Inaccurate modelling of dihydrouridine RNA modifications

[AlexHurtado2000]

Dear Alphafold3 team,

We are interested in modelling RNAs with dihydrouridine modifications. We tried using the H2U modificationType, as well as a CIF file for the modified residue. In both cases AF3 is able to complete the model prediction, but upon visual inspection, the target residues present a planar nitrogenous base similar to uridine and not dihydrouridine (which should be non-planar); these residues successfully show H2U as their descriptor.

Environment:

• OS: Ubuntu 22.04.4 LTS
• MPI runtime: Open MPI 4.1.2
• Memory: 128 GB
• GPUs: 2X Nvidia RTX 4070 Super
• Cuda version: 12.6

The json file inputed should be as below:

  "name": "Example1",
  "modelSeeds": [1],
  "sequences": [
    {
      "rna": {
        "id": "A",
        "sequence": "CCUUCGAUAGCUCAGUUGGUAGAGCGGAGGACUGUAGAUCCUUAGGUCGCUGGUUCGAAUCCGGCUCGAAGGA",
        "modifications": [
          {
            "modificationType": "H2U",
            "basePosition": 16
          }
        ]
      }
    }
  ],
  "dialect": "alphafold3",
  "version": 4

After the above ran to completion, the modification looked to be a uridine however it was assigned as a dihydrouridine, which was not intended. As such, we reattempted this with a user-defined cif file obtained from the PDB H2U.

    "name": "Example2",
    "modelSeeds": [1],
    "sequences": [
        {
            "rna": {
                "id": "A",
                "sequence": "CCUUCGAUAGCUCAGUUGGUAGAGCGGAGGACUGUAGAUCCUUAGGUCGCUGGUUCGAAUCCGGCUCGAAGGA",
                "modifications": [
                    {"modificationType": "MYMOD", "basePosition": 16}
                ],
            }
        }
    ],
    "userCCDPath": "MYMOD.cif",
    "dialect": "alphafold3",
    "version": 4

If more information is required please let me know.

Thank you very much.

Kind regards,
Alex