Is the selectivity just pan-essentiality? No: it explains 0.4% of the variance
also called essentiality confound regression · residual-vs-confound control
The obvious objection to our genome-wide hit list is that genes every cell needs will look important in KMT2A-rearranged cells too, so the ranking is measuring neediness rather than anything about the fusion. We tested that directly and it is not what is happening: how essential a gene is in general predicts almost nothing about how much more the KMT2A-rearranged lines need it.
The same thing, for a scientist
Across 1,379 random protein-coding genes the KMT2A-rearranged-versus-other-leukaemia difference regressed on pan-essentiality (mean Chronos gene effect across all lines) gives difference = -0.0050 + 0.0128 x pan-essentiality, Pearson r = +0.065, r^2 = 0.0043. Residual SD 0.0701 against raw SD 0.0703: removing the confound removes essentially none of the spread. MEN1's residual z is -12.2, the most extreme of 22 probes. The sign of the slope is positive, i.e. more-essential genes show slightly LESS differential requirement, which is the opposite of the direction the objection predicts.
The objection, stated properly so it can be killed
The genome-wide scan ranks genes by how much more the KMT2A-rearranged lines need them than other leukaemia lines. A reviewer's first move is to say that the top of such a list is always contaminated: ribosomal proteins, proteasome subunits and spliceosome components are required by every cell, they have the most negative scores everywhere, and any difference statistic built on them inherits that floor. On this reading MEN1 ranking first is an artefact of MEN1 being broadly required, not evidence of anything fusion-specific.
That is a quantitative claim and it can be measured rather than argued about. If it is right, pan-essentiality should predict the difference statistic, and the top hits should mostly disappear once you subtract what essentiality already explains.
It explains 0.4% of the variance, and with the wrong sign
Regressing the difference on pan-essentiality across 1,379 random protein-coding genes:
| quantity | value |
|---|---|
| slope | +0.0128 |
| intercept | -0.0050 |
| Pearson r | +0.065 |
| r squared | 0.0043 |
| residual SD | 0.0701 |
| raw SD | 0.0703 |
Pan-essentiality accounts for 0.4% of the variance in the difference statistic. Residual SD is 0.0701 against a raw SD of 0.0703 - subtracting the confound removes about a quarter of one percent of the spread. There is nothing there to correct for.
The slope is also positive, meaning more-essential genes show slightly less differential requirement. The objection predicts the opposite sign. Whatever is driving the ranking, it is not a pan-essentiality floor.
Residuals: the controls separate from the hits, in the right direction
Each probe's residual is its difference minus what essentiality predicts, expressed as a z against the random-gene residual distribution. Negative means selectively required in the rearranged lines beyond what general neediness explains.
| probe | pan-essentiality | residual z |
|---|---|---|
| MEN1 | -0.363 | -12.2 |
| FLT3 | -0.147 | -11.9 |
| PSIP1 (LEDGF) | 0.028 | -8.0 |
| KMT2A | -0.206 | -7.6 |
| DOT1L | -0.148 | -3.2 |
| DPY30 | -0.390 | -0.8 |
| ribosome (60S sample) | -2.268 | -1.9 |
| proteasome (20S) | -2.238 | 2.1 |
| GAPDH | -1.543 | 1.1 |
The pan-essential controls behave exactly as they should once essentiality is accounted for. The ribosome sample is the most essential thing in the table (-2.27) and its residual z is only -1.86; the proteasome is equally essential and comes out at +2.12, i.e. slightly less needed by the rearranged lines. GAPDH, profoundly essential, lands at +1.14. Meanwhile MEN1 is barely essential on average (-0.363) and residual z -12.2.
So the genes at the top of the list are not the essential ones. They are a different set, and MEN1 leads it.
What this does not establish
A single linear fit on a single confound is a crude instrument, and it is the only confound tested here. Ruling out pan-essentiality does not rule out lineage, co-occurring mutations, growth rate, culture medium or screen batch - DPY30 has already failed on other designs in this project even though it passes a permutation test, and this analysis does not rescue it (residual z -0.76).
A negative residual means not explained by essentiality. It does not mean caused by the fusion. And the sample is eleven rearranged lines, none of them infant-derived, which is the population the project actually cares about.
| script | analyses/depmap-essentiality-confound/run.py |
|---|---|
| status | executed |
| run date | 2026-09-19 |
| seed | 20260919 |
| n random genes | 1379 |
| n kmt2a rearranged | 11 |
| n other leukemia | 43 |
| slope | 0.01284 |
| intercept | -0.00499 |
| pearson r | 0.0652 |
| r squared | 0.0043 |
| residual sd | 0.0701 |
| raw sd | 0.0703 |
| headline | Pan-essentiality explains 0.4% of the variance in the difference statistic, with a positive slope - the opposite sign to the objection. MEN1 residual z = -12.2. |
| control problem | The pan-essential controls behave correctly: ribosome -1.86, proteasome +2.12, GAPDH +1.14. DPY30 remains unrescued at -0.76, consistent with its failures on the within-subtype, line-normalised and variance-matched designs. |
| caveat | A linear fit on one confound is crude, and essentiality is not the only thing differing between the groups. A negative residual means 'not explained by essentiality', not 'caused by the fusion'. Eleven rearranged lines, none infant-derived. |
residuals — the numbers behind the chart
{
"MEN1": {
"difference": -0.8656,
"pan_essentiality": -0.3631,
"residual": -0.8559,
"residual_z": -12.2,
"kind": "gene"
},
"FLT3": {
"difference": -0.8425,
"pan_essentiality": -0.1466,
"residual": -0.8356,
"residual_z": -11.91,
"kind": "gene"
},
"CCND3": {
"difference": -0.623,
"pan_essentiality": -0.3479,
"residual": -0.6136,
"residual_z": -8.75,
"kind": "gene"
},
"PSIP1": {
"difference": -0.5645,
"pan_essentiality": 0.0278,
"residual": -0.5599,
"residual_z": -7.98,
"kind": "gene"
},
"MEF2C": {
"difference": -0.5601,
"pan_essentiality": -0.0385,
"residual": -0.5546,
"residual_z": -7.91,
"kind": "gene"
},
"KMT2A": {
"difference": -0.5411,
"pan_essentiality": -0.2062,
"residual": -0.5335,
"residual_z": -7.61,
"kind": "gene"
},
"MYB": {
"difference": -0.4746,
"pan_essentiality": -0.384,
"residual": -0.4647,
"residual_z": -6.63,
"kind": "gene"
},
"ZFP64": {
"difference": -0.4098,
"pan_essentiality": -0.0051,
"residual": -0.4047,
"residual_z": -5.77,
"kind": "gene"
},
"BCL2": {
"difference": -0.3959,
"pan_essentiality": -0.0362,
"residual": -0.3904,
"residual_z": -5.57,
"kind": "gene"
},
"MEIS1": {
"difference": -0.3056,
"pan_essentiality": -0.053,
"residual": -0.2999,
"residual_z": -4.28,
"kind": "gene"
},
"KAT7": {
"difference": -0.2963,
"pan_essentiality": -0.2999,
"residual": -0.2875,
"residual_z": -4.1,
"kind": "gene"
},
"DOT1L": {
"difference": -0.231,
"pan_essentiality": -0.148,
"residual": -0.2241,
"residual_z": -3.2,
"kind": "gene"
},
"ACTB": {
"difference": -0.1818,
"pan_essentiality": -0.8145,
"residual": -0.1663,
"residual_z": -2.37,
"kind": "gene"
},
"Mediator": {
"difference": -0.1742,
"pan_essentiality": -0.783,
"residual": -0.1592,
"residual_z": -2.27,
"kind": "complex"
},
"MLLT1": {
"difference": -0.1492,
"pan_essentiality": -0.0065,
"residual": -0.1441,
"residual_z": -2.06,
"kind": "gene"
},
"ribosome (60S sample)": {
"difference": -0.1647,
"pan_essentiality": -2.2685,
"residual": -0.1306,
"residual_z": -1.86,
"kind": "complex"
},
"SEC": {
"difference": -0.0947,
"pan_essentiality": -0.3879,
"residual": -0.0847,
"residual_z": -1.21,
"kind": "complex"
},
"DPY30": {
"difference": -0.0631,
"pan_essentiality": -0.3896,
"residual": -0.0531,
"residual_z": -0.76,
"kind": "gene"
},
"HOXA9": {
"difference": -0.0479,
"pan_essentiality": -0.1725,
"residual": -0.0407,
"residual_z": -0.58,
"kind": "gene"
},
"PAF1C": {
"difference": -0.0232,
"pan_essentiality": -0.7417,
"residual": -0.0087,
"residual_z": -0.12,
"kind": "complex"
},
"GAPDH": {
"difference": 0.055,
"pan_essentiality": -1.5426,
"residual": 0.0798,
"residual_z": 1.14,
"kind": "gene"
},
"proteasome (20S)": {
"difference": 0.1152,
"pan_essentiality": -2.2379,
"residual": 0.1489,
"residual_z": 2.12,
"kind": "complex"
}
}- Refuting the loudest objection is not the same as validating the hit list. Essentiality was the easiest confound to test and it fell; the harder ones - lineage and co-occurring mutation - are still open, and a reviewer who has read this will ask for them next.
- An r squared of 0.004 is such a complete null that it is worth asking whether the regression was set up to be able to fail. It can: the same framework puts the 20S proteasome at +2.12 and GAPDH at +1.14, which is the confound visibly doing work on the genes where it should.
- Does the result hold in infant-derived KMT2A-rearranged lines, of which this sample contains none?
- Which confound does explain the residual spread, given essentiality explains almost none of it - lineage, co-mutation, or proliferation rate?
- Why is the slope positive? A mild ceiling effect on already-lethal genes would do it, but that is a guess and it is testable.
- E1strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.database
Across 1,379 random protein-coding genes, pan-essentiality explains 0.43% of the variance in the KMT2A-rearranged-versus-other difference statistic (r = +0.065, slope +0.0128, residual SD 0.0701 vs raw SD 0.0703). MEN1's essentiality-adjusted residual is z = -12.2, the most extreme of 22 probes, while the ribosome sample reaches only -1.86 and the 20S proteasome is +2.12.
DepMap 24Q2 CRISPR (Chronos) gene effect, retrieved via the Open Targets GraphQL API · Broad Institute DepMap / Open Targets Platform 2026What this evidence cannot show
Computed by the coordinator on 2026-09-19; script and full output are in analyses/depmap-essentiality-confound/. Eleven KMT2A-rearranged lines against 43 other leukaemia lines, rearrangement status from Cellosaurus. A linear fit on one confound is crude and essentiality is not the only thing differing between the groups. None of the eleven rearranged lines is infant-derived.
Evidence that argues against this (1)
- This refutes the pan-essentiality objection specifically, not the general class of confounding. The same project's within-subtype, line-normalised and variance-matched designs disagree with the permutation test about DPY30, and this analysis leaves that disagreement exactly where it was: DPY30's residual z is -0.76, comfortably inside the noise. A control that clears one confound and fails three others has not been vindicated. Broad Institute DepMap / Open Targets Platform 2026
- ← related toRUNX1::RUNX1T1 (AML1-ETO), t(8;21)(q22;q22)
Physiology confound (Goardon 2006, PMID 16407974): ETO2/CBFA2T3 is a core component of the normal erythroid TAL-1/SCL complex whose stoichiometry gates terminal maturation, so the shared ETO module is recurrent physiology, not a leukaemia-specific signature.
Same comparator arm.
Tests and refutes the pan-essentiality objection to this scan's ranking: essentiality explains r^2=0.004 of the difference statistic, with a positive slope.
- supports →The whole protein-coding genome: menin is still first, and a metabolic pathway appears (run)
The full 17,728-gene scan's top hits are not explained by general essentiality; the confound accounts for 0.4% of the variance.
MEN1's selective requirement survives adjustment for pan-essentiality: residual z = -12.2, the most extreme of 22 probes, despite MEN1 being only mildly essential on average (-0.363).
Supplies the essentiality adjustment the complex-level scoring lacked: the ribosome sample falls to residual z -1.86 and the 20S proteasome to +2.12 once essentiality is removed.
- refutes →DPY30 — the smallest COMPASS subunit, and the only one that behaves like a selective dependency
Does NOT rescue DPY30. Adjusted for essentiality its residual z is -0.76, inside the noise, consistent with its failure on the within-subtype, line-normalised and variance-matched designs.
- related to →Six times the null pool, and the criterion I had been using turns out to be the problem (run)
A different control on the same ranking: the deep null pool addresses sampling, this addresses a named biological confound.
Companion confound test. Within-subtype centring removes lineage; this removes general essentiality.
the U2 clause here tests the same confound that entry documents: common-essential genes correlate with everything