Protenix Documentation

Protenix organizes Multiple Sequence Alignment (MSA) data using an integer-indexed directory structure. Each unique protein sequence is assigned an integer index (e.g., 0, 1, 2), which corresponds to a subdirectory under the root MSA directory.

To resolve a sequence to its directory ID, the data loader uses a seq_to_pdb_index.json file located in a common directory.

Directory Layout Example:

├── common
│   ├── seq_to_pdb_index.json # Mapping from sequence to directory ID
│   └── ...                   # Other metadata
├── mmcif_msa_template                 # Root directory for MSA files
│   ├── 0
│   │   ├── pairing.a3m       # MSAs paired by taxonomy ID
│   │   ├── non_pairing.a3m   # Unpaired MSAs
│   │   └── hmmsearch.a3m     # Templates from HMMER search
│   ├── 1
│   │   ├── pairing.a3m
│   │   ├── non_pairing.a3m
│   │   └── hmmsearch.a3m
│   └── ...

The protenix_base_constraint_v0.5.0 model allows you to incorporate experimental structural constraints (such as Pocket or Contact information) during inference.

This model includes specific embedders to handle these inputs:

• pocket_embedder: Processes binding pocket information.
• contact_embedder: Processes contact point information.

The input JSON file is a list of dictionaries, where each dictionary represents a single inference job. Even for a single set of sequences, the top-level structure must be a list.

Each job dictionary contains:

• name: (string) The name of the inference job.
• sequences: (list of dicts) Descriptions of entities like proteins, DNA, RNA, ligands, or ions.
• covalent_bonds: (optional list of dicts) Definitions of covalent bonds between atoms from different entities.

[
  {
    "name": "Test Fold Job Number One",
    "sequences": [...],
    "covalent_bonds": [...]
  }
]

Protenix model names follow a structured format to identify their size, functional features, and version: protenix_{model_size}_{features}_{version}.

• model_size: Defines the scale of the model. Options include base (full accuracy), mini (lightweight), and tiny (minimal).
• features: Describes functional capabilities. Examples include default, constraint (supports distance constraints), and esm (includes ESM embeddings). Multiple features are separated by a hyphen (-).
• version: The specific version number (e.g., v0.5.0, v1.0.0).

ESM and ISM models integrate single-sequence protein language models (ESM2-3B). These are ideal for scenarios where Multiple Sequence Alignments (MSAs) are unavailable.

• ESM models: Use standard ESM2 embeddings.
• ISM models: Use specific ISM embeddings.

Note: For efficiency, these models do not use MSA by default.

Protenix-Mini is a lightweight variant designed for efficient prediction of biomolecular complex structures. It achieves efficiency by using reduced network blocks and a minimal number of ODE steps (as few as one or two steps).

Key Characteristics:

• Efficiency: Optimized for high-throughput and resource-limited environments.
• Accuracy: Maintains a favorable balance, typically showing only a 1–5% drop in metrics like interface LDDT, complex LDDT, and ligand RMSD success rate compared to larger models.
• Use Case: Best suited for large-scale applications where speed and resource management are prioritized over absolute maximum precision.

Protenix supports using constraints to provide extra guidance during structure prediction, which can significantly improve accuracy. You can specify:

• Contacts: At both the residue and atom levels.
• Pocket constraints

For detailed information on the required JSON input format for these constraints, refer to the infer_json_format.md documentation.

To analyze the performance of Protenix v1.0.0, follow these steps:

1. Download: Obtain the relevant .tar.gz file from the benchmark data table.
2. Extract: Decompress the file using tar:

   tar -xzvf <dataset_name>.tar.gz

3. Review Data:
   • Inference Metrics: Look for CSV files containing per-prediction metrics such as LDDT, RMSD, and confidence scores for each sample.
   • Aggregated Results: Review summary reports for statistical overviews across the entire dataset to assess general performance.

After installing the dependencies, clone the ColabFold repository and use the provided setup script to initialize the databases in your target directory.

Steps:

1. Clone the ColabFold repository.
2. Run ./setup_databases.sh with MMSEQS_NO_INDEX=1 to configure the database at the specified path.

git clone https://github.com/sokrypton/ColabFold.git
cd ColabFold
MMSEQS_NO_INDEX=1 ./setup_databases.sh <path/to/colabfold_db>