OpenFold Documentation

OpenFold supports three distinct modes of inference prediction:

• Monomer: Standard single-chain prediction.
• Multimer: Prediction for protein complexes.
• Single Sequence (Soloseq): Prediction based on a single sequence without MSA.

This documentation provides specific guides for each mode (e.g., Multimer_Inference.md and Single_Sequence_Inference.md).

OpenFold's source code is licensed under the Apache Licence, Version 2.0.

DeepMind's pretrained parameters are licensed under CC BY 4.0. A copy of this license is downloaded to openfold/resources/params by the installation script. (Note: This replaced the more restrictive CC BY-NC 4.0 license as of January 2022).

OpenFold uses a condensed filesystem structure to minimize I/O overhead during training. Instead of a directory per protein chain, it uses consolidated database files and index files.

A typical OpenProteinSet directory structure looks like this:

- OpenProteinSet 
  ├── duplicate_pdb_chains.txt
  └── pdb
	  ├── mmcif_cache.json 
	  ├── mmcifs 
	  │   ├── 3lrm.cif
	  │   └── 6kwc.cif
	  └── alignment_db
	      ├── alignment_db_0.db 
	      ├── alignment_db_1.db 
	      │   ...
	      └── alignment_db.index 

Download the required SoloSeq model weights using the provided shell script.

bash scripts/download_openfold_soloseq_params.sh openfold/resources

For systems with I/O bottlenecks, you can use alignment_db files instead of standard alignment directories. This increases training performance. It is recommended to create 10 shards for better filesystem health. The script runs optimally if the number of CPUs is $\ge$ the number of shards.

python $OF_DIR/scripts/alignment_db_scripts/create_alignment_db_sharded.py \
    alignment_data/alignments \
    alignment_data/alignment_dbs \
    alignment_db \
    --n_shards 10 \
    --duplicate_chains_file pdb_data/duplicate_pdb_chains.txt

If the aws CLI is unavailable, use one of the following scripts to download parameters:

• HuggingFace (requires git-lts): ./scripts/download_openfold_params_huggingface.sh
• Google Drive: ./scripts/download_openfold_params_gdrive.sh

If you have checkpoints trained with OpenFold v1 or older and want to resume training on OpenFold v2, you must convert them because certain model layers were renamed (e.g., module.model.template_angle_embedder.* is now module.model.template_embedder.template_single_embedder.*).

Use the scripts/convert_v1_to_v2_weights.py script to perform the conversion.

python scripts/convert_v1_to_v2_weights.py checkpoints/6-209.ckpt checkpoints/6-209.ckpt.converted

To improve I/O performance in bottlenecked environments, you can consolidate alignment directories into a single database and index using scripts/alignment_db_scripts/:

1. Run create_alignment_db.py to compile an alignment directory into database and index files.
2. Run unify_alignment_db_indices.py to unify the indices into a super.index.
3. Pass the resulting index to the training script using flags containing the phrase alignment_index. In this case, alignment_dir flags should point to the directory containing the compiled databases.

Run MSA-free sequence-to-structure prediction using precomputed ESM-1b embeddings. If you wish to use templates, ensure *.hhr files are in the embeddings_output_dir and provide the PDB MMCIF dataset path.

Note: The SoloSeq model is limited to a sequence length of 1022 residues; sequences longer than this will be truncated.

python run_pretrained_openfold.py \
    fasta_dir \
    data/pdb_mmcif/mmcif_files/ \
    --use_precomputed_alignments embeddings_output_dir \
    --output_dir ./ \
    --model_device "cuda:0" \
    --config_preset "seq_model_esm1b_ptm" \
    --openfold_checkpoint_path openfold/resources/openfold_soloseq_params/seq_model_esm1b_ptm.pt