The confounder every previous analysis flagged and none removed, removed (run)
also called within-subtype centring · AML vs ALL control · lineage-within-leukemia null
Every earlier version of this test compared leukemia cell lines as if they were one group, but they are not: most are myeloid and the rest are lymphoid. Two genes that simply matter more in myeloid leukemia would look linked even if they never touch. Comparing each line only against others of its own subtype removes that, and the link between menin and the fusion machinery survives almost untouched.
The same thing, for a scientist
Removes the between-subtype contrast that ana-depmap-codependency, -null-distribution, -null-variance-matched and -null-deep-pool all named as their largest uncontrolled confounder. Gene effects are centred within each leukemia subtype (AML 32, B-ALL 13, T-ALL 8) before correlating with similarly centred MEN1 effect, so only within-subtype covariation contributes. KMT2A retains 99.3% of its raw correlation, PSIP1 99.5%, DOT1L 97.3% and FLT3 97.2%. The tether is not a myeloid-versus-lymphoid artefact.
The objection this answers
Four analyses in this project have restricted to leukemia lines and correlated candidates against MEN1, and every one recorded the same unfixed weakness: "Variance matching does not address lineage restriction within leukemia, which remains the confounder no analysis in this project has removed."
The worry was specific and serious. The 53 usable leukemia lines are 32 AML, 13 B-ALL and 8 T-ALL. Any two genes that are merely more essential in myeloid than in lymphoid disease will correlate strongly across the pooled set with no physical or functional relationship at all. A menin-KMT2A tether that was really an AML-versus-ALL contrast would produce exactly the result we kept getting, four times over.
This was not a hypothetical. The project had already overturned one claim on precisely this ground: ana-depmap-lineage-conditioned killed the MYB co-dependency once tissue was controlled for. There was no reason to assume menin would survive where MYB did not.
Method, and why the null had to be rebuilt
Each gene's effect is centred within each leukemia subtype before correlation, so every line is compared against its own subtype's mean rather than the pooled mean. Between-subtype differences then contribute nothing, and only within-subtype covariation survives. The question becomes: do lines that need MEN1 more than other AMLs do also need KMT2A more than other AMLs do?
The single ambiguous-lineage line was dropped rather than pooled, because putting it in an 'other' bucket would reintroduce exactly the between-group contrast being removed.
A centred correlation has a different null distribution from a raw one, so the 1,090-gene variance-matched null was recomputed on the centred statistic. Comparing a centred r against a raw null would have been meaningless, and would have been an easy mistake to make.
The result: the signal is not lineage2 deeper
| Gene | r raw | r within-subtype | % retained | empirical p | Holm | survives |
|---|---|---|---|---|---|---|
| KMT2A | 0.781 | 0.776 | 99.3 | <0.020 | <0.001 | yes |
| PSIP1 (LEDGF) | 0.684 | 0.681 | 99.5 | <0.016 | <0.001 | yes |
| DOT1L | 0.677 | 0.660 | 97.3 | <0.015 | <0.001 | yes |
| FLT3 | 0.643 | 0.625 | 97.2 | <0.040 | <0.001 | yes |
| MEF2C | 0.531 | 0.478 | 90.1 | <0.024 | <0.001 | yes |
| MEIS1 | 0.464 | 0.440 | 94.9 | 0.0083 | 0.041 | yes |
| KAT7 | 0.528 | 0.486 | 92.1 | 0.0081 | 0.041 | yes |
| DPY30 | 0.439 | 0.402 | 91.6 | 0.028 | 0.085 | no |
| BCL2 | 0.432 | 0.375 | 86.7 | 0.090 | 0.180 | no |
| MLLT1 | 0.421 | 0.345 | 82.1 | 0.099 | 0.180 | no |
| GAPDH (control) | 0.147 | 0.210 | 143 | 0.333 | — | correctly null |
| ACTB (control) | 0.156 | 0.189 | 121 | 0.362 | — | correctly null |
The top four lose between 0.5% and 2.8% of their correlation. That is the headline: the AML-versus-ALL contrast contributes almost nothing to the menin-KMT2A-LEDGF-DOT1L tether. The confounder that four analyses flagged as potentially fatal turns out to be negligible for the genes the project's argument rests on.
The weaker candidates attenuate more — MLLT1 loses 18%, BCL2 13% — which is the expected pattern if part of their raw signal really was lineage. That the attenuation is graded, and largest for the weakest genes, is itself evidence the procedure is doing what it claims.
The controls went up, and that is not a failure
ACTB and GAPDH both show a higher correlation after centring (0.147 to 0.210 for GAPDH). Centring can increase a correlation: removing between-group variance shrinks the denominator as well as the numerator.
What matters is that both remain squarely non-significant, at the 72nd and 73rd percentile of their own matched nulls with p of 0.33 and 0.36. The controls behave like controls. Had they crossed into significance, this entire analysis would have been uninterpretable.
MEIS1 has now flipped three times and should be called unresolved
MEIS1 fails Holm correction in the deep-pool null (0.067) and passes it here (0.041). It has now changed status across three defensible designs. The honest reading is that MEIS1 sits at the threshold, and which side of it lands depends on the null.
Calling it significant because this analysis is the newest would repeat exactly the error ana-depmap-null-deep-pool was written to expose. MEIS1 is unresolved.
What is still not controlled
Finer co-essentiality structure inside AML. Thirty-two AML lines are not homogeneous: they differ by differentiation state, by driver mutation, and by KMT2A-rearrangement status itself. Genes co-essential within a differentiation state will still correlate. Subtype centring is a coarse control, not a complete one.
The lymphoid groups are small. Thirteen B-ALL and eight T-ALL lines centre on means estimated from very few observations, which adds noise. The result is therefore carried mostly by the 32 AML lines.
None of these lines is infant or KMT2A-rearranged-specific, which is the population the project is actually about. Every co-dependency result here is an inference from adult-dominated cell lines to infant leukemia, and no amount of statistical control fixes that.
Correlation is still not interaction. Surviving this does not show menin binds any of these proteins. It shows the co-dependency is not an artefact of the one confounder tested.
| script | analyses/depmap-within-subtype/run.py |
|---|---|
| status | run |
| run date | 2026-09-18 |
| method | gene effects centred within each leukemia subtype of at least 5 lines, then correlated with centred MEN1 effect; variance-matched null recomputed on the centred statistic; Holm over ten candidates |
| seed | 20260918 |
| n lines | 53 |
| n random usable | 1090 |
| figure | analyses/depmap-within-subtype/results.md |
answers open question from
[ "ana-depmap-null-deep-pool", "ana-depmap-null-variance-matched", "ana-depmap-null-distribution", "ana-depmap-codependency-men1-dot1l" ]
subtypes — the numbers behind the chart
{
"acute myeloid leukemia": 32,
"b-cell acute lymphoblastic leukemia": 13,
"t-lymphoblastic leukemia/lymphoma": 8
}dropped
{
"acute leukemias of ambiguous lineage": 1
}PCT of raw retained — the numbers behind the chart
{
"KMT2A": 99.3,
"PSIP1": 99.5,
"DOT1L": 97.3,
"FLT3": 97.2,
"MEIS1": 94.9,
"KAT7": 92.1,
"DPY30": 91.6,
"MEF2C": 90.1,
"BCL2": 86.7,
"MLLT1": 82.1
}survives holm
[ "KMT2A", "PSIP1", "DOT1L", "FLT3", "MEF2C", "MEIS1", "KAT7" ]
does not survive holm
[ "DPY30", "BCL2", "MLLT1" ]
controls
[
{
"gene": "ACTB",
"r_raw": 0.1562,
"r_within": 0.189,
"empirical_p": 0.36207
},
{
"gene": "GAPDH",
"r_raw": 0.1474,
"r_within": 0.2104,
"empirical_p": 0.33333
}
]- This is the result the project wanted, which is exactly why it deserves the most suspicion. It was run to try to kill the tether and failed to; it was not run to confirm it.
- Surviving one confounder is not surviving all of them. The structure remaining inside 32 AML lines is the same kind that killed the MYB claim, and it is untested.
- Nothing here shows menin binds KMT2A, LEDGF or DOT1L. It shows the co-dependency is not explained by myeloid versus lymphoid lineage. Treating it as validation of a physical complex would overstate it badly.
- KAT7 and MEIS1 clearing Holm at 0.041 is a hair's breadth. Two genes at p just under threshold in a set of ten is what borderline noise looks like.
- AML lines still differ by differentiation state, driver genotype and KMT2A-rearrangement status. Does the tether survive centring within KMT2A-rearranged versus wild-type AML, and is there even enough n to ask?
- If the co-dependency is genuinely stronger in KMT2A-rearranged lines specifically, that is a much stronger claim than anything shown so far. Nobody in this project has tested it, and the annotation to do so may exist.
- Every line here is adult-dominated and none is infant. What is the strongest available check that adult cell-line co-dependency transfers to infant KMT2A-r leukemia, and has anyone published it?
- MEIS1 has flipped three times across three designs. Is there a test that would actually settle it, or is it simply a borderline effect that no reanalysis of 53 lines can resolve?
- E1moderate confidence 3 of 5: One strong study, or several smaller consistent ones.dataset
With DepMap gene effects centred within leukemia subtype (32 AML, 13 B-ALL, 8 T-ALL; one ambiguous-lineage line dropped) and correlated against similarly centred MEN1 effect, KMT2A retains 99.3% of its raw correlation (0.781 to 0.776), PSIP1 99.5% (0.684 to 0.681), DOT1L 97.3% (0.677 to 0.660) and FLT3 97.2% (0.643 to 0.625). Against a 1,090-gene variance-matched null recomputed on the centred statistic and Holm-corrected, KMT2A, PSIP1, DOT1L, FLT3, MEF2C, MEIS1 and KAT7 survive; DPY30, BCL2 and MLLT1 do not. Controls ACTB and GAPDH give p 0.36 and 0.33.
DepMap CRISPR (Chronos) gene effect via the Open Targets Platform GraphQL API, centred within leukemia subtype, with a variance-matched null recomputed on the centred statistic; analysis in analyses/depmap-within-subtype/ · Broad Institute DepMap; HGNC; Open Targets Platform; analysis run by this project · Databases and derived analysis 2026What this evidence cannot show
Our own computation, not peer reviewed. Subtype centring removes the between-subtype contrast but not finer co-essentiality structure within AML, such as shared differentiation state or driver genotype. The B-ALL (13) and T-ALL (8) groups are small, so their subtype means are noisy and the result is carried mainly by the 32 AML lines. No line is infant or KMT2A-rearranged-specific, so this is an inference from adult-dominated cell lines to infant leukemia. Correlation across lines is not evidence of physical interaction. DepMap release version is not stated by the API.
Evidence that argues against this (2)
- MEIS1 fails Holm correction in the deep-pool null (0.067) and passes here (0.041), having now changed status across three defensible designs. Treating this result as settling MEIS1 would repeat the error that the deep-pool analysis exists to expose. It is unresolved. analysis run by this project 2026
- This project has already seen a co-dependency die once the right covariate was added: the MYB claim did not survive lineage conditioning. That a different claim survives a different control is not a general warrant, and the remaining uncontrolled structure inside AML is precisely the kind that killed MYB. analysis run by this project 2026
Companion confound test. Within-subtype centring removes lineage; this removes general essentiality.
- ← supportsThe assumption the whole project rests on, tested at last: menin dependency is fusion-specific (run)
answers that entry's first open question: the dependency is not just non-lineage, it tracks the rearrangement itself
the same confound at finer grain, and here the panel is too small for the control that worked there
- supports →Six times the null pool, and the criterion I had been using turns out to be the problem (run)
removes the lineage confounder that entry named as its largest remaining weakness; the top four survive nearly untouched
the original co-dependency is not an AML-versus-ALL artefact: KMT2A retains 99.3% of its correlation after subtype centring
the same control at finer grain: that entry conditioned on tissue, this one on subtype within leukemia
rebuilds the variance-matched null on the centred statistic, because a centred r cannot be compared against a raw null