Phase-2 dossier: the CXXC-DNA interface - two structures, a genetically validated requirement, and a selectivity problem that decides everything

Verified against RCSB this cycle 1: 2J2S, solution structure of the non-methyl-CpG-binding CXXC domain of the leukaemia-associated MLL histone methyltransferase (NMR); 4NW3, crystal structure of the MLL CXXC domain in complex with a CpG DNA at 2.82 A. Both are isolated-domain structures with their ligand, which is exactly what a structure-based campaign needs, and a rare completeness for this beat.

What is missing is any structure of the CXXC domain in the context of the fusion protein or on a nucleosome, so the geometry of the real target is unknown.

A protein-DNA interface. The CXXC zinc-finger presents a positively charged, shallow surface that reads the unmethylated CpG dinucleotide. Three consequences follow directly:

1. Small molecules are poorly suited. There is no buried hydrophobic pocket; affinity is distributed over electrostatics and a few base contacts. The field's successes against nucleic-acid-binding proteins are overwhelmingly allosteric or degradative, not orthosteric.
2. Selectivity is a family problem, not a pair problem. The zinc-coordinating CXXC fold is conserved across roughly a dozen human proteins with essential and opposing functions - CXXC1/CFP1 in H3K4 methylation, TET1/TET3 in demethylation, DNMT1 in maintenance methylation. Hitting the family would scramble DNA methylation genome-wide.
3. A DNA-competitive ligand is a DNA-mimetic. Molecules that mimic the phosphodiester backbone tend to be polyanionic, cell-impermeant, and promiscuous across nucleic-acid-binding proteins.

ChEMBL holds 577 activity rows at pChEMBL 6 or above against KMT2A (CHEMBL1293299) and 1,014 against the menin/MLL protein-protein-interaction target entity (CHEMBL2093861) 2. Read carelessly, the first number says KMT2A is a well-drugged protein. It is not. Those rows are dominated by the WDR5-MLL and menin-MLL interaction programmes, which engage the partner proteins. To this review's knowledge no published series engages the CXXC domain itself. The number is a warning about how easily a druggability score computed from ChEMBL target-level counts can mislead - and that is a caution that applies to any ML model trained on such counts.

CORRECTED 2026-09-21. This section said 'To this review's knowledge no published series engages the CXXC domain itself.' No series is right; no molecule is not. Liang et al. report dimethyl fumarate, an approved oral therapy for multiple sclerosis and psoriasis, inhibiting the MLL1 CXXC-DNA complex with a relative IC50 of 0.91 +/- 0.09 uM in a single-molecule hairpin-unfolding assay, with the active metabolite monomethyl fumarate inactive at 10 uM 4. It strengthens the STOP verdict rather than weakening it. DMF covalently modifies more than 2,400 cysteines and the CXXC fold coordinates zinc through eight of them, so this is the family-wide reactive-cysteine liability, demonstrated. It was not counter-screened against CXXC1, KDM2B or TET1 and would not be expected to pass gate G1. See ana-the-cxxc-interface-has-exactly-one-published-inhibitor-and-it-is-an-approved-drug.

The selectivity gate must be passed before any potency work. Concretely: a candidate must show at least 30-fold discrimination between the KMT2A CXXC domain and CXXC1/CFP1, KDM2B and TET1 in a biophysical panel. If it cannot, the programme stops, because a pan-CXXC agent is a genome-wide DNA-methylation disruptor with no plausible paediatric safety case.

The realistic alternative modality, and the one this dossier recommends evaluating first because it is cheap: do not drug the domain, degrade the protein that carries it. A KMT2A-directed degrader engaging any accessible surface would remove the CXXC function as a by-product, and the selectivity problem becomes a ligandability problem on a different, possibly better, surface. That routes this candidate into the degrader dossier.

A potency gate cannot be set from the literature, and 2026-09-21 established why. The natural ligand's affinity is not one number: ten published measurements of the CXXC-DNA contact span 44 uM to 33 nM, driven by construct span (1147-1203 against the 1147-1244 MT domain), tag-borne valency, DNA valency (one CpG against six) and ionic strength (150 against 300 mM) 5. The nearest-to-physiological tagless value is 5.7 uM at 150 mM NaCl; the most multivalent geometry gives 33 nM. No measurement uses full-length KMT2A, a fusion protein or a nucleosome, so the species a competitive ligand must displace has never been characterised at all.

Two consequences for this dossier. First, any target-product profile stating a required potency would be guessing across three orders of magnitude, so the honest G0 is measure the affinity of the relevant species before setting a potency gate. Second, and more decisive: in the same MLL-AF9 swap series that supplies this dossier's functional-substitution warning, affinity rank is inverted against functional rank - the tighter-binding CGBP/CXXC1 domain fails to support leukaemogenesis and the 34-fold weaker DNMT1 domain succeeds - and every swapped domain including two that bind no DNA still targeted the fusion to Hoxa9 by ChIP 6. So a compound could achieve full biochemical engagement and move neither localisation nor function, which means gate G2 as written (chromatin displacement by CUT&RUN) may be unreachable in principle rather than merely demanding. G2 should be re-specified around CpG methylation at the target locus, which is the property that actually tracked function.