Oncofusion

start / graph / localization-cxxc-cpg-address

Mechanismmoderate confidence 3 of 5: One strong study, or several smaller consistent ones.Q4b-localization

The CXXC domain (aa 1147-1195, PDB 2KKF) is a chromatin address, not a compartment address — CpG mutants lose transformation but keep their subnuclear position

also called MLL CXXC · MT domain · methyltransferase homology domain · 2KKF · nonmethylated CpG binding

The CXXC domain is the part of KMT2A that reads unmethylated CG letters in DNA and so decides which genes the fusion lands on. The decisive experiment showed that breaking it stops the leukemia without moving the protein anywhere else in the nucleus — so it tells the fusion which genes to sit on, not which part of the nucleus to live in.

The same thing, for a scientist

The CXXC-type zinc finger of human KMT2A occupies residues **1147-1195** (UniProt Q03164) [E1], and the solution NMR structure of the domain bound to palindromic CpG DNA is **PDB 2KKF**, chain A = Q03164 residues 1147-1203, Homo sapiens, verified at RCSB [E2]. Birke et al. measured the recombinant MT/CXXC domain binding CpG-containing DNA with a KD of approximately 3.3 x 10^-8 M and showed 5-methylcytosine on either strand blocks the interaction [E3]. Cierpicki et al. solved the complex and used it to design point mutants that were introduced into MLL-AF9: disrupting CpG binding caused CpG hypermethylation and H3K9 methylation gain at Hoxa9, Hoxa9 downregulation, loss of immortalisation and failure to induce leukemia in mice [E2]. The observation that matters most for a localisation beat comes from Ayton, Chen & Cleary: high-resolution mutagenesis identified CXXC point mutants that abolish non-methylated-CpG binding, transactivation and myeloid transformation **while leaving subnuclear localisation intact**, assessed by fluorescence microscopy in COS cells [E4]. That is a clean dissociation. Where the fusion sits inside the nucleus and which loci it occupies are separately encoded: the SNL elements and the N-terminal region supply the former, CXXC supplies the latter. Any model that treats 'the fusion is mislocalised' as the explanation for its transforming activity has to survive this result, and at present it does not.

loading 2KKF…
PDB 2KKF
The CXXC domain of human KMT2A gripping a short piece of DNA at an unmethylated CG site.
Tell me more about this picture

Solution NMR structure; chain A is UniProt Q03164 residues 1147-1203, Homo sapiens.

Tell me more about this picture

Point mutations designed on this structure abolish CpG binding, transformation and leukemia induction, but do not move the protein within the nucleus.

What it looks like
loading 2KKF…
PDB 2KKF ↗
shape

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

I have never read one of these pictures before

A protein is one long chain of amino acids that folds up. The ribbon traces that chain from one end to the other; it is not a wire, it is a simplification of where the backbone runs. Drag to rotate, scroll to zoom.

Domains are the stretches of chain that do a particular job, and they are what the coloured blocks above the picture mark out. Click one and the view flies to it. In this project the important question about any domain is usually whether the fusion still has it — the labels say lost where it does not.

Anything in yellow sticks is a bound small molecule, which is usually the drug.

Go deeper
CORRECTION (cycle 22 counter-evidence pass): the second half of this entry's claim does not survive

This entry made two claims. The first — that CXXC is not a compartment address, because CXXC point mutants lose CpG binding, transactivation and transformation while keeping their subnuclear position 4 — stands. The second — that CXXC therefore is the locus address, the module that decides which genes the fusion lands on — is contradicted by a direct experiment that this entry did not cite.

Risner et al. replaced the MLL CXXC domain inside MLL-AF9 with the CXXC domains of DNMT1, CGBP/CFP1 and MBD1, and with the MBD1 methyl-CpG-binding domain. Every substituted fusion still targeted the Hoxa9 locus in primary bone-marrow progenitor cells, including the MBD1 domains, which do not bind the same DNA in vitro at all. What separated the leukemogenic swaps from the non-leukemogenic ones was not targeting but which specific CpG residues were protected from de novo methylation, and hence whether H3K9me3 was kept off the locus and Hoxa9 stayed expressed. MLL(DNMT1 CXXC)-AF9 transformed and caused leukemia like MLL-AF9; MLL(CGBP CXXC)-AF9 did not, although CGBP CXXC binds unmethylated DNA perfectly well (see counter-evidence on 2 and 4).

So CXXC is best described as a local methylation guard at an address supplied by something else — the menin-LEDGF tether and the pre-existing chromatin state of the cell of origin — rather than as the addressing module itself. The measured 33 nM CpG affinity 3 and the 2KKF structure 2 are unaffected as biophysics; what changes is the function they are assigned.

There is a second, softer problem with the compartment/locus dichotomy itself. Yano et al., the study that defined MLL's subnuclear localisation signals, interpreted the punctate pattern as clusters of ALL-1 protein positioned at ALL-1 target sites on the chromatin, and concluded that the role of the N terminus is to direct the protein to its target genes. On that reading a punctum is a set of occupied loci seen at low resolution, not a separate compartment, and 'position' and 'locus' are one variable rather than two. A CXXC mutant scored as still-punctate in COS monkey kidney cells — which have no Hoxa9 program and none of the relevant target set — is then a weak test of separability.

confidence_overall has been lowered from strong to moderate for this reason. The dissociation result is intact; the addressing model built on top of it is not.

The dissociation experiment, stated precisely

Ayton et al. performed high-resolution mutagenesis across MLL fusion constructs and scored four things in parallel: in-vitro non-methylated CpG binding, luciferase transactivation, in-vitro myeloid transformation after retroviral transduction of mouse bone marrow, and subnuclear localisation by fluorescence microscopy in COS cells. CXXC point mutants lost the first three and retained the fourth 4.

This is the single most important result in the graph for calibrating localisation claims about KMT2A fusions, and it cuts against the beat's own subject matter. It says the protein's address inside the nucleus is not the variable that carries transformation. It does not say localisation is irrelevant — a protein must be nuclear and chromatin-associated to do anything — but it removes 'gross subnuclear position' from the list of candidate mechanisms for CXXC-dependent transformation.

Method caveat that keeps the result honest: the localisation readout is fluorescence microscopy of an expressed construct in COS cells, a monkey kidney fibroblast line. A speckled-versus-diffuse call at that resolution would not detect a subtler change (a shift in the fraction at the nucleolar periphery, a change in cluster number or dwell time). The claim that is safe is 'no gross redistribution'; the claim that is not safe is 'no change in localisation'.

What CXXC actually reads, with numbers1 deeper

Birke et al. used recombinant GST-MT protein with EMSA, surface plasmon resonance, SELEX site selection and DNase I footprinting: the domain binds CpG-containing DNA with KD approximately 3.3 x 10^-8 M (about 33 nM), and 5-methylcytosine on either strand blocks it 3. Cierpicki's NMR structure of the domain-DNA complex explains the discrimination structurally and licenses structure-guided mutation 2.

The consequence for targeting is the one the graph already records under mechanism-fusion-dna-targeting-cxxc-athooks: the fusion's destination set is a read-out of the cell's pre-existing DNA-methylation landscape, i.e. the unmethylated CpG-island promoter set of the cell of origin. What this entry adds is that this is a chromatin-level address with a measured affinity and a structure, operating independently of the protein's compartment-level address.

Why a 33 nM CpG reader cannot by itself specify a target set

There are roughly 30,000 CpG islands in the human genome and the CXXC domain does not discriminate among them by sequence. A 33 nM affinity for a ubiquitous feature specifies a class, not a list. The selectivity within that class has to come from somewhere else: the menin-LEDGF tether, the AT-hook contribution, H3K36me2/3 read by the LEDGF PWWP domain, local chromatin accessibility, and the cell-of-origin methylome. This is the structural reason the graph's hyp-methylome-of-origin-sets-fusion-target-set is plausible — CXXC is a filter, not an addressing scheme, so whatever sets the methylome sets the target list.

The residue-level picture of 'what decides where this protein goes'

Reading the N-terminal retained fragment as a stack of separable addressing modules, with the coordinates verified rather than recalled:

elementresidueswhat it specifiesevidence grade
MBM1/MBM2 (menin-binding)~6-10 and downstreamtether to menin-LEDGF on chromatinestablished
AT-hook 1/2/3169-180, 217-227, 301-309DNA affinity; structure- or AT-selective (disputed)moderate
SNL-1388-432punctate subnuclear distributionmoderate
RKRKRK nuclear targeting734-739nuclear importmoderate
SNL-21051-1089punctate subnuclear distributionmoderate
CXXC1147-1195unmethylated-CpG locus selection; not compartmentstrong

The honest summary of the beat's central question: compartment position is set by the SNLs and the N-terminal region; locus selection is set by CXXC plus the menin-LEDGF tether; and the two are experimentally separable. Nobody has performed the reciprocal experiment — an SNL-deleted, CXXC-intact fusion scored for transformation — which would close the logic.

Facts and figures
uniprotQ03164
kd NM33
kd noteapproximately 3.3e-8 M for the recombinant MT/CXXC domain on CpG-containing DNA
structure species verifiedHomo sapiens (RCSB entry 2KKF, UniProt cross-reference Q03164)
separable from subnuclear localisationtrue
PDB IDS
[
  "2KKF",
  "2J2S",
  "2JYI",
  "4NW3"
]
domains
[
  {
    "name": "CXXC (CpG-binding zinc finger)",
    "start": 1147,
    "end": 1195
  },
  {
    "name": "CXXC construct in PDB 2KKF",
    "start": 1147,
    "end": 1203
  }
]
cxxc dna kd conditions
{
  "_why": "Added by king sweep 2026-09-21 (cycle 46). This entry quotes the KMT2A CXXC-DNA 44 uM. That number is a property of five choices, not of the molecule, and quoting it bare is how this graph manufactured a 1,300-fold 'discrepancy' with the menin contact.",
  "value": "44 uM",
  "construct": "isolated CXXC domain, KMT2A 1147-1203, untagged, 15N-labelled",
  "ligand": "DNA pal oligonucleotide, one CpG element",
  "method": "NMR titration",
  "buffer": "50 mM phosphate, pH 7.1, 300 mM NaCl, 2 mM DTT, 25 C",
  "nacl_mM": 300,
  "source_pmid": "20010842",
  "source_quote": "\"Determination of the dissociation constants by NMR for the CXXC domain (MLL 1147-1203) and mutants was performed by titration of 15N-labeled protein with concentrated solutions of DNA pal oligonucleotide. Experiments were carried out in 50 mM phosphate buffer (pH 7.1), 300 mM NaCl, 2 mM DTT at 25 C.\" -- Cierpicki 2010 NSMB, Methods, fetched and read from PMC2908503 on 2026-09-21. The result sentence reads \"The wild type protein yielded a Kd of 44 uM\".",
  "this_is_the_high_salt_end": "300 mM NaCl is roughly twice physiological ionic strength and SIX TIMES the 50 mM the menin-MBM1 53 nM was measured in. The same domain measured tagless at 150 mM NaCl gives 5.7 uM -- 7.7-fold tighter from salt alone.",
  "do_not_compare_naively": {
    "against": "menin-MBM1 53 nM (PMID 20961854), measured at 50 mM NaCl by FP",
    "best_matched_pair": "108-fold, not ~1,000-fold",
    "see": "ana-the-menin-affinity-comparison-was-made-at-one-sixth-the-salt"
  },
  "added": "2026-09-21"
}
How this could bite us back
  • A CXXC-DNA interface inhibitor would face a large family of paralogous CXXC domains including DNMT1 and the TET enzymes; off-target engagement would perturb the methylome globally, which in an infant is a developmental as well as an oncological risk.
  • Because CXXC selects a class (unmethylated CpG islands) rather than a list, partial inhibition redistributes rather than silences the fusion: the highest-avidity loci retain occupancy while marginal ones are lost, which predicts a graded transcriptional response and a surviving core program — the same residual-occupancy failure mode the graph records for menin inhibition.
  • If the methylome of the cell of origin sets the target set, then a therapy that changes methylation (hypomethylating agents, already used in AML) could *expand* the fusion's accessible target set rather than shrink it.
Still open
  1. Risner et al. showed heterologous CXXC and even non-DNA-binding MBD domains still deliver MLL-AF9 to Hoxa9. What then does supply the address? The obvious test is a menin-binding-motif mutant and a LEDGF-PWWP mutant scored for Hoxa9 occupancy in the same assay, which would tell the lab whether the druggable tether is also the addressing module — the single most decision-relevant experiment in this beat for the menin hypothesis.
  2. The reciprocal experiment has never been done: delete or mutate SNL-1 (388-432) and SNL-2 (1051-1089) while leaving CXXC intact, and ask whether transformation survives. If it does, subnuclear position is not load-bearing for this disease and the whole compartment programme should be re-scoped.
  3. Does the CXXC mutant that preserves gross localisation also preserve fine localisation — cluster number, nucleolar fraction, chromatin dwell time — when re-scored by live single-molecule imaging rather than 2004-era fluorescence microscopy?
  4. Does CXXC binding measured on naked DNA (KD ~33 nM) hold on nucleosomal templates, where the CpG may be occluded? If affinity collapses on nucleosomes, the addressing logic runs through accessibility rather than through CXXC.
  5. Can a small molecule occupy the CXXC-DNA interface selectively over the ten other CXXC-domain proteins (DNMT1, TET1/3, KDM2A/B, CFP1, MLL1-4)? The graph's `target-kmt2a-cxxc-dna-interface` entry treats this as open and the structure 2KKF is the obvious starting point.
Every source behind this entry (4)
  1. E1established confidence 5 of 5: Multiple independent lines of evidence; treated as settled in the field.database

    The CXXC-type zinc finger of human KMT2A (UniProt Q03164, 3,969 aa) is annotated at residues 1147-1195.

    UniProtKB Q03164 (KMT2A_HUMAN) — Histone-lysine N-methyltransferase 2A · UniProt Consortium · UniProt Knowledgebase 2026
    What this evidence cannot show

    Curated feature boundary; constructs used experimentally span slightly different ranges (2KKF chain A covers 1147-1203; the 2J2S construct 1143-1214), so the functional domain is a little larger than the annotated zinc finger.

    Evidence that argues against this (1)
    • The annotated zinc finger identifies membership of a class whose defining property is NON-selective genome-wide CpG-island binding, which is the opposite of an address. Blackledge, Thomson and Skene describe how CpG islands act as nucleation sites for the zinc-finger CxxC-domain proteins CFP1 and KDM2A, each of which carries an associated enzymatic activity that modulates histone lysine methylation, thereby imposing a defined chromatin architecture on CpG islands that distinguishes them from the surrounding genome. So a CXXC domain buys you every nonmethylated CpG island in the genome - tens of thousands of sites - and KMT2A's occupancy is a small subset of those. Whatever selects that subset is not the CXXC domain, and this entry's title claim of a 'CpG address' overstates what the module can specify. Blackledge NP et al. 2013
  2. E2strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.paper

    The solution NMR structure of the human MLL CXXC domain bound to palindromic CpG DNA (PDB 2KKF, chain A = Q03164 residues 1147-1203) shows how the domain reads non-methylated CpG, and structure-guided point mutants introduced into MLL-AF9 cause Hoxa9 CpG hypermethylation, H3K9 methylation gain, Hoxa9 downregulation, loss of immortalisation and failure to induce leukemia in mice.

    Structure of the MLL CXXC domain-DNA complex and its functional role in MLL-AF9 leukemia · Cierpicki T, Risner LE, Grembecka J, Lukasik SM, Popovic R, Omonkowska M, Shultis DD, Zeleznik-Le NJ, Bushweller JH · Nature Structural & Molecular Biology 2010
    PMID 20010842DOI
    What this evidence cannot show

    The structure is of the isolated domain with a short palindromic duplex, not of the full protein on a nucleosome. The functional arm uses retrovirally expressed MLL-AF9 in mouse cells, and the fusion is MLL-AF9 rather than KMT2A::AFF1. PDB 2KKF was checked at RCSB and is Homo sapiens, not a murine construct. LADDER POINTER added 2026-09-21 by Q82: this contact has TEN published dissociation constants spanning 44 uM to 33 nM across construct (1147-1203 to 1147-1244), tag (none / GST), DNA (12-bp one-CpG to 52-bp six-CpG) and ionic strength (150 to 300 mM). The nearest-to-physiological tagless value is 5.7 uM (PMID 34280370, 150 mM NaCl). Do not quote any single figure as the affinity - see ana-cxxc-dna-affinity-ladder.

    Evidence that argues against this (1)
    • Qualifies the structure-guided-mutant argument. Cierpicki's point mutants abolish CpG binding and leukemia together, which is read here as CpG binding being the targeting function. Risner et al. dissociate the two: heterologous CXXC domains that bind unmethylated CpG perfectly well (CGBP/CFP1) fail to support leukemogenesis, while a heterologous domain that binds with different affinity (DNMT1 CXXC) supports it fully - and every swap, binding or not, still delivered the fusion to Hoxa9. The functional variable they isolate is protection of particular CpGs from de novo methylation, which is downstream of binding and not the same thing as target recognition. Read: abstract only. Risner LE et al. 2013
  3. E3strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.paper

    The MT (CXXC) domain of MLL binds CpG-containing DNA with a dissociation constant of approximately 3.3 x 10^-8 M and discriminates against methylation, with 5-methylcytosine on either strand blocking the interaction.

    The MT domain of the proto-oncoprotein MLL binds to CpG-containing DNA and discriminates against methylation · Birke M, Schreiner S, Garcia-Cuellar MP, Mahr K, Titgemeyer F, Slany RK · Nucleic Acids Research 2002
    PMID 11842107DOI
    What this evidence cannot show

    Recombinant GST-MT fusion protein in vitro (EMSA, SPR, SELEX, DNase I footprinting) plus an MT-VP16 hybrid reporter assay; affinity measured on naked DNA, not nucleosomal DNA, and GST dimerisation can inflate apparent avidity. GST dimerisation is the explanation everyone reaches for and it is now bounded: a same-laboratory, same-construct, same-oligonucleotide pair puts GST plus a halving of ionic strength at 5.5-fold, and a tagless pair puts about 8-fold on the salt alone, so the tag's own contribution is small. LADDER POINTER added 2026-09-21 by Q82: this contact has TEN published dissociation constants spanning 44 uM to 33 nM across construct (1147-1203 to 1147-1244), tag (none / GST), DNA (12-bp one-CpG to 52-bp six-CpG) and ionic strength (150 to 300 mM). The nearest-to-physiological tagless value is 5.7 uM (PMID 34280370, 150 mM NaCl). Do not quote any single figure as the affinity - see ana-cxxc-dna-affinity-ladder.

    Evidence that argues against this (1)
    • A 33 nM affinity for CpG-containing naked DNA cannot be a selectivity mechanism when the same chemistry is deployed genome-wide by other proteins. Blackledge et al. review the evidence that CpG islands serve as nucleation sites for the ZF-CxxC proteins CFP1 and KDM2A, which recruit methylation-modulating activities to essentially the whole nonmethylated CpG-island compartment and give it a distinct chromatin architecture. Methylation discrimination therefore separates CGIs from the rest of the genome, not KMT2A's targets from other CGIs, and the affinity measured on naked oligonucleotides by EMSA, SPR and footprinting says nothing about competition with a nucleus full of CxxC proteins for the same sites. The selectivity term is missing from this entry and from the field. Blackledge NP et al. 2013
  4. E4strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.paper

    High-resolution mutagenesis identified CXXC point mutants of an MLL oncoprotein that abolish non-methylated-CpG binding, transactivation and myeloid transformation while leaving subnuclear localisation intact as assessed by fluorescence microscopy in COS cells.

    PMID 15542854DOI
    What this evidence cannot show

    The localisation readout is fluorescence microscopy of an expressed construct in COS monkey kidney fibroblasts, which supports 'no gross redistribution' but not 'no change in localisation'; subtler changes in cluster number, nucleolar fraction or dwell time would not be detected. Transformation is scored in mouse bone marrow in vitro.

    Evidence that argues against this (2)
    • This is the strongest counter-evidence in the localization beat and it contradicts the second half of this entry's central claim. The entry says compartment position is set by the SNLs while LOCUS selection is set by CXXC. Risner et al. replaced the MLL CXXC domain in MLL-AF9 with the CXXC domains of DNMT1, CGBP/CFP1 and MBD1, and with the MBD1 methyl-CpG-binding domain, and found that ALL of the substituted fusions - including the MBD1 domains, which do not bind the same DNA at all in vitro - still targeted the fusion to the functionally important Hoxa9 locus in primary bone-marrow progenitors. What separated leukemogenic from non-leukemogenic swaps was not locus targeting but which specific CpG residues were protected from methylation, and hence whether H3K9me3 was prevented and Hoxa9 stayed expressed. MLL(DNMT1 CXXC)-AF9 transformed and caused leukemia as well as MLL-AF9; MLL(CGBP CXXC)-AF9 did not, despite binding unmethylated DNA. So CXXC is not the address either: it is a local methylation guard at an address supplied by something else (menin-LEDGF, the pre-existing chromatin state). Ayton's dissociation between compartment and transformation still stands; the positive assignment of locus selection to CXXC does not. Read: abstract only. Risner LE et al. 2013
    • The compartment-versus-locus dichotomy this entry builds on may be a distinction without a difference, on the authority of the paper that defined the subnuclear localisation signals in the first place. Yano et al. mapped the punctate determinants to two short sequences in the N-terminal approximately 1100 residues that are conserved with TRITHORAX, showed that enforced nuclear expression of the PHD fingers or the SET motif gives a uniform non-punctate pattern, and then interpreted the speckles explicitly as protein complexes containing multiple copies of ALL-1 positioned AT ALL-1 TARGET SITES ON THE CHROMATIN, concluding that the role of the N terminus is to direct the protein to its target genes. On that reading a punctum is a cluster of occupied loci, not a separate compartment, and a CXXC mutant that keeps a punctate pattern in COS cells - a monkey kidney line with no Hoxa9 program and no leukemic target set - is weak evidence that position and locus are separately encoded. Read: abstract only. Yano T et al. 1997
Connected to