OpenFold3

OpenFold3 provides structure prediction capabilities for various molecule types. While all features described in the AlphaFold3 Nature paper are supported in training, inference support is being rolled out incrementally.

Supported Molecule Types

Protein

• MSA-based prediction: Supports the ColabFold MSA pipeline, pre-computed MSAs, and OpenFold3's native MSA generation pipeline.
• MSA-free prediction: Prediction without Multiple Sequence Alignments.
• Template-based prediction: Supports ColabFold template alignments, pre-computed template alignments, and direct CIF template files (no alignments required).
• Non-canonical residues.

DNA

• MSA-free prediction: Prediction without MSA (following AF3 defaults).
• Non-canonical residues.

RNA

• MSA-based prediction: Uses OpenFold3's native MSA generation pipeline.
• MSA-free prediction.
• Non-canonical residues.

Ligand

• Non-covalent ligands.

Coming Soon

• Covalently modified residues and cross-chain covalent bonds (Protein, DNA, RNA).
• Template-based prediction for RNA.
• Protein-RNA MSA pairing.
• Covalently bound ligands and polymeric ligands (e.g., glycans).

When --use-msa-server=True is enabled, OpenFold3 generates processed MSAs as .npz files. These are stored in a dedicated MSA output directory.

Directory Structure

• main/: Contains .npz files for each unique chain (unpaired MSAs).
• paired/: Contains .npz files for heteromeric protein complexes (paired MSAs). The filename is a hash of the concatenated sequences.
• mappings/: Contains metadata to map hashes back to sequences and chain IDs:
  • chain_id_to_rep_id.json
  • query_name_to_complex_id.json
  • rep_id_to_seq.json (hash to sequence)
  • seq_to_rep_id.json (sequence to hash)
• raw/: Contains the original ColabFold .a3m files. Note: This directory is deleted by default after processing to save space. To keep them, you must modify the source code in colabfold_msa_server.py.

<msa_output_directory>
├── main
│   ├── <hash of sequence A>.npz
│   └── <hash of sequence B>.npz
├── mappings
│   ├── chain_id_to_rep_id.json
│   ├── query_name_to_complex_id.json
│   ├── README.md
│   ├── rep_id_to_seq.json
│   └── seq_to_rep_id.json
└── paired
    └── <hash of concatenation of sequences A and B>
        ├── <hash of sequence A>.npz
        └── <hash of sequence B>.npz

Reference molecules are stored as SDF files in the reference_molecules/ folder using the V2000 molfile format. They include two custom annotation fields for integration with AtomArrays:

1. atom_annot_atom_name: Canonical atom names for the ligand.
2. atom_annot_used_atom_mask: A per-atom mask. True for all atoms if a conformer was successfully generated. If CCD-deposited fallback coordinates were used, atoms with missing coordinates are marked False.

     RDKit          3D

  4  3  0  0  0  0  0  0  0  0999 V2000
    0.8140   -0.0385    0.0035 C   0  0  0  0  0  0  0  0  0  0  0  0
    1.5232   -1.0718    0.0134 O   0  0  0  0  0  0  0  0  0  0  0  0
    1.5226    1.1766   -0.0223 O   0  0  0  0  0  1  0  0  0  0  0  0
   -0.6545   -0.0200    0.0167 C   0  0  0  0  0  0  0  0  0  0  0  0
  1  2  2  0
  1  3  1  0
  1  4  1  0
M  CHG  1   3  -1
M  END

>  <atom_annot_atom_name>  (1)
C O OXT CH3

>  <atom_annot_used_atom_mask>  (1)
True True True True

$$$$

OpenFold3 supports two distinct modes for using template information during inference. Choosing between them depends on whether you have precomputed alignments or only have structure files.

1. Alignment-Based Mode (Traditional): This mode is used when you have existing query-to-template alignments. It requires two components:

   • Template Alignments: Files describing how the query sequence relates to the template sequences.
   • Template Structures: The actual structural files (currently only .cif format supported).

2. CIF Direct Mode (Simplified): This mode is used when you only have template structure files. The system automatically handles the alignment of template chains to your query sequence and selects the best match. You only need to provide the .cif files.

Note: The template pipeline currently supports monomeric templates and is tested for protein chains only.

OpenFold allows the use of LLMs and coding agents, but contributors must adhere to the following quality and engagement standards to conserve core team resources:

• Human Responsibility: All contributions (AI-generated or otherwise) are the sole responsibility of the human contributor. You must fully understand, review, and verify that the code works as intended before submission. Always provide test examples demonstrating functionality.
• Human-Written Communication: Issues and Pull Requests must be written by humans to facilitate meaningful dialogue with the development team. Exceptions are permitted for translations or generating failing test examples, provided the use of AI tools is disclosed.
• Good First Issues: These issues are reserved for new human developers to learn the workflow. Agentic/AI contributions on issues labeled good first issue will be closed.

Failure to meet these standards may result in the closure of contributions.

The validation cache (validation_cache.json) is used for proper model evaluation. It follows a similar structure to the training cache but includes additional fields for homology and quality assessment.

To create this cache, you must use the scripts/data_preprocessing/create_pdb_validation_dataset_cache.py script. Note that this process requires the training cache as input to perform homology comparisons.

Key characteristics of the validation cache include:

• Stricter filtering: Requires resolution $\le$ 4.5 Angstrom and adheres to specific token limits.
• Homology detection: Uses a 40% sequence identity threshold for proteins and a 0.85 Tanimoto similarity threshold for ligands.
• Subsets: Contains both multimer and monomer validation sets.

"7vl5": {
    "release_date": "2022-03-09",
    "resolution": 1.93,
    "token_count": 1463,
    "chains": {
        "1": {
            "label_asym_id": "A",
            "auth_asym_id": "A",
            "entity_id": 1,
            "molecule_type": "PROTEIN",
            "reference_mol_id": null,
            "alignment_representative_id": "7vl6_A",
            "template_ids": [
                "5gsl_A",
                "5gsl_B",
                "6jow_A"
            ],
            "cluster_id": "2024",
            "cluster_size": null,
            "low_homology": true,
            "metric_eligible": true,
            "use_metrics": false,
            "ranking_model_fit": null,
            "source_subset": "base"
        },
        "3": {
            "label_asym_id": "C",
            "auth_asym_id": "A",
            "entity_id": 2,
            "molecule_type": "LIGAND",
            "reference_mol_id": "BOG",
            "alignment_representative_id": null,
            "template_ids": null,
            "cluster_id": "5086",
            "cluster_size": null,
            "low_homology": false,
            "metric_eligible": false,
            "use_metrics": false,
            "ranking_model_fit": 0.9312,
            "source_subset": "base"
        }
    },
    "interfaces": {
        "1_3": {
            "cluster_id": "2024_5086",
            "cluster_size": null,
            "low_homology": true,
            "metric_eligible": true,
            "use_metrics": false,
            "source_subset": "base"
        }
    }
}