Scoring complexes instead of genes, and the controls disagree with each other (run)
also called complex-level selectivity · PAF1C flat · ribosome vs proteasome control
Every dependency test here has scored one gene at a time, which is the wrong unit for a disease about protein machines and is blind to genes that cover for each other. Scoring whole machines instead gives a different picture — and two control machines that should have behaved the same way behaved oppositely, which undermines an assumption this project had been making.
The same thing, for a scientist
Scores 12 complexes by the mean KMT2A-rearranged-versus-other difference over their subunits, against a 732-gene random background. PAF1C — which this project has repeatedly asked about as a second druggable handle — is **flat** (mean -0.023, 36th percentile). Mediator is the strongest non-trivial signal (31 subunits, 87% leaning the same way, z = -2.18). The decisive result is in the controls: the ribosome sample scores -0.165 and the proteasome **+0.115**, so 'pan-essential genes score negative because the panel is more sensitive' is not a sufficient explanation, and the interpretation this project applied to the genome-wide scan's top 20 needs qualifying.
Why the unit was wrong
Single-gene ranks are the wrong grain for this project twice over.
It is a project about complexes. The phase-2 register asks 'is PAF1C a handle?', and no single-gene rank answers that. And single-gene knockout is blind to paralogue buffering — HOXA9 ranks 3,649 of 17,728 because HOXA10 covers for it. A complex whose subunits are individually buffered can still be required as a unit, and every subunit rank will look unremarkable.
So each subunit was scored on the same difference statistic used throughout, and the complex summarised by the mean over its scored subunits, read against a 732-gene random background 1.
The result that matters for the target register1 deeper
| Complex | subunits | mean difference | % of subunits negative | percentile | z |
|---|---|---|---|---|---|
| COMPASS/WRAD + KMT2A | 7 | -0.238 | 56% | 1.2 | -3.0 |
| Mediator | 31 | -0.174 | 87% | 1.9 | -2.18 |
| CONTROL ribosome | 11 | -0.165 | 91% | 2.2 | -2.06 |
| SAGA | 14 | -0.123 | 86% | 4.6 | -1.53 |
| Cohesin | 8 | -0.117 | 62% | 5.2 | -1.44 |
| SEC | 8 | -0.095 | 75% | 7.8 | -1.16 |
| CONTROL olfactory receptors | 7 | -0.045 | 86% | 21.5 | -0.52 |
| DOT1L complex | 5 | -0.040 | 40% | 25.1 | -0.45 |
| PRC2 | 5 | -0.033 | 60% | 28.8 | -0.36 |
| PAF1C | 6 | -0.023 | 67% | 35.8 | -0.24 |
| PRC1 | 7 | +0.026 | 43% | 68.6 | +0.40 |
| CONTROL proteasome | 12 | +0.115 | 17% | 97.1 | +1.55 |
PAF1C is flat. This project has asked three times whether PAF1C is the second druggable handle on the retained fragment, and the complex-level answer is that it shows no selective requirement at all — 36th percentile, indistinguishable from a random gene. That is consistent with the literature finding recorded elsewhere in this graph that the field disagrees even on the sign of PAF1 perturbation, and with two of its six named subunits having disputed membership. It does not support PAF1C as a handle.
COMPASS/WRAD tops the table but is not what it looks like. That set contains KMT2A, MEN1 and PSIP1, which are ranks 1, 7 and 11 of the genome individually. The complex mean is carried by them, not by WRAD. This is the failure mode the caveat predicts: a complex mean assumes subunits are exchangeable and they are not.
Mediator is the most interesting genuinely new signal — 31 subunits, 87% leaning the same way, which is harder to produce by a few outliers than a 5-subunit complex is. The genome-wide scan independently put MED13L at rank 9 and MED23 at 16, and MED26 is already in this graph as a SEC link. Mediator may be under-represented in this project's interactome.
The controls disagree, and that breaks an assumption I was using
The two pan-essential controls were included to measure the same artefact and point in opposite directions: the ribosome sample at -0.165 (2nd percentile, 91% of subunits negative) and the proteasome at +0.115 (97th percentile, only 17% negative). Both are about equally pan-essential (mean effect -2.27 and -2.24).
This matters because of how this project read the genome-wide scan. That entry says eight ribosomal and core-machinery genes in the top 20 are "more depleted in rearranged lines because those lines are more sensitive overall, not because of the fusion". If that were the whole story, the proteasome would be depleted too. It is not. A uniform panel-sensitivity offset cannot produce opposite signs in two equally essential complexes.
So one of two things is true, and this analysis cannot distinguish them:
- the artefact is real but not uniform, acting on some essential complexes and not others — in which case 'pan-essential therefore artefact' is an unreliable filter and the genome-wide scan's discard list needs re-examining; or
- ribosome dependence is genuinely elevated in KMT2A-rearranged lines, which would make it biology rather than noise.
The second is not idle. This same cycle established that WDR5 WIN-site inhibitors act in MLL-rearranged cells by suppressing ribosome protein gene transcription and inducing nucleolar stress. Two independent lines arriving at the ribosome in one cycle is worth a look — but a convergence is not evidence, and 11 lines with 11 ribosomal genes is thin.
The olfactory-receptor negative control behaved acceptably (21st percentile, z = -0.52) though with 86% of its seven members nominally negative, which at n = 7 is noise.
What a complex mean cannot do
It assumes subunits are exchangeable. They are not, and COMPASS/WRAD shows how badly that fails — three members carry the whole score. A high complex mean should always be decomposed before it is believed.
Membership is a choice. The lists here are conventional, and this graph separately records that two named PAF1C subunits, RTF1 and LEO1, are not stably part of the human complex. Changing the membership list changes the answer.
Small complexes are noisy. PRC2 and DotCom have five scored subunits each; their positions should not be read as meaningful.
Eleven rearranged lines, none infant-derived, as in every dependency conclusion in this project.
| script | analyses/depmap-complex-level-selectivity/run.py |
|---|---|
| status | run |
| run date | 2026-09-19 |
| seed | 20260919 |
| n kmt2a rearranged | 11 |
| n other leukemia | 43 |
| null n | 732 |
| null mean | -0.005 |
| null sd | 0.0774 |
| headline | PAF1C shows no complex-level selective requirement (36th percentile) |
| control problem | ribosome and proteasome are equally pan-essential and score in opposite directions, so a uniform panel-sensitivity offset cannot be the whole explanation for pan-essential genes scoring negative in the genome-wide scan |
| figure | analyses/depmap-complex-level-selectivity/results.md |
complex means — the numbers behind the chart
{
"COMPASS/WRAD+KMT2A": -0.2375,
"Mediator": -0.1742,
"CONTROL_ribosome": -0.1647,
"SAGA": -0.1231,
"Cohesin": -0.1169,
"SEC": -0.0947,
"CONTROL_olfactory": -0.0454,
"DotCom": -0.0395,
"PRC2": -0.0329,
"PAF1C": -0.0232,
"PRC1": 0.0256,
"CONTROL_proteasome": 0.1152
}- The most dangerous use of this table is to read COMPASS/WRAD's top rank as evidence for WRAD. It is carried by KMT2A, MEN1 and PSIP1, three genes already known to rank at the top individually, and says nothing new.
- A flat complex mean is not a clearance. PAF1C shows no complex-level signal; that is different from saying no PAF1C subunit could ever be a target, and the register should record the narrower claim.
- This analysis was built to test complexes and its most useful output is a problem with its own controls. That is worth more than the ranking, and it retroactively weakens a reading this project applied to an earlier scan.
- Ribosome and proteasome score oppositely despite equal pan-essentiality. Is the panel-sensitivity artefact non-uniform, or is ribosome dependence genuinely elevated in KMT2A-rearranged lines? A per-line normalisation of the kind already used elsewhere in this project would separate them.
- WDR5 WIN-site inhibitors act through ribosome protein gene suppression, and the ribosome scores selectively required here. Convergence or coincidence? Both rest on small samples.
- Mediator has 31 subunits with 87% leaning the same way, and MED13L and MED23 rank 9 and 16 genome-wide. Is Mediator under-represented in this project's interactome, and is any Mediator surface tractable?
- The genome-wide scan discarded eight pan-essential genes from its top 20 on an explanation this analysis undermines. Should that discard list be re-examined?
- PAF1C is flat at complex level, but the membership list includes two subunits this graph says are not stably part of the human complex. Does the result change with a four-subunit core?
- E1moderate confidence 3 of 5: One strong study, or several smaller consistent ones.dataset
Scoring 12 complexes by the mean difference in DepMap gene effect between 11 KMT2A-rearranged and 43 other leukemia lines, against a 732-gene random background (mean -0.0050, SD 0.0774): COMPASS/WRAD+KMT2A -0.238 (z -3.0), Mediator -0.174 over 31 subunits with 87% negative (z -2.18), SAGA -0.123, cohesin -0.117, SEC -0.095, DOT1L complex -0.040, PRC2 -0.033, PAF1C -0.023 (36th percentile), PRC1 +0.026. Controls: ribosome sample -0.165 (2nd percentile), proteasome +0.115 (97th percentile), olfactory receptors -0.045 (21st percentile). Mean pan-essentiality is -2.27 for the ribosome sample and -2.24 for the proteasome.
DepMap CRISPR (Chronos) gene effect via the Open Targets Platform GraphQL API, grouped by conventional complex membership and split by Cellosaurus KMT2A gene-fusion annotation; analysis in analyses/depmap-complex-level-selectivity/ · Broad Institute DepMap; Cellosaurus (SIB); HGNC; Open Targets Platform; analysis run by this project · Databases and derived analysis 2026What this evidence cannot show
Our own computation, not peer reviewed. A complex mean assumes subunits are exchangeable, which they are not: COMPASS/WRAD's rank is carried by KMT2A, MEN1 and PSIP1 rather than by WRAD. Membership lists are the conventional ones, and this graph separately records RTF1 and LEO1 as not stably part of human PAF1C, so the PAF1C result depends on a contested list. Complexes with five scored subunits are noisy. Eleven rearranged lines, none infant-derived. DepMap release version not stated by the API.
Evidence that argues against this (2)
- The two pan-essential controls point in opposite directions - ribosome -0.165, proteasome +0.115 - despite near-identical pan-essentiality. That is not consistent with a uniform panel-sensitivity offset, which is the explanation this project applied to the pan-essential genes in the genome-wide scan's top 20. Either the artefact is non-uniform, in which case 'pan-essential therefore artefact' is an unreliable filter, or ribosome dependence is genuinely elevated here. This analysis cannot distinguish them and should not be cited as if it had. analysis run by this project 2026
- A flat complex mean does not exonerate every subunit. PAF1C scores 36th percentile overall, but the question the register asks is whether any single PAF1C surface is druggable, and a mean over six subunits can hide one that matters. The correct reading is that PAF1C shows no COMPLEX-level selective requirement, not that no PAF1C subunit could be a target. analysis run by this project 2026
- refutes →The whole protein-coding genome: menin is still first, and a metabolic pathway appears (run)
its pan-essential discards were explained as panel sensitivity, but ribosome and proteasome score oppositely at equal essentiality
PAF1C shows no complex-level selective requirement in KMT2A-rearranged lines (36th percentile)
Mediator is the strongest non-trivial complex signal, 31 subunits with 87% concordant
built to address exactly that: a complex whose subunits are individually buffered can still be required as a unit
answers the register's recurring PAF1C question at complex level, and narrows rather than clears it
- related to →WDR5 — the partner that is not a partner: it binds the half of KMT2A the fusion throws away
WIN-site inhibitors act via ribosome protein gene suppression, and the ribosome scores selectively required here; convergence, not evidence
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.