Oncofusion

start / graph / ana-prc2-dependency-lineage-split

Analysismoderate confidence 3 of 5: One strong study, or several smaller consistent ones.Q41-prc2-arm

The PRC2 dependency profile, lineage-split and expression-controlled - and the first time this project's paralogue rule came back positive

also called PRC2 DepMap panel · EZH1/EZH2 redundancy signature · RBBP4/RBBP7 expression control

We scored every part of the Polycomb machine against a large public database of which genes cancer cells need, split the leukaemia lines by lineage, and then checked whether the parts that looked unnecessary were actually being made at all. The pattern exactly matches what the mouse genetics predicted: the interchangeable parts look unnecessary, the irreplaceable ones do not.

The same thing, for a scientist

Run 2026-09-20 for the Q41 Polycomb beat. **Method:** 24 genes (PRC2 core, PRC2.1/PRC2.2 accessory, PRC1 and control genes) scored on DepMap Chronos gene effect via Open Targets, index arm = 11 Cellosaurus-annotated KMT2A-rearranged DepMap leukaemia lines, comparator = 43 other leukaemia lines, index arm split 7 myeloid : 4 B-lineage and each half compared to the matching half of the comparator. Raw effect sizes only - no z, no division by comparator SD, no permutation null - per this project's three DepMap corrections. A second script checked Human Protein Atlas cell-line RNA for the seven panel lines HPA carries. **Result 1, the redundancy signature.** EZH2 -0.201 and EZH1 -0.100 are both non-dependencies with differences of +0.004 and +0.002; EED is -0.458 (difference **-0.117**, rank **824 of 17,728**), SUZ12 -0.482, RBBP4 -1.486. **Neither catalytic paralogue is required alone while the shared, non-paralogous subunits are** - the exact CRISPR signature of the EZH1-compensation result from mouse MLL-AF9 genetics. **Result 2, the expression control passes.** EZH1 is transcribed in 7/7 panel lines (median 13.7 nTPM) and **RBBP7 at median 139.0 nTPM, higher than RBBP4's 77.5 in every line**, while RBBP7's gene effect is -0.013 against RBBP4's -1.486. **This is the first time this project's paralogue-expression control has returned a positive**, in both cases: the low scorers are transcribed, so the asymmetries are functional. **Result 3, the lineage split does not show the usual myeloid artefact.** EED's difference is -0.058 myeloid against **-0.224 lymphoid**, and EZH2's index mean is -0.112 myeloid against -0.355 lymphoid. Four lymphoid lines, so this is weak - but it is the first epigenetic panel in this project where the split pointed toward the index lineage rather than away from it.

Go deeper
The full table

Index = 11 KMT2A-rearranged DepMap lines (ALL-PO, HB11;19, KOPN-8, Karpas-45, ML-2, MOLM-13, MOLM-14, MV4-11, Mono-Mac-6, NOMO-1, OCI-AML-2, RS4;11, SEM, SEMK2, SHI-1, THP-1 were returned by Cellosaurus; 11 have DepMap screens). Comparator = 43 other leukaemia lines. diff = index minus comparator; negative = more required in the rearranged arm.

genediffindexcomparatorpan-essentialitymyeloid difflymphoid diff
MEN1-0.866-1.263-0.397-0.363-0.890-0.833
KMT2A-0.541-0.811-0.270-0.206-0.581-0.463
DOT1L-0.231-0.604-0.373-0.148-0.257-0.212
EED-0.117-0.458-0.340-0.277-0.058-0.224
MTF2-0.072-0.097-0.025-0.086-0.042-0.125
CDKN1C-0.063-0.013+0.050+0.073-0.073-0.063
PHF1-0.047-0.096-0.049-0.069+0.029-0.159
LCOR-0.040-0.069-0.029+0.001-0.030-0.061
SUZ12-0.030-0.482-0.452-0.405-0.116+0.105
BMI1-0.029-0.036-0.007+0.059+0.005-0.080
PHF19-0.027+0.045+0.072+0.045+0.013-0.089
RBBP4-0.024-1.486-1.462-1.347-0.007-0.047
JARID2-0.022-0.046-0.024-0.008-0.012-0.043
CBX8-0.005+0.068+0.072+0.101-0.049+0.067
EZH1+0.002-0.100-0.102-0.134+0.008+0.003
RBBP7+0.003-0.013-0.016-0.1550.000+0.005
EZH2+0.004-0.201-0.204-0.049+0.062-0.109
AEBP2+0.014-0.044-0.058-0.035-0.024+0.084
KDM6B+0.023-0.063-0.086-0.039+0.013+0.039
EPOP+0.050-0.367-0.417-0.549+0.120-0.061
RNF2+0.096-0.028-0.125-0.047+0.108+0.075
KDM6A+0.143+0.239+0.097+0.096+0.110+0.214
PCGF1+0.145+0.059-0.086-0.038+0.232-0.008
CDKN2A+0.177+0.274+0.097+0.160+0.308-0.034

PALI1 did not map to an HGNC-derived Ensembl identifier and was scored as LCOR instead; that substitution is not exact and no claim rests on it.

MEN1 at -0.866 is the internal positive control and it behaves, which is the reason to take the shape of the rest seriously even though none of the PRC2 numbers is a hit.

Four things this analysis is careful not to say
  1. No selectivity claim for any PRC2 gene. EED's -0.117 is below this project's -0.5 convention. EED is described as 'the best-scoring PRC2 subunit', which is a within-panel ranking, not a hit.
  2. RBBP4 carries no selectivity reading at all. Pan-essentiality -1.347 is at the Chronos floor, so its -0.024 difference is distance-to-floor and is uninterpretable rather than small - the HCFC1/KAT8 caution.
  3. The catalytic nulls say nothing about PRC2 catalysis. A single-gene knockout screen cannot detect a redundant pair. EZH2 at +0.004 is evidence about EZH2 alone; the mouse genetics that used a combined Ezh1/Ezh2 arm reached the opposite conclusion about the complex, and the two are not in conflict.
  4. CDKN2A's +0.177 is probably not biology. CDKN2A is deleted in a large fraction of leukaemia lines and a Chronos score for a deleted gene is not a dependency measurement. It is in the table because it is the published mediator of PRC2 loss and a reader will look for it, not because the number means anything.

And one thing it does say that was not expected: the lineage split failed to reveal a myeloid artefact. Every previous epigenetics panel in this project that was split this way found the index signal tracking myeloid identity (CEBPA at rank 25, SPI1 at 43, the monocytic haem-pathway confound). Here the EED and EZH2 signals are larger in the B-lineage half. With four lines that is likely noise, and it is recorded rather than suppressed because it agrees in direction with the two index-disease pharmacology results and because an unrecorded observation cannot be checked later.

Facts and figures
scriptanalyses/prc2-dependency-lineage-split/run.py
statusrun
n index lines11
n comparator lines43
n index myeloid7
n index lymphoid4
genes scored24
result summaryEED is the best-scoring PRC2 subunit (-0.117, rank 824/17,728); neither EZH2 (+0.004) nor EZH1 (+0.002) is a dependency, which is the CRISPR signature of paralogue redundancy and matches the Ezh2-null/EZH1-compensation genetics; RBBP4 is pan-essential and floor-compressed; RBBP7 is dispensable DESPITE being the more abundant transcript in 7/7 lines, so the paralogue-expression control passes; the lineage split shows the EED and EZH2 signals larger in the B-lineage half, which is the opposite of this project's usual finding and rests on four lines.
inputs
[
  "DepMap CRISPR (Chronos) gene effect via Open Targets GraphQL",
  "Cellosaurus KMT2A fusion annotation (var:KMT2A AND dr:DepMap)",
  "HGNC complete set for symbol-to-Ensembl mapping",
  "Human Protein Atlas rna_celline.tsv (downloaded 2026-09-20)",
  "analyses/depmap-kmt2a-selective-scan-full/results.json for genome-wide ranks"
]
outputs
[
  "analyses/prc2-dependency-lineage-split/results.json",
  "analyses/prc2-dependency-lineage-split/expression_control.json"
]
conventions applied
[
  "raw effect sizes only; no z, no division by comparator standard deviation",
  "index arm split by lineage and compared to the matching comparator half",
  "pan-essentiality reported for every gene; no selectivity claim above the project's -0.5 convention",
  "paralogue-expression control run against an independent resource (HPA) rather than DepMap's own expression"
]
supports
[
  "gene-ezh2",
  "gene-ezh1",
  "gene-eed",
  "gene-suz12",
  "gene-rbbp4",
  "gene-rbbp7",
  "protein-prc2-complex",
  "hyp-prc2-is-the-fusions-repressive-arm"
]
How this could bite us back
  • The headline of this analysis - neither EZH2 nor EZH1 is a dependency - is the number most likely to be quoted out of context as 'PRC2 is not a target in this disease'. It is not that claim. A single-gene screen cannot see a redundant pair, and the complex-level subunits in the same table ARE required.
  • The lymphoid lean is four cell lines and will be over-read if it is repeated without its n. It is recorded at confidence weak for that reason.
Still open
  1. Is there any screen anywhere containing an EZH1/EZH2 DOUBLE knockout in a KMT2A-rearranged line? Every number in this analysis is a single-gene knockout, which is the one design guaranteed to miss the pair that matters, and a combinatorial screen would convert the strongest inference in this beat into a measurement.
  2. Would a permutation null change the EED result? None was computed here. The project's sibling-triage experience is that shortlists do not survive a 50,000-label permutation, so EED's -0.117 should be tested before it is treated as anything more than consistent with the genetics.
  3. Is RBBP7 protein in PRC2 in these lines? The expression control shows the transcript is abundant; only a proteomic measurement of complex composition would show whether the dispensability is because RBBP7 is absent from PRC2 or present and inert - and those imply different resistance risks.
Every source behind this entry (3)
  1. E1moderate confidence 3 of 5: One strong study, or several smaller consistent ones.dataset

    Across 11 Cellosaurus-annotated KMT2A-rearranged DepMap leukaemia lines and 43 other leukaemia lines, EED (index -0.458, difference -0.117, rank 824 of 17,728) and SUZ12 (-0.482) are required while EZH2 (-0.201, difference +0.004) and EZH1 (-0.100, +0.002) are not; RBBP4 is pan-essential (-1.486 index, -1.347 across 1,208 lines) and RBBP7 is dispensable (-0.013); all PRC2.1/PRC2.2 accessory subunits are mid-genome; MEN1 at -0.866 is rank 1 as an internal control.

    What this evidence cannot show

    11 index lines, none infant-derived and none a true infant KMT2A::AFF1 model; the lymphoid half is 4 lines. Cellosaurus fusion annotation may be incomplete, biasing toward the null. DepMap release version is not stated by the Open Targets API. Raw effect sizes only - no z, no permutation null - so nothing here is a hit and the comparator-arm ranks are taken from the project's existing 17,728-gene scan rather than recomputed. Single-gene knockouts systematically under-score every paralogue pair in the panel (EZH1/EZH2, RBBP4/RBBP7, MTF2/PHF19/PHF1). PALI1 did not map and LCOR was substituted.

    Evidence that argues against this (1)
    • The published genetics reach the opposite conclusion about PRC2 catalysis using a combined arm: shRNA suppression of Eed, Suz12 OR Ezh1/Ezh2 together arrested proliferation and forced differentiation in MLL-AF9;Nras-G12D AML. A DepMap profile in which no catalytic subunit scores is fully compatible with PRC2 catalysis being required, and this analysis must not be cited as evidence against it. Shi J et al. 2013
  2. E2moderate confidence 3 of 5: One strong study, or several smaller consistent ones.dataset

    In the seven KMT2A-rearranged lines Human Protein Atlas carries (SEM, RS4;11, MV4-11, THP-1, NOMO-1, MOLM-13, OCI-AML-2), EZH1 is detected in 7/7 at median 13.7 nTPM and RBBP7 in 7/7 at median 139.0 nTPM - higher than RBBP4's 77.5 in every line - so the low CRISPR scores of both are functional non-requirement rather than absence of transcript, and the paralogue-expression control this project requires is satisfied.

    What this evidence cannot show

    nTPM is bulk RNA from immortalised cell lines: it is not protein, not complex occupancy, and cannot show whether EZH1 or RBBP7 protein is incorporated into PRC2. HPA carries 7 of the 11 index lines (SHI-1 and HB11;19 absent; SEMK2 and MOLM-14 are the same patients as SEM and MOLM-13). Detection threshold 1.0 nTPM. EPOP was the only gene below threshold in any line (OCI-AML-2, 0.4).

    Evidence that argues against this (1)
    • Expression rules out one artefact and proves nothing positive. A transcribed methyltransferase can still be excluded from the relevant loci, and the compensation reported in Ezh2-null MLL-AF9 leukaemia was LOCUS-SPECIFIC rather than global - so EZH1 may cover only part of the PRC2 landscape, and the residual pattern after an EZH2-selective inhibitor need not resemble the untreated one. Neff T et al. 2012
  3. E3weak confidence 2 of 5: Preliminary, small, indirect, or not yet peer reviewed.dataset

    Splitting the index arm 7 myeloid : 4 B-lineage and comparing each half to the matching half of the comparator, EED's difference is -0.058 in the myeloid half and -0.224 in the lymphoid half, and EZH2's index mean is -0.112 myeloid against -0.355 lymphoid - the first epigenetic panel in this project whose lineage split did not attribute the index signal to myeloid identity.

    What this evidence cannot show

    **Four lymphoid lines.** No statistical test was applied and none would be meaningful at that n. No infant KMT2A::AFF1 line exists in DepMap, so the lymphoid half is paediatric or adult B-ALL. Graded weak deliberately: this is an observation recorded so it can be checked, not a finding.

    Evidence that argues against this (1)
    • This project's lineage splits have previously existed to KILL apparent findings - CEBPA at rank 25 and SPI1 at 43 were both artefacts that moved +0.15 and +0.92 in the B-lineage half, and a monocytic confound explained the haem-pathway signal. A split that for once supports the index lineage is exactly the result most likely to be over-read, and four lines cannot support it. Q37-histone-chaperones 2026
Connected to