MBD3 (O95983) - the methyl-CpG-binding subunit that does NOT bind methylated DNA, and a validated KMT2A::AFF1-selective CRISPR vulnerability that is absent from every column of the pan-fusion interactome

This is real experimental data — atoms placed by crystallography or cryo-EM, deposited by the people who did the experiment.

• The one KMT2A::AFF1 phenotype measured for MBD3 loss is slower proliferation and explicitly NOT cell death. A cytostatic effect in a disease whose problem is relapse from residual disease is the least useful shape of response, and the screen cannot distinguish it from a temporary arrest that reverses on withdrawal.
• MBD3 is the subunit that couples NuRD's remodelling arm to its deacetylase arm. Removing it does not remove a function, it UNCOUPLES two active modules that this graph's own analysis shows are independently competent - so an MBD3 degrader could release a free, active CHD4 remodeller rather than silencing NuRD, and the direction of that effect is unknown.
• CHD4 disruption in MLL/AF4 models induces myeloid gene expression and is recurrently associated with myeloid lineage switch at relapse. A drug aimed at the same complex in the same disease has to be tested against lineage switch as a primary endpoint, not as a safety afterthought.
• There is no MBD3 chemical matter. The only published route into this module is a peptide that disrupts the p66alpha-MBD2 coiled coil, which targets the MBD2 form of the complex, not MBD3 - so the validated target and the available tool are in mutually exclusive complexes.
• MBD3 is required for normal development in the systems where it has been tested and sits in a complex present in every cell. A near-zero pan-essentiality score in immortalised lines is not a safety signal for an infant.

1. Does MBD3 knockout in SEM change the KMT2A::AFF1 transcriptional programme, or a NuRD programme that runs alongside it? The screen measured proliferation only, so the mechanism is entirely open and a single RNA-seq after MBD3 knockout in SEM would start it.
2. Is the MBD3 dependency B-lineage-specific or infant-specific? KOPN-8 and ALL-PO were screened but only SEM and 697 appear in the validation described, so whether a KMT2A::MLLT1 line or an adult KMT2A-r myeloid line depletes is unknown - and DepMap's myeloid-weighted arm says it does not.
3. Does MBD3 loss phenocopy or oppose CHD4 loss in a KMT2A::AFF1 line? CHD4 disruption in MLL/AF4 models induces myeloid gene expression and accompanies lineage switch at relapse; if MBD3 loss did the same, an MBD3-directed therapy would carry the same escape risk, and nobody has looked.
4. Are MBD2 and MBD3 doing different things in KMT2A-rearranged B-ALL? MBD2 is in the Skucha network (MLL-EEN only, p = 0.0827) and MBD3 is not, while MBD3 is the CRISPR hit - a paralogue swap experiment in SEM would say whether the dependency is on MBD3 specifically or on an MBD-containing NuRD.
5. Does MBD3 have a subnuclear address in a leukaemic cell? Nothing places it in a named compartment in any system, and its partner GATAD2A is annotated to speckles - so a two-colour image in a t(4;11) line is a cheap, genuinely unmeasured result.

1. E1established confidence 5 of 5: Multiple independent lines of evidence; treated as settled in the field.paper

   MBD3 is a subunit of the NuRD complex that mediates the association of MTA2 with the HDAC1/2-RbAp48/RbAp46 core complex, and MBD3 does NOT directly bind methylated DNA although it is highly related to MBD2, which does and which directs NuRD to methylated DNA.

   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 1999
   PMID 10444591DOI
   ▸What this evidence cannot show

   1999 biochemistry from HeLa nuclear extract, read from the abstract. THE LOAD-BEARING HALF OF THIS CLAIM IS AN ABSENCE - 'does not directly bind methylated DNA' - and the assay behind it (which probe, what affinity floor) is not stated in the abstract and was not verified here. Nothing in this paper concerns leukaemia, MLL or any fusion protein.

   ▸Evidence that argues against this (1)

   • Mouse MBD3 does retain a functional MBD in some vertebrates and the non-binding phenotype is allele- and species-scoped; more importantly for this graph, MBD3 has since been reported to associate with 5-hydroxymethylcytosine-marked genes, so 'does not read a DNA modification' is narrower than it sounds - it is specific to 5-methyl-CpG. I searched for a paper directly contradicting the 5mC non-binding result and did not find one; what I found instead were papers extending MBD3's specificity to a different modification, which is a qualification rather than a refutation, and I state it as such. Le Guezennec X et al. 2006
2. E2strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.paper

   MBD2 and MBD3 assemble into mutually exclusive, distinct Mi-2/NuRD-like complexes (MBD2/NuRD and MBD3/NuRD) with different biochemical and functional properties; DOC-1 is a core subunit of both, while PRMT5 and its cofactor MEP50 are specific MBD2/NuRD interactors, and PRMT5 methylates the RG-rich N terminus of MBD2 and is recruited with MBD2 to CpG islands in a methylation-dependent manner.

   MBD2/NuRD and MBD3/NuRD, two distinct complexes with different biochemical and functional properties. · Le Guezennec X, Vermeulen M, Brinkman AB, Hoeijmakers WA, Cohen A, Lasonder E, Stunnenberg HG · Molecular and Cellular Biology 2006
   PMID 16428440DOI
   ▸What this evidence cannot show

   Tagged-protein mass spectrometry in a human cell line (not specified in the abstract), read from the abstract. 'Mutually exclusive' is a statement about the purified complexes recovered, not a proof that no particle ever contains both. This graph already carries gene-prmt5, which is the MBD2-side interactor, so the paralogue choice has a consequence this corpus can already see.

   ▸Evidence that argues against this (1)

   • MUTUAL EXCLUSIVITY CUTS AGAINST TREATING MBD3 AS A NuRD MARKER. If MBD2/NuRD and MBD3/NuRD are separate complexes, then MBD2's presence in the Skucha network (MLL-EEN, p = 0.0827, above 0.05) and MBD3's total absence from it are consistent with the interactome containing the MBD2 form and not the MBD3 form - in which case the MBD3 dependency in SEM and the NuRD signal in the interactome are about two different complexes, and this entry must not fuse them. This repository 2026
3. E3strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.paper

   MBD3 is the subunit whose presence or absence defines how NuRD's deacetylase and chromatin-remodelling sub-modules associate in vivo: single-molecule imaging of CHD4 in live mouse embryonic stem cells with and without MBD3, together with combinatorial reconstitution of all subunit combinations, supports a pathway in which NuRD assembles from preformed, catalytically active sub-modules present in both nucleus and cytosol, with CHD4 associating transiently.

   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 2016
   PMID 27117189DOI
   ▸What this evidence cannot show

   Three systems in one paper and none of them human haematopoietic: endogenous purification in DROSOPHILA, reconstitution in recombinant MultiBac insect cells, imaging in MOUSE embryonic stem cells. Read from the abstract. No fusion protein anywhere in it. The claim it supports is architectural, not about leukaemia.

   ▸Evidence that argues against this (1)

   • The Drosophila arm of this same paper reports that CHD4 associates with the NuRD core only TRANSIENTLY, which weakens any argument that MBD3 loss disables a stable fusion-associated NuRD particle - if the remodeller is already transiently attached, the MBD3 dependency in SEM need not act through a NuRD holo-complex at all. The system is also invertebrate for the composition and murine pluripotent for the imaging, and a cell state further from infant B-ALL is hard to construct. Zhang W et al. 2016
4. E4moderate confidence 3 of 5: One strong study, or several smaller consistent ones.paper

   In CRISPR-Cas9 epigenome and kinome library knockout screens across KMT2A-rearranged ALL lines SEM (KMT2A::AFF1), ALL-PO (KMT2A::AFF1) and KOPN-8 (KMT2A::MLLT1) against KMT2A-wildtype lines 697 and NALM-6, MBD3 was identified and validated as a novel molecular vulnerability alongside ARID4B and BMPR2; sgRNAs targeting MBD3 progressively and significantly reduced GFP-positive cells in SEM but not in 697, an effect the authors attribute to inhibition of cell proliferation rather than leukaemic cell death.

   CRISPR-Cas9 Library Screening Identifies Novel Molecular Vulnerabilities in KMT2A-Rearranged Acute Lymphoblastic Leukemia. · Schneider P, Wander P, Arentsen-Peters STCJM, Vrenken KS, Rockx-Brouwer D, Adriaanse FRS, Hoeve V, Paassen I, Drost J, Pieters R, Stam RW · International Journal of Molecular Sciences 2023
   PMID 37686014DOI
   ▸What this evidence cannot show

   SUPPORT LOCATION MATTERS HERE: THE MCP TOOL'S RENDERING, NOT PUBMED, DELETES THE SYMBOLS: through that tool the abstract reads 'the epigenetic regulators and, as well as the receptor kinase as novel molecular vulnerabilities' with the names missing, but the raw PubMed record contains MBD3, ARID4B and BMPR2 (efetch XML, 2026-10-03) and MBD3[tiab] finds the paper. The symbols MBD3, ARID4B and BMPR2 were read from the PMC full text and from the record's own keyword list on 2026-09-24; an auditor searching the PubMed abstract for 'MBD3' will not find it. CELL LINES ONLY - no primary infant sample, no xenograft, and no normal haematopoietic comparator, so there is no therapeutic-window measurement. The validated readout is a GFP competition assay in TWO lines (SEM and 697); I did not read a quantitative effect size, a replicate count or a statistical test from the figure, and no per-gene depletion score is tabulated. IJMS is a lower-scrutiny venue than the other primary sources in this entry.

   ▸Evidence that argues against this (2)

   • DIRECTLY OPPOSING QUANTITATIVE EVIDENCE: in this project's genome-wide DepMap KMT2A-selectivity scan MBD3 ranks 13,274 of 17,728 with a difference of +0.034 - nominally the wrong direction - and its pan-essentiality is -0.111, so it is not a dependency in any arm. My reading is that the DepMap figure is two nulls rather than an anti-dependency, but a reader who takes DepMap at face value would conclude that MBD3 is not a KMT2A-r vulnerability at all, and that reader has 11 KMT2A-rearranged cell lines behind them against this paper's two. DepMap via this repository 2026
   • And a second, independent one: MBD3 is ABSENT from all seven bait columns of the Skucha 960-protein pan-MLL-fusion interactome, including the MLL-AF4 column. So the one NuRD subunit with a dedicated index-fusion dependency result is also the one with no co-purification evidence of any kind, which is the opposite pattern from CHD4. This repository 2026
5. E5moderate confidence 3 of 5: One strong study, or several smaller consistent ones.database

   MBD3 is absent from the Skucha 2018 960-protein pan-MLL-fusion AP-MS network in all seven bait columns, including MLL-AF4 (= KMT2A::AFF1); it is one of two members of the NuRD remodelling module not recovered, the other being CHD5.

   NuRD sub-module split and Sin3/HDAC scoring over Skucha 2018 Supplementary Table 5 (PMID 29777171), run 2026-09-24: analyses/q193-nurd-module-split/run.py, with load-time assertions on sheet layout and a 960-row count. · This repository (script and results.json committed) 2026
   ▸What this evidence cannot show

   ABSENCE FROM THIS TABLE IS NOT A ZERO. The paper retains the ~300 most significant preys per bait (291 for MLL-AF4), so MBD3's per-bait spectral count exists in the raw data and was never tabulated: absence_class measured-not-reported, upgradeable to measured-not-analysed because the raw files are public at PXD009338. The host is JURKAT, a T-ALL line, and the bait is a doxycycline-induced Strep-HA transgene, so a partner requiring a B-lineage or infant proteome could not appear.

   ▸Evidence that argues against this (1)

   • The same dataset's negative control undermines any confident reading of its absences as well as its presences: the CCT/TRiC chaperonin, a canonical affinity-purification contaminant, scores 8 of 9 subunits at p = 2.0e-10 and 8 of 8 in the MLL-AF4 column. A table on which the strongest complex signal is an artefact is not a table whose silence about one subunit should carry much weight either way. Skucha A et al. 2018
6. E6weak confidence 2 of 5: Preliminary, small, indirect, or not yet peer reviewed.database

   In this project's genome-wide DepMap KMT2A-selectivity scan MBD3 ranks 13,274 of 17,728 with a KMT2A-rearranged-versus-other-leukaemia difference of +0.0337; both arm means (-0.0752 and -0.1089) and the pan-essentiality (-0.1112) are within the noise band around zero, so the ranking orders two null effects rather than establishing an anti-dependency.

   DepMap KMT2A-selective dependency scan over 17,728 genes: analyses/depmap-kmt2a-selective-scan-full/results.json, generated 2026-09-19. · DepMap via this repository 2026
   ▸What this evidence cannot show

   THREE CAVEATS THAT THE CORPUS HAS ALREADY MEASURED AND THAT APPLY IN FULL. (1) The scan FIXES BOTH COMPARISON ARMS from the first gene that returns screen data, so 5 of 16 eligible KMT2A-rearranged lines are dropped from every gene comparison; the effect on any one rank is unquantified. (2) The scan reports a uniform n_all_lines of 1,208 for every gene, whereas DepMap screens each gene in a gene-specific number of lines, and it records NO DepMap RELEASE - only a generation date - so per rule 83 this number cannot be told apart from drift. (3) Rank above the median is a 50% null by construction; 13,274 of 17,728 is percentile 74.9, which is not far enough from the base rate to mean anything on its own. I re-read these figures from results.json on 2026-09-24 rather than quoting them from the earlier entry that cited them.

   ▸Evidence that argues against this (1)

   • A dedicated CRISPR screen in the relevant lineage finds the opposite: MBD3 knockout depletes the KMT2A::AFF1 line SEM and not the KMT2A-wildtype line 697. DepMap's KMT2A-rearranged arm here is 11 lines that are predominantly myeloid and adult-derived, so the disagreement may be lineage rather than error - which is exactly the reading that would make the DepMap rank uninformative for infant B-ALL. Schneider P et al. 2023
7. E7moderate confidence 3 of 5: One strong study, or several smaller consistent ones.database

   A bounded literature search on 2026-09-24 returns 1 record for MBD3 with KMT2A in title or abstract and 0 for MBD3 with MLL, against 55 for the MEN1/MLL positive control, with the single KMT2A record being the CRISPR screen cited in E4; MBD3 with leukemia returns 7.

   NCBI esearch via scripts/pubmed_safe.py on 2026-09-24. MBD3[tiab] AND KMT2A[tiab] -> 1 (translation "MBD3"[Title/Abstract] AND "KMT2A"[Title/Abstract], PMID 37686014). MBD3[tiab] AND MLL[tiab] -> 0. MBD3[tiab] AND leukemia[tiab] -> 7. Positive control MEN1[tiab] AND MLL[tiab] -> 55. · NCBI PubMed E-utilities 2026
   ▸What this evidence cannot show

   One tag per term (rule 94: never the grouped `(A OR B)[tiab]` form), and the querytranslation is recorded beside every count because that is the only way a later reader can see which search ran. pubmed_safe raises rather than returning a fake zero on an empty translation, so the 0 for MLL is a real zero. BUT NOTE THE RECALL LESSON IN MY OWN NUMBERS: at the complex level, NuRD[tiab] AND KMT2A[tiab] returns 0 while NuRD[tiab] AND MLL[tiab] returns 6 - the current and retired symbols give DISJOINT hit sets, and a beat that queried only KMT2A would have concluded the NuRD literature in this disease does not exist.

   ▸Evidence that argues against this (1)

   • A PubMed count bounds what has been REPORTED, not what has been measured, and this specific absence is the weaker kind. MBD3 is in every genome-wide CRISPR library, so it has been silently screened in dozens of KMT2A-rearranged contexts whose per-gene scores were never written into a title or abstract; the DepMap row is one such measurement and it did not reach this query. The honest class for 'MBD3 in KMT2A::AFF1' is measured-not-reported, not never-measured. DepMap via this repository 2026

• depends on →KMT2A::AFF1 (MLL-AF4), t(4;11)(q21;q23)

  MBD3 knockout depletes the KMT2A::AFF1 line SEM and not the KMT2A-wildtype line 697 (GFP competition)

• related to →CHD4 (Q14839) - the NuRD ATPase, a pan-MLL-fusion co-purifier, and the gene whose disruption drives lineage switching in KMT2A::AFF1 relapse

  MBD3 couples the NuRD deacetylase module to the CHD4 remodelling module; answers gene-chd4's open question in the negative for MBD3

• related to →Infant acute lymphoblastic leukaemia (ALL diagnosed under 1 year of age)

  the validating line SEM and the co-screened ALL-PO are infant KMT2A::AFF1 B-ALL lines

• related to →Plain-English: chromatin and epigenetics — the sticky notes on the DNA, and which ones matter here

  the methyl-CpG-binding subunit that does not read 5-methyl-CpG

• ← supportsThe 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

  MBD3 absent from all seven bait columns of the 960-protein network

• ← bindsGATAD2A / p66-alpha (Q86YP4) - the subunit that bridges NuRD's two modules, the best-ranked NuRD gene in this project's KMT2A-selectivity scan, and present in the pan-fusion interactome under MLL-AF9 but NOT the index fusion

  p66-alpha binds directly to both MBD2 and MBD3 (yeast two-hybrid and co-precipitation)

• ← bindsMTA2 / MTA1L1 (O94776) - the NuRD scaffold that clamps HDAC1, recruited in T lymphocytes by BCL11B, and absent together with MTA1 and MTA3 from every column of the pan-MLL-fusion interactome

  MBD3 mediates the association of MTA2 with the HDAC1/2-RBBP4/7 core

• part of →The NuRD complex - two separable modules, three paralogue choices that change what it does, and a graph that carried its ATPase and its shared core but never the scaffolds that make it NuRD

  NuRD-specific; couples MTA2 to the HDAC1/2 core

• ← co-occurs withARID4B / SAP180 (Q4LE39) - the Sin3A-complex subunit that shares SIN3A's HDAC1-binding region, is in the MLL-fusion interactome, and is the second of only two epigenetic vulnerabilities a CRISPR screen found in KMT2A-rearranged ALL

  the two epigenetic hits of the same epigenome-wide CRISPR screen: one Sin3 subunit, one NuRD subunit

• ← supportsNo experiment has ever ASKED whether MTA2, MBD3, GATAD2A or SIN3A binds a KMT2A fusion - four of them were nonetheless in the lysate of a seven-fusion pulldown, and the class of that absence is measured-not-reported, not never-measured

  no binding experiment; one dependency screen; absent from all seven bait columns

• ← bindsGATAD2B / p66-beta (Q8WXI9) - the NuRD scaffold with the broadest presence in the pan-MLL-fusion interactome, five of seven baits and still not the index fusion, and a haploinsufficiency syndrome that caps any therapeutic idea

  p66-beta binds directly to both MBD2 and MBD3, at lower affinity and through one interaction domain

• ← related toNo histone-mark readout has ever been measured after perturbing ANY NuRD or Sin3 subunit in a KMT2A-rearranged cell - the two complexes whose entire job is acetylation have zero acetylation data in this disease

  the CRISPR screen that made MBD3 a KMT2A::AFF1 dependency measured proliferation only