The MLL-fusion interactome recovers corepressor SCAFFOLDS and not their HDAC1/2 catalytic core: NuRD splits 5/7 remodelling arm versus 0/7 deacetylase arm (Fisher p = 0.021), and HDAC1, HDAC2, RBBP4 and RBBP7 are absent from all seven baits
also called NuRD module split in Skucha 2018 · Q193 module-split analysis
NuRD is a two-part machine: a motor that slides nucleosomes and a separate deacetylase that strips acetyl marks off histones. Sorting the published MLL-fusion pulldown by which part each protein belongs to shows that the motor half is well represented and the deacetylase half is completely missing - and the same is true of the rival Sin3 complex, whose scaffold shows up while its deacetylase does not.
The same thing, for a scientist
The existing corpus analysis `analyses/mll-fusion-interactome-complexes` scored NuRD as one 14-member set and reported 5 of 14 present in the Skucha 2018 960-protein network (CHD3, CHD4, GATAD2A, GATAD2B, MBD2; p = 3.5e-4). NuRD is not one object: it assembles from two biochemically separable sub-modules that can be produced independently and are catalytically active on their own [E4]. Re-scored by module, all five hits fall in the remodelling arm (CHD3/4/5, GATAD2A/B, MBD2/3) at 5 of 7, and the deacetylase arm (MTA1/2/3, HDAC1/2, RBBP4/7) is 0 of 7. Fisher exact on that 2x2 gives one-sided p = 0.0105, two-sided p = 0.0210 - a test that needs no background choice, which is why it and not the hypergeometric p-value is the headline [E1]. The Sin3/HDAC complex, which the earlier analysis never tested even though it shares HDAC1/2 and RBBP4/7 with NuRD [E2], scores 4 of 17 (SIN3A, SUDS3, ARID4B, ING2) - and all four are Sin3-SPECIFIC subunits, with the shared HDAC1/2 + RBBP4/7 core at 0 of 4. So the pattern is not about NuRD: across two different HDAC1/2 corepressor complexes this dataset recovers the scaffolds and never the catalytic core. In the MLL-AF4 (= KMT2A::AFF1) column specifically the count is 1 (CHD4) of the remodelling arm, 0 of the deacetylase arm and 0 of Sin3. The controls behave as in the earlier analysis: the SEC positive control is 10 of 16 and the CCT/TRiC contaminant negative control is 8 of 9, so complex-wide enrichment on this table is not by itself evidence of specific binding.
The measurement
Run on the cached Skucha Supplementary Table 5 (MOESM4) on 2026-09-24, with load-time assertions that the sheet has the seven expected bait columns, gene_name/uniprot_id in columns 0/1, neg_ctrl in column 17 and exactly 960 data rows 1.
| set | found | of | in the MLL-AF4 column | absent from the 960 |
|---|---|---|---|---|
| NuRD remodelling module | 5 | 7 | 1 (CHD4) | CHD5, MBD3 |
| NuRD deacetylase module | 0 | 7 | 0 | HDAC1, HDAC2, MTA1, MTA2, MTA3, RBBP4, RBBP7 |
| Sin3/HDAC, all subunits | 4 | 17 | 0 | 13, incl. SIN3B, SAP30, SAP18 |
| Sin3/HDAC, Sin3-specific only | 4 | 13 | 0 | 9 |
| shared HDAC1/2 core (both complexes) | 0 | 4 | 0 | HDAC1, HDAC2, RBBP4, RBBP7 |
| CONTROL negative, CCT/TRiC | 8 | 9 | 8 | CCT6B |
| CONTROL positive, SEC | 10 | 16 | 10 | EAF1/2, ELL2/3, ICE1/2 |
The found members, with UniProt accession, number of the seven baits, negative-control spectra and per-bait p and spectral counts, are persisted in results.json (rule 90: the identities, not only the count).
Per-bait retained counts are 291-293 proteins, so "absent from the 960" means outside the top ~300 per bait by the paper's own combined p-value, not scored as zero.
Why the module split is the result and the p-values are not
Three backgrounds are reported (proteome 20,431; nuclear 5,754; recovered 4,600) because the honest denominator - the ~4,600 proteins the study recovered before filtering - is not in the table, so every hypergeometric p-value here is an upper bound on significance. The Fisher test on the module 2x2 needs no background at all: it compares seven subunits against seven subunits inside the same dataset, so a wrong denominator cannot move it. 5/7 versus 0/7, two-sided p = 0.0210 1.
This also matters for reading the earlier pooled figure. analyses/mll-fusion-interactome-complexes is not wrong - 5 of 14 is 5 of 14 - but a single number over a set that contains two separable machines cannot express which machine the five came from, and here they all came from one.
The reading I resisted, and the caveat that carries the weight
The tempting reading is mechanistic: MLL fusions engage the nucleosome-sliding arm of NuRD and leave the deacetylase behind, which would fit the standard picture of a fusion as an activator that has no use for HDAC1/2. The data cannot support that, for one specific reason. The table retains the top ~300 preys per bait ranked by a bait-specificity model, and HDAC1, HDAC2, RBBP4 and RBBP7 are among the most abundant and most promiscuous nuclear proteins in any lysate - RBBP4/7 alone are shared between NuRD, Sin3/HDAC, PRC2 and CAF-1. A statistical filter for bait specificity removes promiscuous abundant proteins first. So a 0 of 4 on the shared core is exactly what a specificity filter would produce whether or not the biology is there, and the module split inherits that: the deacetylase arm is the more abundant and more shared arm.
What survives that caveat is narrower and still worth having: within this dataset, no NuRD or Sin3 catalytic core subunit is recoverable, so no claim in this graph about deacetylase activity at a KMT2A fusion can be sourced to this dataset, and any future entry that reads "NuRD is in the pan-fusion core" is reading a scaffold signal. The corroborating detail is that CHD4 itself carries 2|2 negative-control spectra and was retained anyway, so the filter is not simply excluding everything with background.
What this does and does not say about KMT2A::AFF1
In the MLL-AF4 column the whole of NuRD is one protein, CHD4, and the whole of Sin3/HDAC is nothing. This reproduces, by an independent route, the correction already recorded inside gene-chd4: the 5-subunit NuRD signature is a union over seven baits, MLL-AF9 and MLL-EEN contribute 4 each, and MLL-AF4 contributes one. Reading the NuRD rows as a statement about the index fusion is the error that correction was written to stop, and this analysis agrees with it.
GATAD2A is the sharpest case. It is in the network - Q86YP4, one of seven baits, negative control 0|0, p = 0.0361 with 10|12 spectra - and that one bait is MLL-AF9, not MLL-AF4 1. So GATAD2A has been assayed against a KMT2A fusion and scored; it has not been shown to associate with the index fusion.
| script | analyses/q193-nurd-module-split/run.py |
|---|---|
| status | run |
| result summary | NuRD remodelling arm 5/7, deacetylase arm 0/7, Fisher two-sided p = 0.0210. Sin3/HDAC 4/17 (all Sin3-specific). Shared HDAC1/2 + RBBP4/7 core 0/4. MLL-AF4 column: 1, 0, 0. Controls unchanged: SEC 10/16, CCT/TRiC 8/9. |
| relationship to existing entries | EXTENDS analyses/mll-fusion-interactome-complexes and ana-interactome4-enumeration; does not replace either. The earlier analysis scored NuRD as one 14-member set (5/14, p = 3.5e-4) and never tested Sin3/HDAC at all. This one splits NuRD by sub-module and adds Sin3. It AGREES with the per-bait correction already stored inside gene-chd4 that the 5-subunit NuRD signature is a union over seven baits and the MLL-AF4 column contains CHD4 alone. |
| absence class | measured-not-reported |
| absence class note | For any single absent subunit - MTA2, MBD3, HDAC1, HDAC2, RBBP4, RBBP7, MTA1, MTA3, CHD5, SIN3B, SAP30, SAP18 - the mass spectrometer generated a per-bait spectral count and the paper tabulated only the top ~300 per bait, so the class is measured-not-reported. It upgrades to measured-not-analysed because the raw files are public at ProteomeXchange PXD009338 and a re-search would return the number. It is NOT never-measured, and an absence entry calling it that would overstate the cost of closing it by years. |
| shown for fusion | MLL-AF4 (= KMT2A::AFF1) plus MLL-AF1p, MLL-AF9, MLL-CBP, MLL-EEN, MLL-ENL, MLL-GAS7 - all seven Skucha baits, in Jurkat. The module split is a pan-fusion statement; in the index-fusion column alone the NuRD count is 1 and the Sin3 count is 0. |
| epigenetic readout | NONE MEASURED HERE, and that is the point of the entry. No histone-acetylation, deacetylation or nucleosome-positioning readout exists in this dataset; it is affinity purification followed by mass spectrometry. The absence of HDAC1/2 from the network means no deacetylase-activity claim at a KMT2A fusion can be sourced to it. |
inputs
[ "analyses/skucha-mll-fusion-interactome/cache/41467_2018_4329_MOESM4_ESM.xlsx (Skucha 2018 Supplementary Table 5, PMID 29777171)" ]
measurements — the numbers behind the chart
[
{
"quantity": "subunits_present",
"subject": "NuRD remodelling module (CHD3 CHD4 CHD5 GATAD2A GATAD2B MBD2 MBD3)",
"value": 5,
"units": "subunits",
"n": 7,
"universe": "Skucha 2018 960-protein pan-MLL-fusion network, union over seven baits, top ~300 preys per bait",
"null": 0.33,
"note": "null is 7 x 960 / 20431 = 0.33 expected by chance against the proteome background; found members CHD3, CHD4, GATAD2A, GATAD2B, MBD2"
},
{
"quantity": "subunits_present",
"subject": "NuRD deacetylase module (MTA1 MTA2 MTA3 HDAC1 HDAC2 RBBP4 RBBP7)",
"value": 0,
"units": "subunits",
"n": 7,
"universe": "Skucha 2018 960-protein pan-MLL-fusion network, union over seven baits, top ~300 preys per bait",
"null": 0.33,
"note": "all seven absent; rule-of-three bound on the per-subunit presence rate in this arm is under 3/7 = 43%, which is a weak bound and is why the Fisher test on the 2x2 is quoted instead"
},
{
"quantity": "fisher_two_sided_p",
"subject": "NuRD remodelling (5 of 7) versus NuRD deacetylase (0 of 7) presence in the 960",
"value": 0.021,
"units": null,
"n": 14,
"universe": "the 14 canonical NuRD subunits, seven per module, scored in one dataset",
"null": 1,
"note": "one-sided p = 0.0105; background-free by construction, which is why it is the headline instead of the hypergeometric p"
},
{
"quantity": "subunits_present",
"subject": "Sin3/HDAC complex, all 17 subunits",
"value": 4,
"units": "subunits",
"n": 17,
"universe": "Skucha 2018 960-protein pan-MLL-fusion network, union over seven baits",
"null": 0.8,
"note": "SIN3A, SUDS3, ARID4B, ING2 - every one of them Sin3-SPECIFIC; the four subunits shared with NuRD (HDAC1, HDAC2, RBBP4, RBBP7) are 0 of 4"
},
{
"quantity": "subunits_present_in_index_bait",
"subject": "NuRD, all 14 subunits, MLL-AF4 (= KMT2A::AFF1) column only",
"value": 1,
"units": "subunits",
"n": 14,
"universe": "the 291 proteins retained in the MLL-AF4 column of Skucha Supplementary Table 5",
"null": null,
"note": "CHD4 alone, p = 0.0116, 7|9 spectra, negative control 2|2. Sin3/HDAC is 0 of 17 in the same column."
}
]per bait retained proteins — the numbers behind the chart
{
"MLL-AF1p": 293,
"MLL-AF4": 291,
"MLL-AF9": 293,
"MLL-CBP": 293,
"MLL-EEN": 293,
"MLL-ENL": 293,
"MLL-GAS7": 293
}found members with detail
{
"CHD3": "Q12873, 5 of 7 baits, neg_ctrl 0|0, NOT in MLL-AF4",
"CHD4": "Q14839, 7 of 7 baits, neg_ctrl 2|2, MLL-AF4 p=0.0116 (7|9)",
"GATAD2A": "Q86YP4, 1 of 7 baits, neg_ctrl 0|0, MLL-AF9 only, p=0.0361 (10|12)",
"GATAD2B": "Q8WXI9, 5 of 7 baits, neg_ctrl 0|0, NOT in MLL-AF4",
"MBD2": "Q9UBB5, 1 of 7 baits, neg_ctrl 0|0, MLL-EEN only, p=0.0827 (4|5) - above 0.05",
"SIN3A": "Q96ST3, 3 of 7 baits, neg_ctrl 1|2, MLL-AF9/EEN/ENL, NOT in MLL-AF4",
"SUDS3": "Q9H7L9, 1 of 7 baits, neg_ctrl 0|0, MLL-EEN only, p=0.0355 (6|5)",
"ARID4B": "Q4LE39, 2 of 7 baits, neg_ctrl 0|0, MLL-AF9 p=0.0138, MLL-EEN p=0.0553",
"ING2": "Q9H160, 1 of 7 baits, neg_ctrl 0|0, MLL-EEN only, p=0.0147 (1|1)"
}- If this graph ever proposes an HDAC inhibitor on the strength of 'NuRD is in the KMT2A-fusion interactome', the proposal is built on a scaffold signal with no catalytic subunit in the dataset, and in the index fusion's own column on exactly one protein.
- The absent arm is the DRUGGABLE arm. HDAC1/2 have clinical inhibitors and MTA/GATAD2/MBD scaffolds do not, so the module this dataset supports is the one with no chemical matter and the module with chemical matter is the one it cannot see.
- HDAC1/2 sit in at least five distinct corepressor complexes, so an HDAC1/2 inhibitor given for a NuRD hypothesis also hits Sin3, CoREST, NCoR and MiDAC in every cell in the body. The specificity a mechanistic claim needs does not exist at the level of the catalytic subunit.
- Re-searching the public PXD009338 raw files for MTA2, MBD3, HDAC1, HDAC2, RBBP4 and RBBP7 peptides would convert twelve measured-not-reported absences into numbers in an afternoon of compute. Do those proteins have zero spectra in the MLL-AF4 pulldown, or non-zero spectra that the top-300 filter discarded?
- Is the absence of the shared HDAC1/2 + RBBP4/7 core an artefact of the bait-specificity model? The test is to score a set of proteins matched to HDAC1/2 and RBBP4/7 on abundance and complex promiscuity but with no expected relationship to a fusion - if that matched set is also 0 of 4, the filter explains it and the biology does not.
- GATAD2A is in the network in MLL-AF9 only and GATAD2B in five baits but not MLL-AF4. Is the paralogue choice bait-dependent, or is one of them simply the better-detected protein? A spectral-count comparison in the raw files would settle it.
- Does a t(4;11) B-lineage line, rather than Jurkat, recover the NuRD deacetylase arm at the fusion? Every statement here inherits Skucha's T-ALL host proteome.
- E1moderate confidence 3 of 5: One strong study, or several smaller consistent ones.database
Re-scoring the Skucha 2018 960-protein MLL-fusion network by NuRD sub-module gives 5 of 7 for the remodelling arm (CHD3, CHD4, GATAD2A, GATAD2B, MBD2) and 0 of 7 for the deacetylase arm (MTA1, MTA2, MTA3, HDAC1, HDAC2, RBBP4, RBBP7), Fisher two-sided p = 0.0210; the Sin3/HDAC complex scores 4 of 17, all Sin3-specific, with the HDAC1/2 + RBBP4/7 core shared by both complexes at 0 of 4; in the MLL-AF4 column the counts are 1, 0 and 0.
NuRD sub-module split and Sin3/HDAC scoring over Skucha 2018 Supplementary Table 5, run 2026-09-24: analyses/q193-nurd-module-split/run.py, with load-time assertions on the sheet layout and row count, three backgrounds, a background-free Fisher test on the module 2x2, and the SEC and CCT/TRiC controls carried over unchanged from analyses/mll-fusion-interactome-complexes. · This repository (script and results.json committed; subunit lists written out in source and auditable in the diff) 2026What this evidence cannot show
THE SUBUNIT LISTS ARE MY CURATION, not a database's, and the module boundary is the one thing a reader could reject: MBD2/MBD3 are assigned to the remodelling arm because they bridge GATAD2 to the CHD ATPase, and a reader who assigns them to the core would make the split 3/5 versus 0/7 (still one-sided, weaker). The seven-versus-seven balance is a coincidence of the canonical lists and was not engineered; the Fisher test is on those lists as written in the script. The deeper limit is the rank truncation: the table keeps the top ~300 preys per bait, so 'absent' means outside that cut, and the deacetylase arm is the more abundant and more promiscuous arm, which is what a bait-specificity filter removes first. Three backgrounds are reported because the honest denominator (~4,600 recovered proteins) is not in the table; the Fisher test is background-free and is the reason this is reported at all.
Evidence that argues against this (2)
- THE STRONGEST OBJECTION TO MY OWN RESULT, and it is quantitative. On this same table the CCT/TRiC chaperonin - a canonical affinity-purification contaminant that nobody believes is a specific MLL-fusion partner - scores 8 of 9 subunits at p = 2.0e-10, five orders of magnitude stronger than the remodelling arm's 4.4e-06, and 8 of those 8 are in the MLL-AF4 column against the remodelling arm's 1. So complex-wide enrichment on this dataset is demonstrably NOT evidence of specific binding, and the ceiling that places on the NuRD remodelling row is low. The control and its interpretation are the earlier corpus analysis's, not mine, and I carried both over unchanged rather than re-deriving a friendlier one. Skucha A et al. 2018
- A second, independent objection: the absent arm is the SHARED arm. RBBP4 and RBBP7 belong to NuRD, Sin3/HDAC, PRC2 and CAF-1 simultaneously, and HDAC1/2 to at least NuRD, Sin3, CoREST, NCoR and MiDAC - the point Bantscheff et al established by profiling HDAC inhibitors against megadalton complexes rather than catalytic subunits. Promiscuity is exactly the property a bait-specificity model penalises, so 0 of 4 on the shared core is the expected artefact of the filter as well as a possible fact about the biology, and this analysis cannot separate them. Bantscheff M et al. 2011
- E2established confidence 5 of 5: Multiple independent lines of evidence; treated as settled in the field.paper
HDAC1, HDAC2 and the histone-binding proteins RbAp48 (RBBP4) and RbAp46 (RBBP7) form a core complex that is SHARED between NuRD and the Sin3-histone deacetylase complex, and the histone deacetylase activity of that isolated core is severely compromised; MTA2 modulates its enzymatic activity and MBD3 mediates the association of MTA2 with it.
Analysis of the NuRD subunits reveals a histone deacetylase core complex and a connection with DNA methylation. · Zhang Y, Ng HH, Erdjument-Bromage H, Tempst P, Bird A, Reinberg D · Genes & Development 1999What this evidence cannot show
Biochemical purification from HeLa cells, 1999; read from the abstract. The 'shared core' claim is a statement about co-purifying polypeptide composition, not about a single physical particle interconverting between the two complexes. Nothing in this paper is about leukaemia, MLL or any fusion protein.
Evidence that argues against this (1)
- A later chemoproteomic census of HDAC1/2-containing complexes finds MORE of them than the NuRD/Sin3 pair this paper contrasts - including NCoR, CoREST and a then-novel mitotic deacetylase complex (MiDAC) - and shows they are pharmacologically distinguishable, with aminobenzamides binding the NCoR complex more strongly than the Sin3 complex. So 'shared between NuRD and Sin3' understates the sharing, and any claim that an HDAC1/2 signal identifies one of these two complexes is underdetermined. Bantscheff M et al. 2011
- E3established confidence 5 of 5: Multiple independent lines of evidence; treated as settled in the field.paper
The mammalian Sin3 corepressor exists in a complex with the histone deacetylases HDAC1 and HDAC2 together with further associated polypeptides including SAP18 and SAP30.
Histone deacetylases and SAP18, a novel polypeptide, are components of a human Sin3 complex. · Zhang Y, Iratni R, Erdjument-Bromage H, Tempst P, Reinberg D · Cell 1997What this evidence cannot show
1997 biochemistry in HeLa cells, read from the abstract. It establishes co-complex membership, not stoichiometry, and predates the modern subunit inventory (SUDS3/SDS3, ARID4A/B, ING1/2, BRMS1/BRMS1L, PHF12 and MORF4L1 are not in it), so the 17-member Sin3 list scored in E1 is assembled from later work and is my curation.
Evidence that argues against this (1)
- SAP30 and SAP18 - two of the three Sin3-associated polypeptides this paper names - are ABSENT from the Skucha 960-protein network in my own measurement, and SIN3B is absent as well. So the four Sin3 hits (SIN3A, SUDS3, ARID4B, ING2) are not a coherent Sin3 complex any more than the five NuRD hits are a coherent NuRD complex, and this evidence item should not be read as support for 'the Sin3 complex is in the MLL-fusion interactome'. This repository 2026
- E4strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.paper
NuRD is built from preformed, stably separable sub-modules that retain enzymatic activity on their own: purification of endogenous Drosophila NuRD shows a stable core of subunits with the chromatin remodeller CHD4 associating transiently, and combinatorial reconstitution plus single-molecule imaging of CHD4 in live mouse embryonic stem cells with and without MBD3 shows how the deacetylase and remodelling sub-modules associate in vivo.
The Nucleosome Remodeling and Deacetylase Complex NuRD Is Built from Preformed Catalytically Active Sub-modules. · Zhang W, Aubert A, Gomez de Segura JM, Karuppasamy M, Basu S, et al. · Journal of Molecular Biology 2016What this evidence cannot show
THE MIXED-SYSTEM CAVEAT MATTERS HERE: the endogenous purification is DROSOPHILA, the reconstitution is recombinant MultiBac, and the live imaging is MOUSE embryonic stem cells. No part of it is human haematopoietic, and none of it involves a fusion protein. Read from the abstract. It licenses the CLAIM THAT THE MODULES ARE SEPARABLE - which is what E1's module split needs - and nothing about which module any particular cellular activity belongs to.
Evidence that argues against this (1)
- This paper's own finding cuts against the interpretation my analysis is tempted toward. If CHD4 associates with the NuRD core TRANSIENTLY and the sub-modules exist free in both nucleus and cytosol, then recovering CHD4 without the deacetylase module in an affinity purification is the EXPECTED behaviour of the complex rather than evidence that a fusion protein selected one arm. The module split may therefore be a property of NuRD's assembly dynamics, not of MLL fusions. Zhang W et al. 2016
agrees with gene-chd4's own per-bait correction: the MLL-AF4 column contains CHD4 alone of 14 NuRD subunits
- related to →KMT2A::AFF1 (MLL-AF4), t(4;11)(q21;q23)
in the index fusion's own AP-MS column, NuRD is one protein and Sin3/HDAC is nothing
HDAC1 is absent from all seven bait columns of the 960-protein network
HDAC2 is absent from all seven bait columns of the 960-protein network
RBBP4 is shared between NuRD, Sin3/HDAC, PRC2 and CAF-1 and is absent from the network
ING2 is one of the four Sin3-specific subunits recovered (MLL-EEN only, p = 0.0147)
MBD3 absent from all seven bait columns of the 960-protein network
GATAD2A is the one NuRD scaffold subunit recovered, and only under MLL-AF9
Sin3/HDAC scores 4 of 17, all Sin3-specific; the shared HDAC1/2 core is 0 of 4
MTA1, MTA2 and MTA3 all absent from all seven bait columns; the deacetylase module is 0 of 7
one of four Sin3-specific subunits recovered; the shared HDAC1/2 core is 0 of 4
the module split, 5/7 versus 0/7, Fisher two-sided p = 0.0210, is the observation the hypothesis explains
the most broadly recovered NuRD scaffold subunit, and still absent from the index fusion's column