative = more required.
| gene | KMT2A-r index | OCI-AML3 (NPM1c) | U-937 (PICALM-MLLT10) |
|---|---|---|---|
| MEN1 | -2.42 | -2.98 | +0.33 |
| PSIP1 | -2.49 | -2.80 | +0.49 |
| DOT1L | -0.70 | -1.77 | -0.54 |
| MEIS1 | -1.28 | -1.59 | -0.24 |
| HOXA9 | -0.37 | -1.45 | -0.42 |
| KMT2A | -1.72 | -2.32 | -0.42 |
| AFF1 | -1.89 | -2.71 | +0.28 |
| MLLT1 | -0.68 | -2.04 | -0.34 |
MEN1 and PSIP1 move together and U-937 has neither. PICALM-MLLT10 reaches HOXA9/MEIS1 through a DOT1L-dependent route that does not pass through the menin-LEDGF module, and its DOT1L dependency is intact while its menin dependency is absent. NPM1c does the opposite: it exceeds the index arm on menin, LEDGF and DOT1L alike.
That ordering is consistent with where the drugs are approved: revumenib for KMT2A-translocated acute leukaemia and ziftomenib for NPM1-mutant AML. No menin inhibitor is approved for, or has reported a subgroup in, PICALM-MLLT10 disease - but that is an absence of testing rather than a reported negative, and it is weaker corroboration than it first looks. Care is needed here because KMT2A::MLLT10 and PICALM-MLLT10 share a partner gene and are routinely conflated: they are t(10;11)(p12;q23) and t(10;11)(p13;q14) respectively, and only the first is a KMT2A rearrangement.
This corrects the coordinator's own previous analysis
Cycle 39 found that the comparator arm of every DepMap number in this project contained two lines running the same programme by another route, and called both pathway-equivalent contaminants: U-937 and OCI-AML3. Cycle 40 measured the bias and found it ran toward the null.
The label was too coarse, and this analysis shows where. Only OCI-AML3 is contamination for a menin question. U-937 is HOXA/MEIS1-programme-equivalent but menin-independent, so leaving it in the comparator does not blunt a menin contrast the way removing it assumed. The cycle-39 rule — screen a control group for pathway-equivalent lesions — still stands, but "pathway-equivalent" has to be defined against the pathway being tested, not against the disease in general.
The practical effect is small and in the same direction as before: removing a menin-independent line from the comparator slightly understates menin selectivity rather than overstating it. No published conclusion changes.
What this design can and cannot support
n=1 per lesion. DepMap's entire leukaemia panel contains two pathway-equivalent lines, so this is not a sample, it is the population, and the population is two. The formal power calculation is brutal: at n=2 the minimum detectable difference for MEN1 is 0.762 on the Chronos scale, and the pooled programme-arm difference was 0.498. Pooled, not one of the 17 panel genes reaches significance.
So the numbers above are read as a dissociation in direction, corroborated by the clinical approvals, and not as an effect size. Two things make the direction harder to dismiss than n=1 usually is: MEN1 and PSIP1 agree with each other independently, and the same line is simultaneously DOT1L-dependent, which rules out a flat 'U-937 is a poor screen' explanation.
Pooling the two lines destroys the result, which is the methodological lesson worth keeping. The arm mean of two unlike lesions estimates neither. An earlier version of this script reported only the pooled arm and showed a bland 'programme reproduces 0.55 of the index effect'.
| script | analyses/menin-programme-vs-rearrangement/run.py |
|---|---|
| status | executed |
| run date | 2026-09-19 |
| min detectable diff MEN1 at n2 | 0.7622 |
| n panel genes significant pooled | 0 |
| n panel genes ok | 17 |
| headline | OCI-AML3 (NPM1c) exceeds the KMT2A-r arm on menin and LEDGF; U-937 (PICALM-MLLT10) has neither while staying DOT1L-dependent. The programme is not sufficient for menin dependence. n=1 per lesion. |
| caveat | THE PROGRAMME ARM IS n=2, and that is the entire supply in DepMap's leukaemia panel, not a sampling choice. Any 'programme reproduces X% of the index effect' figure below is two cell lines and should be read as a direction to test, never as an estimate. The minimum-detectable-difference column is the honest output of this analysis: it says what an n=2 arm could ever have shown. It is also computed on the optimistic normal approximation, so the real power at n=2 is worse than stated. Two further limits: OCI-AML3 carries DNMT3A R882C alongside NPM1c and U-937 carries a CDKN2A deletion and TP53 mutation, so neither line isolates its pathway-equivalent lesion; and a Chronos gene effect is a knockout, which models menin LOSS rather than menin-KMT2A interface blockade. |
arms
{
"index": "KMT2A-rearranged, accession-verified (n=11)",
"programme": {
"ACH-000336": "OCI-AML3 - NPM1c (menin-inhibitor-sensitive genotype)",
"ACH-000406": "U-937 - PICALM-MLLT10 (DOT1L-dependent HOXA/MEIS1 driver)"
},
"negative": "remaining leukaemia lines (n=44)"
}z vs negative — the numbers behind the chart
{
"MEN1": {
"index": -2.42,
"OCI-AML3_NPM1c": -2.98,
"U-937_PICALM-MLLT10": 0.33
},
"PSIP1": {
"index": -2.49,
"OCI-AML3_NPM1c": -2.8,
"U-937_PICALM-MLLT10": 0.49
},
"DOT1L": {
"index": -0.7,
"OCI-AML3_NPM1c": -1.77,
"U-937_PICALM-MLLT10": -0.54
},
"MEIS1": {
"index": -1.28,
"OCI-AML3_NPM1c": -1.59,
"U-937_PICALM-MLLT10": -0.24
},
"HOXA9": {
"index": -0.37,
"OCI-AML3_NPM1c": -1.45,
"U-937_PICALM-MLLT10": -0.42
},
"KMT2A": {
"index": -1.72,
"OCI-AML3_NPM1c": -2.32,
"U-937_PICALM-MLLT10": -0.42
},
"AFF1": {
"index": -1.89,
"OCI-AML3_NPM1c": -2.71,
"U-937_PICALM-MLLT10": 0.28
},
"MLLT1": {
"index": -0.68,
"OCI-AML3_NPM1c": -2.04,
"U-937_PICALM-MLLT10": -0.34
}
}- This analysis produces a conclusion its own power calculation says it cannot support: not one of 17 genes reaches significance pooled, and the headline rests on two cell lines. It is published here as a direction with the arithmetic attached, and anyone quoting the SD figures without the n is misusing it.
- It also retracts part of the reasoning behind two previous cycles' work. U-937 was called comparator contamination on the basis of its HOXA/MEIS1 programme; on its own dependency profile it is menin-independent and was never contamination for a menin question.
- Is PICALM-MLLT10 leukaemia clinically menin-inhibitor resistant, as this DepMap signal predicts? NOTE, corrected on writing: AUGMENT-101's seven t(10;11) patients are NOT the test. They sit in a KMT2A-REARRANGED cohort, so they carry KMT2A::MLLT10 at t(10;11)(p12;q23); PICALM-MLLT10 is t(10;11)(p13;q14), a different translocation involving neither KMT2A nor that breakpoint. The two share a partner gene and nothing else relevant. Answering this needs a CALM-AF10 cohort, and no menin-inhibitor trial has enrolled one.
- Does LEDGF/PSIP1 dependency track menin dependency across genotypes generally? Here they agree in all three arms, which would make PSIP1 a cheap surrogate for menin dependence if it holds up beyond two lines.
- Every DepMap analysis in this project used a two-arm design. Which of them would change if re-run with the programme arm separated, and is separating it even possible at n=2?
- E1weak confidence 2 of 5: Preliminary, small, indirect, or not yet peer reviewed.analysis
Splitting DepMap leukaemia lines three ways rather than two dissociates menin dependency from HOXA/MEIS1 programme membership. Relative to the cleaned negative arm, OCI-AML3 (NPM1c) is -2.98 SD on MEN1 and -2.80 SD on PSIP1, both exceeding the KMT2A-rearranged index arm (-2.42 and -2.49 SD). U-937 (PICALM-MLLT10) is +0.33 SD on MEN1 and +0.49 SD on PSIP1, indistinguishable from an average comparator line, while remaining DOT1L-dependent at -0.54 SD. Running the HOXA/MEIS1 programme is therefore not sufficient for menin dependence.
Three-arm DepMap analysis: menin dependency against KMT2A rearrangement versus HOXA/MEIS1 programme membership · Oncofusion knowledge graph (internal analysis); DepMap Chronos via Open Targets 2026What this evidence cannot show
Run by the coordinator on 2026-09-19. n=1 per lesion type, because DepMap's leukaemia panel contains exactly two pathway-equivalent lines in total. Pooled, NONE of the 17 panel genes clears the minimum detectable difference at n=2 (0.762 for MEN1 against a pooled difference of 0.498), and the MDD is computed on the optimistic normal approximation so true power is worse. Neither line isolates its lesion: OCI-AML3 carries DNMT3A R882C alongside NPM1c, and U-937 carries CDKN2A deletion and mutant TP53. A Chronos gene effect is a knockout and models menin LOSS, not menin-KMT2A interface blockade, which the USP22 precedent in this graph shows can differ in sign. Read as a direction to test, never as an effect size. One further correction made while writing: an earlier draft of this entry proposed AUGMENT-101's seven t(10;11) patients as the clinical test of the prediction. They are not - they are KMT2A::MLLT10 in a KMT2A-rearranged cohort, not PICALM-MLLT10.
Evidence that argues against this (3)
- A single cell line can fail to show a dependency for reasons that have nothing to do with its driver: copy number, screen quality, growth rate, or simple Chronos noise. With one line per lesion there is no way to separate 'PICALM-MLLT10 leukaemia does not need menin' from 'this particular U-937 culture did not score'. The honest status of the headline is a hypothesis generated by two data points. Oncofusion knowledge graph (internal analysis); DepMap Chronos via Open Targets 2026
- The result is also convenient for this project, which should lower rather than raise confidence in it. A menin-centred graph benefits from a finding that menin dependence is specific rather than a generic consequence of a HOXA-high state, and the coordinator both chose the panel and read the output. The pre-registered version of this analysis does not exist. Oncofusion knowledge graph (internal analysis); DepMap Chronos via Open Targets 2026
- It cuts against a claim this project has made twice. Cycle 39 called U-937 a pathway-equivalent contaminant of the comparator arm and cycle 40 removed it when measuring bias. On this analysis U-937 is menin-INDEPENDENT, so it was never contamination for a menin contrast and removing it was not the correction it was described as. The direction of that error is conservative, and no published number changes, but the reasoning was wrong. Oncofusion knowledge graph (internal analysis); DepMap Chronos via Open Targets 2026
The NPM1c line OCI-AML3 is the programme arm of that analysis and is MORE menin-dependent than the KMT2A-rearranged arm.
Shows the 'pathway-equivalent contaminant' label was too coarse: U-937 is menin-independent, so it was never contamination for a menin contrast.
Supplies the two pathway-equivalent lines this analysis splits out, and shows they are not equivalent to each other.
Menin dependency dissociates from HOXA/MEIS1 programme membership across the two non-KMT2A lesions.
LEDGF/PSIP1 dependency tracks menin dependency in all three arms, absent in the menin-independent line.
DOT1L dependency is retained in the menin-independent programme line, which is what makes the dissociation interpretable.
Menin dependence is not a generic consequence of a HOXA-high state, which argues the interface is doing something specific.
Companion test: that one separates programme from rearrangement, this one separates lineage from rearrangement.
Third of three tests separating the menin axis from a confound: programme, lineage, and now a random-gene null.
SAME LINE, PRIOR ANALYSIS, AND THE POWER BOUND: that three-arm split's PROGRAMME arm is exactly {OCI-AML3, U-937}; mean_index -1.2629 matches, and min_detectable_diff_at_n2 = 0.762 against an observed 0.498 means the arm is not powered.