Oncofusion

start / graph / ana-phase2-dataset-pdb-alphafold

Analysisstrong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.Q14-phase2-readiness

Dataset: PDB and AlphaFold - the structural census, including the two absences that shape every plan

also called PDB access plan · AlphaFold census · structure availability audit

Before designing a drug you need to know the shape of what you are aiming at. We checked every structure this project relies on against the official databases, and also checked what is missing. Two things are missing that matter: there is no computed model of the full KMT2A protein, and no experimental structure of any intact fusion protein.

The same thing, for a scientist

A structural census was run this cycle against the RCSB data API and the AlphaFold prediction API, and every PDB identifier cited in this beat's dossiers was verified by fetching its entry record rather than recalled. Two absences were confirmed by direct query and are load-bearing for the whole phase-2 plan: the AlphaFold database returns HTTP 404 for UniProt Q03164 (KMT2A), so no full-length model exists, and no PDB entry exists for an intact KMT2A fusion protein or for any junction-spanning construct. The AlphaFold confidence census is equally informative: menin 84.44 global pLDDT, KAT7 72.56, ENL 64.69, LEDGF 62.62, CBX8 61.75, DOT1L 52.50, AFF1 52.34. The ordering of those numbers is almost the ordering of the portfolio, which is not a coincidence - folded proteins have pockets and pockets have drugs.

Go deeper
What was verified, and how

Every structure cited in this beat was fetched from https://data.rcsb.org/rest/v1/core/entry/<ID> and its title, method and resolution read from the response 1:

PDBTitleMethodResolution
4GQ6Human menin in complex with MLL peptideX-ray1.55 A
7UJ4Inhibition of Human Menin by SNDX-5613X-ray1.96 A
3U88Human menin in complex with MLL1 and LEDGFX-ray3.00 A
2MSRLEDGF/p75 IBD with MLL1 peptide 140-160NMR-
4HRADOT1L in complex with EPZ-5676X-ray3.15 A
5J9SENL YEATS with H3K27acX-ray2.70 A
6HT1MLLT1 (ENL) YEATS with SGC-iMLLTX-ray2.10 A
6MAJHBO1/KAT7 HAT domain with inhibitorX-ray2.14 A
3I91CBX8 with H3K9 peptideX-ray1.55 A
5EQ0CBX8 chromodomain with UNC3866X-ray1.18 A
2J2SMLL CXXC domain, non-methyl-CpG bindingNMR-
4NW3MLL CXXC domain with CpG DNAX-ray2.82 A

This project has already been bitten by unverified structural identifiers: the graph's own interface entry records that 4GQ3 and 4GQ4 were previously miscited as MBM1 peptide complexes when they are inhibitor complexes from a different paper, and that 6E1A is menin with compound M-89 rather than revumenib. Verification by fetch is not pedantry here; it is a correction of a known failure mode.

The AlphaFold confidence census

Fetched from https://alphafold.ebi.ac.uk/api/prediction/<UniProt> 2:

ProteinUniProtModelGlobal pLDDT
meninO00255AF-O00255-F184.44
KAT7O95251AF-O95251-F172.56
ENLQ03111AF-Q03111-F164.69
LEDGFO75475AF-O75475-F162.62
CBX8Q9HC52AF-Q9HC52-F161.75
DOT1LQ8TEK3AF-Q8TEK3-F152.50
AFF1P51825AF-P51825-F152.34
KMT2AQ03164none - HTTP 404n/a

A global pLDDT near 52 means the model is largely a disorder prediction with coordinates. A whole-protein pLDDT is nonetheless a crude summary: DOT1L at 52.5 has a perfectly well-folded catalytic domain inside a long disordered tail. The number should be used to decide which fragment to work on, never to decide whether a protein is druggable.

The access barrier

None for either resource. Both APIs are open, key-free and safe to automate. The barrier here is entirely the content: the two absences above cannot be worked around by asking for permission, only by computing or determining a structure.

What can be done about the KMT2A absence: the public AlphaFold database omits KMT2A because at 3,969 residues it exceeds the length cut-off for bulk prediction. A local prediction of selected KMT2A fragments is feasible on the Explorer GPU cluster, and that is the correct phase-2 response - fragment-wise, with the fragment boundaries chosen on domain annotation rather than arbitrarily.

Facts and figures
question answeredWhat is the shape of the target, at what resolution, with what ligand - and what is missing?
access barrierNone. The barrier is content: no full-length KMT2A model and no intact fusion structure.
recommended actionCompute KMT2A fragment models locally on the Explorer GPU cluster, with fragment boundaries chosen on domain annotation (AT-hooks, CXXC, MBM) rather than arbitrary windows.
statusrun
computelaptop for the census; Explorer Slurm GPU for local KMT2A fragment prediction
PDB verified this cycle
5EQ01.1805EQ0: 1.184GQ61.5504GQ6: 1.553I911.5503I91: 1.557UJ41.9607UJ4: 1.966HT12.1006HT1: 2.16MAJ2.1396MAJ: 2.1395J9S2.7025J9S: 2.7024NW32.8204NW3: 2.823U883.0003U88: 34HRA3.1504HRA: 3.15
resources
[
  "RCSB PDB",
  "AlphaFold Protein Structure Database"
]
access
{
  "rcsb": "https://data.rcsb.org/rest/v1/core/entry/<ID>",
  "alphafold": "https://alphafold.ebi.ac.uk/api/prediction/<UniProt>",
  "auth": "none for both"
}
PDB verified this cycle — the numbers behind the chart
[
  {
    "pdb": "4GQ6",
    "resolution_A": 1.55,
    "method": "X-ray"
  },
  {
    "pdb": "7UJ4",
    "resolution_A": 1.96,
    "method": "X-ray"
  },
  {
    "pdb": "3U88",
    "resolution_A": 3,
    "method": "X-ray"
  },
  {
    "pdb": "2MSR",
    "method": "NMR"
  },
  {
    "pdb": "4HRA",
    "resolution_A": 3.15,
    "method": "X-ray"
  },
  {
    "pdb": "5J9S",
    "resolution_A": 2.702,
    "method": "X-ray"
  },
  {
    "pdb": "6HT1",
    "resolution_A": 2.1,
    "method": "X-ray"
  },
  {
    "pdb": "6MAJ",
    "resolution_A": 2.139,
    "method": "X-ray"
  },
  {
    "pdb": "3I91",
    "resolution_A": 1.55,
    "method": "X-ray"
  },
  {
    "pdb": "5EQ0",
    "resolution_A": 1.18,
    "method": "X-ray"
  },
  {
    "pdb": "2J2S",
    "method": "NMR"
  },
  {
    "pdb": "4NW3",
    "resolution_A": 2.82,
    "method": "X-ray"
  }
]
alphafold census
{
  "O00255_menin": 84.44,
  "O95251_KAT7": 72.56,
  "Q03111_ENL": 64.69,
  "O75475_LEDGF": 62.62,
  "Q9HC52_CBX8": 61.75,
  "Q8TEK3_DOT1L": 52.5,
  "P51825_AFF1": 52.34,
  "Q03164_KMT2A": null
}
structural absences
[
  "no AlphaFold model of full-length KMT2A (Q03164, HTTP 404; 3,969 aa exceeds the public bulk-prediction length cut-off)",
  "no PDB structure of an intact KMT2A fusion protein",
  "no junction-spanning structure of any kind",
  "no experimental structure of KMT2A residues 1215-1563, the window containing every major-BCR fusion junction"
]
prior ID errors in this project
[
  "4GQ3 and 4GQ4 were previously miscited as MBM1 peptide complexes; they are inhibitor complexes from a different paper",
  "6E1A was miscited as menin with revumenib; it is menin with compound M-89"
]
kmt2a structural coverage
{
  "uniprot": "Q03164",
  "length_aa": 3969,
  "residues_solved": 629,
  "coverage_pct": 15.8,
  "n_islands": 7,
  "islands": [
    [
      4,
      25
    ],
    [
      103,
      160
    ],
    [
      747,
      755
    ],
    [
      1143,
      1214
    ],
    [
      1564,
      1784
    ],
    [
      2839,
      2869
    ],
    [
      3754,
      3969
    ]
  ],
  "largest_unsolved_gap": [
    1215,
    1563
  ],
  "note": "Computed 2026-09-19 by merging the 70 chain mappings returned by the PDBe SIFTS best_structures endpoint for Q03164. The 1215-1563 gap sits between the end of CXXC (~1195-1204) and the start of PHD1 (1431) and therefore contains the fusion junction of every major-BCR KMT2A rearrangement: the part of KMT2A that breaks has never been solved.",
  "source": "https://www.ebi.ac.uk/pdbe/api/mappings/best_structures/Q03164"
}
How this could bite us back
  • Verifying that an identifier resolves is not verifying that the structure supports the claim attached to it. Both errors already found in this project were of the second kind - real structures, wrong papers, wrong ligands.
  • A local AlphaFold or Boltz model of a KMT2A fragment will come back with coordinates regardless of whether the fragment folds. A low-confidence model rendered in a 3D viewer looks exactly as authoritative as a crystal structure, and the dashboard makes that risk worse, not better.
Still open
  1. Which KMT2A fragments are worth predicting locally, and is there a fragment boundary that would reveal a folded module near a recurrent breakpoint?
  2. Has any group deposited a fusion-protein structure under embargo, and would a query to the PDB pre-release listings reveal it?
  3. Should this project maintain an automated structural-identifier verifier that re-checks every PDB id in the graph on each build, given that two identifiers have already been found wrong?
Every source behind this entry (2)
  1. E1established confidence 5 of 5: Multiple independent lines of evidence; treated as settled in the field.database

    Twelve PDB entries cited across this beat were each verified by fetching their RCSB entry record this cycle, confirming title, experimental method and resolution; no entry exists for an intact KMT2A fusion protein or a junction-spanning construct.

    What this evidence cannot show

    Verification confirms that an identifier resolves to the structure described; it does not confirm the biological interpretation attached to it elsewhere in this graph. The absence of a fusion-protein structure is an absence in the PDB, which is comprehensive for deposited structures but says nothing about unpublished work.

    Evidence that argues against this (1)
    • Twelve verified entries is a census of what exists, and what exists for KMT2A is very little. The PDBe SIFTS best_structures mapping for Q03164, re-fetched on 2026-09-19, returns 70 chain mappings that collapse to seven disconnected islands - residues 4-25, 103-160, 747-755, 1143-1214, 1564-1784, 2839-2869 and 3754-3969 - covering 629 of 3,969 residues, 15.8% of the protein. More pointedly, nothing at all is solved between residue 1215 and residue 1563. CXXC ends around 1195-1204 and PHD1 begins at 1431, so the entire window in which every major-BCR KMT2A fusion junction falls is structurally blank. There is no structure of the part of KMT2A that actually breaks. EMBL-EBI PDBe 2026
  2. E2strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.database

    The AlphaFold prediction API returns models for menin (84.44 global pLDDT), KAT7 (72.56), ENL (64.69), LEDGF (62.62), CBX8 (61.75), DOT1L (52.50) and AFF1 (52.34), and returns HTTP 404 for KMT2A (Q03164), so no full-length KMT2A model exists in the public database.

    What this evidence cannot show

    Global pLDDT averages over the whole chain and hides well-folded domains inside disordered proteins; DOT1L at 52.5 has a fully ordered catalytic domain. The KMT2A 404 reflects the public database's length cut-off, not a claim that the protein cannot be modelled.

    Evidence that argues against this (2)
    • The KMT2A 404 was re-confirmed on 2026-09-19: the AlphaFold prediction API returns HTTP 404 with an empty body for Q03164. This is not a gap that computing a model locally fully closes. Ruff and Pappu warn that AlphaFold assigns very low confidence precisely to intrinsically disordered regions, that over 30% of the human proteome is disordered, and that treating those low-confidence coordinates as structures invites misinterpretation. KMT2A is largely disordered outside its small folded modules, so a locally computed fragment model will return coordinates for the breakpoint region that look authoritative in a 3D viewer and mean very little - which is this entry's own second bite_back item, now with a citation behind it. EMBL-EBI 2026
    • Ruff and Pappu, on the specific failure mode: many AlphaFold predictions feature regions of very low confidence that largely overlap intrinsically disordered regions, and drawing inferences about the 'structures' of IDRs and their interactomes from those predictions is a misinterpretation the field should avoid. For KMT2A this is not a hypothetical - the unsolved 1215-1563 window that contains the fusion junction is exactly the kind of region where a prediction will be confidently drawn and epistemically empty. Ruff KM et al. 2021
Connected to