Does the menin co-dependency survive lineage adjustment? (run)
also called lineage-conditioned co-dependency · within-tissue correlation MEN1
Our earlier scan found that cells needing menin also need KMT2A and LEDGF, but that could have been a trick of the data: two genes can look related simply because both matter in the same tissue. We removed that effect and the result held almost perfectly, then got noticeably stronger when we looked only at leukemia.
The same thing, for a scientist
Tests the lineage confounder flagged in ana-depmap-codependency-men1-dot1l. Three views of DepMap CRISPR gene effect against MEN1: pooled across all lines; centred within tissue so only within-lineage variance contributes; and restricted to leukemia lines. The menin-LEDGF-KMT2A tether retains 97-99% of its correlation under within-tissue centring and roughly doubles in leukemia, while the pan-essential genes and the negative controls stay flat throughout.
The confounder is ruled out1 deeper
A correlation across cell lines can arise because two genes act in the same process, or merely because both happen to be required in the same lineage. Centring each gene on its own tissue mean removes everything the second explanation could contribute. If a correlation is a lineage artefact, it collapses.
It did not collapse 1.
| Gene | pooled r | within-tissue r | retained | leukemia-only r |
|---|---|---|---|---|
| KMT2A | 0.58 | 0.58 | 99% | 0.78 |
| DOT1L | 0.57 | 0.54 | 95% | 0.68 |
| PSIP1 (LEDGF) | 0.53 | 0.52 | 97% | 0.69 |
| GAPDH (control) | 0.04 | 0.001 | — | 0.15 |
| ACTB (control) | 0.01 | 0.01 | — | 0.15 |
The tether module keeps essentially all of its signal, and the negative controls stay at zero under both treatments. This is about as clean as an observational dependency result gets.
And it concentrates where the fusion operates
Restricting to the 54 leukemia lines, the same three correlations rise sharply: KMT2A from 0.58 to 0.78, LEDGF from 0.53 to 0.69, DOT1L from 0.57 to 0.68.
That is the direction the mechanism predicts. If menin, LEDGF and KMT2A form an obligate module whose importance is greatest where a KMT2A fusion is driving the cell, the co-dependency should be tightest in leukemia. It is.
The output-program genes behave the same way, from almost nothing pan-cancer to a real signal in leukemia: HOXA9 0.003 to 0.37, MEIS1 0.08 to 0.47, FLT3 0.15 to 0.65, MEF2C 0.27 to 0.53. Those genes are only co-required where the program is actually running, which is exactly right and which a pan-cancer analysis almost completely hides.
This corrects our own claim from the previous cycle
The earlier analysis singled out MYB as the most interesting orthogonal candidate: strongest leukemia selectivity in the table while barely correlating with menin, which made it look like a genuinely independent dependency worth combining against.
Lineage adjustment does not support that reading. MYB's pooled correlation of 0.11 reverses to -0.05 once tissue is controlled for, meaning its apparent pan-cancer relationship to menin was itself a lineage effect. Its leukemia-only correlation is 0.35, so within leukemia it does track menin somewhat.
MYB may still be a good combination target on selectivity grounds, but the specific argument we made for it, that it is uncorrelated with menin and therefore hedges an independent risk, does not survive this analysis. Recording the correction rather than quietly dropping it is the point of keeping these entries.
What stayed flat, and why that is informative
The pan-essential genes remain near zero under every treatment: CDK9 -0.00 pooled, -0.03 within, -0.09 in leukemia; WDR5 -0.02, -0.03, 0.00; CTR9 0.03, 0.01, 0.01; PAF1 0.04, 0.02, -0.10.
This confirms the structural point from the first analysis with a much harder test. A gene required by every cell has almost no variance to correlate with, in any stratum. The elongation machinery the fusion depends on is invisible to co-dependency analysis not because the biology is absent but because the statistic cannot see it.
A few genes turn mildly negative within leukemia (AFF4 -0.11, MLLT3 -0.03, RBBP5 -0.08, PAF1 -0.10). With 54 lines these are not distinguishable from noise and should not be interpreted.
Limits
Fifty-four leukemia lines is a small stratum, so the leukemia-only column has wide uncertainty that this analysis does not quantify with confidence intervals. None of the lines is an infant KMT2A-rearranged line. Within-tissue centring removes between-lineage variance but not other shared structure such as correlated expression or growth rate. Tissues with fewer than three lines are dropped from the centred analysis. The DepMap release version is not stated by the API.
No bootstrap or permutation test was run, so the correlations carry no p-values. For the three headline genes the effect sizes are large enough that this is unlikely to change the conclusion; for everything below MEF2C it matters.
| script | analyses/depmap-lineage-conditioned/run.py |
|---|---|
| status | run |
| run date | 2026-09-18 |
| method | Pooled Pearson vs MEN1; within-tissue-centred Pearson (tissues with >=3 lines); Pearson restricted to leukemia lines |
| n genes | 30 |
| n leukemia lines | 54 |
| figure | analyses/depmap-lineage-conditioned/results.md |
headline — the numbers behind the chart
[
{
"gene": "KMT2A",
"pooled_r": 0.5834,
"within_tissue_r": 0.5767,
"leukemia_only_r": 0.7831,
"fraction_retained": 0.989
},
{
"gene": "DOT1L",
"pooled_r": 0.5703,
"within_tissue_r": 0.5439,
"leukemia_only_r": 0.6808,
"fraction_retained": 0.954
},
{
"gene": "PSIP1",
"pooled_r": 0.5288,
"within_tissue_r": 0.515,
"leukemia_only_r": 0.686,
"fraction_retained": 0.974
}
]output program emerges in leukemia — the numbers behind the chart
[
{
"gene": "FLT3",
"pooled_r": 0.1516,
"leukemia_only_r": 0.6452
},
{
"gene": "MEF2C",
"pooled_r": 0.2678,
"leukemia_only_r": 0.526
},
{
"gene": "MEIS1",
"pooled_r": 0.0771,
"leukemia_only_r": 0.468
},
{
"gene": "HOXA9",
"pooled_r": 0.0029,
"leukemia_only_r": 0.372
}
]correction to prior claim
{
"gene": "MYB",
"prior_claim": "uncorrelated with menin and therefore an orthogonal combination candidate",
"pooled_r": 0.1051,
"within_tissue_r": -0.0515,
"leukemia_only_r": 0.3534,
"verdict": "the pan-cancer correlation was itself a lineage effect; within leukemia MYB does track menin, so the orthogonality argument does not hold as stated"
}negative controls — the numbers behind the chart
[
{
"gene": "GAPDH",
"pooled_r": 0.0443,
"within_tissue_r": 0.0009
},
{
"gene": "ACTB",
"pooled_r": 0.0071,
"within_tissue_r": 0.0118
},
{
"gene": "TP53",
"pooled_r": -0.017,
"within_tissue_r": -0.0131
}
]- A co-dependency that is strongest in leukemia lines still says nothing about whether the same module is required in normal blood progenitors, which is where the toxicity would be.
- Acting on a mid-table correlation from 54 cell lines without confidence intervals would be over-reading the data.
- With only 54 leukemia lines, how wide are the confidence intervals on the leukemia-only correlations? A bootstrap would cost nothing and would tell us which of the mid-table genes are real.
- Does restricting further to KMT2A-rearranged lines specifically, rather than all leukemia, sharpen the tether correlation further or exhaust the sample?
- The output-program genes only appear as co-dependencies in leukemia. Is that a general property, meaning pan-cancer dependency screens systematically hide lineage-specific programs from this kind of analysis?
- Would an unbiased genome-wide within-tissue scan against MEN1 surface anything in the top fifty that is not already in the known interactome?
- E1moderate confidence 3 of 5: One strong study, or several smaller consistent ones.dataset
Correlation of DepMap CRISPR gene effect with MEN1 retains 95-99% of its magnitude for KMT2A, PSIP1 and DOT1L after within-tissue centring, and rises to 0.78, 0.69 and 0.68 respectively when restricted to leukemia lines, while essential negative controls ACTB and GAPDH remain near zero throughout.
DepMap CRISPR (Chronos) gene effect per cell line via the Open Targets Platform GraphQL API; analysis script and results in analyses/depmap-lineage-conditioned/ · Broad Institute DepMap; Open Targets Platform; analysis run by this project · Database and derived analysis 2026What this evidence cannot show
Our own computation, not peer reviewed and with no confidence intervals or permutation testing. Only 54 leukemia lines pass the disease-string filter, and none is an infant KMT2A-rearranged line. Within-tissue centring removes lineage but not other shared structure such as correlated expression. DepMap release version unverified.
Evidence that argues against this (1)
- Cell lines drift genetically and transcriptionally in culture, so dependency correlations measured across panels may not reflect primary disease biology. Ben-David U et al. 2018
- E2established confidence 5 of 5: Multiple independent lines of evidence; treated as settled in the field.paper
Menin bridges KMT2A to LEDGF/PSIP1 as a chromatin adaptor, which is the physical basis for expecting these three genes to be co-required.
Menin critically links MLL proteins with LEDGF on cancer-associated target genes. · Yokoyama A, Cleary ML · Cancer Cell 2008What this evidence cannot show
Establishes the physical module; does not predict the quantitative dependency correlation measured here.
Evidence that argues against this (1)
- Citing the mechanism as the reason a co-dependency is EXPECTED turns a prior into support. The analysis is only informative if it could have come out the other way, and the physical bridge does not in fact predict correlated gene effects across a panel: two proteins in one complex show co-dependency only when both are rate-limiting in the same lines, which is an empirical question the mechanism does not settle. A chromatin adaptor bridging two proteins predicts that removing any one breaks the tether; it does not predict that a CRISPR screen sees all three as equally limiting, and in this project's own runs PSIP1 and MEN1 separate by a factor of two on effect size. Oncofusion knowledge graph (internal analysis) 2026
- ← depends onSix times the null pool, and the criterion I had been using turns out to be the problem (run)
the lineage confounder this analysis cannot remove is the one that analysis addressed for a different question
rules out the lineage confounder the earlier scan flagged, and corrects its MYB claim
the tether co-dependency survives lineage adjustment and strengthens in leukemia
menin and DOT1L co-depend at r=0.68 within leukemia
- supports →Step 5 — The output program: HOXA9/HOXA10/MEIS1/PBX3 plus MEF2C, MYB, BCL2, FLT3 and PROM1/CD133
HOXA9 and MEIS1 only appear as co-dependencies in leukemia lines, where the program runs
FLT3 co-dependency rises from 0.15 pan-cancer to 0.65 within leukemia
puts confidence intervals and corrected p-values on the leukemia column
- ← supportsPhase-2 dossier: CBX8 - the candidate this review would kill, and the reasons are all numeric
CBX8 sits at negative-control level in both analyses (-0.002 within-tissue)
names the single biggest hole in this result: ARID1A-mutant lines are lineage-enriched and a lineage-conditioned version of this contrast was NOT run; that analysis is the template
the template that entry named; this applies within-stratum permutation rather than adjustment, so lineage is held fixed by construction
the leukaemia-restricted arm applies that entry's conditioning logic and finds the negative survives it
THE COMPANION CONFOUND. Lineage is the other categorical predictor this project conditions on, and it has not been tested for the same definition-sensitivity.
tension to resolve: MYB ranks 9 here but was overturned as a MEN1 co-dependency there
the same control at finer grain: that entry conditioned on tissue, this one on subtype within leukemia
G3 thresholds are quoted against this executed lineage-controlled analysis
essentiality numbers taken from this executed analysis
selectivity numbers taken from this executed analysis
co-dependency numbers taken from this executed analysis
lineage-conditional co-dependency numbers taken from this executed analysis
lineage-controlled selectivity taken from this executed analysis
the executed analysis this access plan generalises