Skip to content

Trainable, memory-efficient, and GPU-friendly PyTorch reproduction of AlphaFold 2

License

Notifications You must be signed in to change notification settings

johnypark/openfold

 
 

Repository files navigation

header Figure: Comparison of OpenFold and AlphaFold2 predictions to the experimental structure of PDB 7KDX, chain B.

OpenFold

A faithful but trainable PyTorch reproduction of DeepMind's AlphaFold 2.

Features

OpenFold carefully reproduces (almost) all of the features of the original open source inference code (v2.0.1). The sole exception is model ensembling, which fared poorly in DeepMind's own ablation testing and is being phased out in future DeepMind experiments. It is omitted here for the sake of reducing clutter. In cases where the Nature paper differs from the source, we always defer to the latter.

OpenFold is trainable in full precision, half precision, or bfloat16 with or without DeepSpeed, and we've trained it from scratch, matching the performance of the original. We've publicly released model weights and our training data — some 400,000 MSAs and PDB70 template hit files — under a permissive license. Model weights are available via scripts in this repository while the MSAs are hosted by the Registry of Open Data on AWS (RODA). Try out running inference for yourself with our Colab notebook.

OpenFold also supports inference using AlphaFold's official parameters, and vice versa (see scripts/convert_of_weights_to_jax.py).

OpenFold has the following advantages over the reference implementation:

  • Faster inference on GPU, sometimes by as much as 2x. The greatest speedups are achieved on (>= Ampere) GPUs.
  • Inference on extremely long chains, made possible by our implementation of low-memory attention (Rabe & Staats 2021). OpenFold can predict the structures of sequences with more than 4000 residues on a single A100, and even longer ones with CPU offloading.
  • Custom CUDA attention kernels modified from FastFold's kernels support in-place attention during inference and training. They use 4x and 5x less GPU memory than equivalent FastFold and stock PyTorch implementations, respectively.
  • Efficient alignment scripts using the original AlphaFold HHblits/JackHMMER pipeline or ColabFold's, which uses the faster MMseqs2 instead. We've used them to generate millions of alignments.
  • FlashAttention support greatly speeds up MSA attention.

Installation (Linux)

All Python dependencies are specified in environment.yml. For producing sequence alignments, you'll also need kalign, the HH-suite, and one of {jackhmmer, MMseqs2 (nightly build)} installed on on your system. You'll need git-lfs to download OpenFold parameters. Finally, some download scripts require aria2c and aws.

For convenience, we provide a script that installs Miniconda locally, creates a conda virtual environment, installs all Python dependencies, and downloads useful resources, including both sets of model parameters. Run:

scripts/install_third_party_dependencies.sh

To activate the environment, run:

source scripts/activate_conda_env.sh

To deactivate it, run:

source scripts/deactivate_conda_env.sh

With the environment active, compile OpenFold's CUDA kernels with

python3 setup.py install

To install the HH-suite to /usr/bin, run

# scripts/install_hh_suite.sh

Usage

If you intend to generate your own alignments, e.g. for inference, you have two choices for downloading protein databases, depending on whether you want to use DeepMind's MSA generation pipeline (w/ HMMR & HHblits) or ColabFold's, which uses the faster MMseqs2 instead. For the former, run:

bash scripts/download_alphafold_dbs.sh data/

For the latter, run:

bash scripts/download_mmseqs_dbs.sh data/    # downloads .tar files
bash scripts/prep_mmseqs_dbs.sh data/        # unpacks and preps the databases

Make sure to run the latter command on the machine that will be used for MSA generation (the script estimates how the precomputed database index used by MMseqs2 should be split according to the memory available on the system).

If you're using your own precomputed MSAs or MSAs from the RODA repository, there's no need to download these alignment databases. Simply make sure that the alignment_dir contains one directory per chain and that each of these contains alignments (.sto, .a3m, and .hhr) corresponding to that chain. You can use scripts/flatten_roda.sh to reformat RODA downloads in this way. Note that the RODA alignments are NOT compatible with the recent .cif ground truth files downloaded by scripts/download_alphafold_dbs.sh. To fetch .cif files that match the RODA MSAs, once the alignments are flattened, use scripts/download_roda_pdbs.sh. That script outputs a list of alignment dirs for which matching .cif files could not be found. These should be removed from the alignment directory.

Alternatively, you can use raw MSAs from ProteinNet. After downloading that database, use scripts/prep_proteinnet_msas.py to convert the data into a format recognized by the OpenFold parser. The resulting directory becomes the alignment_dir used in subsequent steps. Use scripts/unpack_proteinnet.py to extract .core files from ProteinNet text files.

For both inference and training, the model's hyperparameters can be tuned from openfold/config.py. Of course, if you plan to perform inference using DeepMind's pretrained parameters, you will only be able to make changes that do not affect the shapes of model parameters. For an example of initializing the model, consult run_pretrained_openfold.py.

Inference

To run inference on a sequence or multiple sequences using a set of DeepMind's pretrained parameters, run e.g.:

python3 run_pretrained_openfold.py \
    fasta_dir \
    data/pdb_mmcif/mmcif_files/ \
    --uniref90_database_path data/uniref90/uniref90.fasta \
    --mgnify_database_path data/mgnify/mgy_clusters_2018_12.fa \
    --pdb70_database_path data/pdb70/pdb70 \
    --uniclust30_database_path data/uniclust30/uniclust30_2018_08/uniclust30_2018_08 \
    --output_dir ./ \
    --bfd_database_path data/bfd/bfd_metaclust_clu_complete_id30_c90_final_seq.sorted_opt \
    --model_device "cuda:0" \
    --jackhmmer_binary_path lib/conda/envs/openfold_venv/bin/jackhmmer \
    --hhblits_binary_path lib/conda/envs/openfold_venv/bin/hhblits \
    --hhsearch_binary_path lib/conda/envs/openfold_venv/bin/hhsearch \
    --kalign_binary_path lib/conda/envs/openfold_venv/bin/kalign \
    --config_preset "model_1_ptm" \
    --openfold_checkpoint_path openfold/resources/openfold_params/finetuning_ptm_2.pt

where data is the same directory as in the previous step. If jackhmmer, hhblits, hhsearch and kalign are available at the default path of /usr/bin, their binary_path command-line arguments can be dropped. If you've already computed alignments for the query, you have the option to skip the expensive alignment computation here with --use_precomputed_alignments.

--openfold_checkpoint_path or --jax_param_path accept comma-delineated lists of .pt/DeepSpeed OpenFold checkpoints and AlphaFold's .npz JAX parameter files, respectively. For a breakdown of the differences between the different parameter files, see the README downloaded to openfold/resources/openfold_params/. Since OpenFold was trained under a newer training schedule than the one from which the model_n config presets are derived, there is no clean correspondence between config_preset settings and OpenFold checkpoints; the only restraints are that *_ptm checkpoints must be run with *_ptm config presets and that _no_templ_ checkpoints are only compatible with template-less presets (model_3 and above).

Note that chunking (as defined in section 1.11.8 of the AlphaFold 2 supplement) is enabled by default in inference mode. To disable it, set globals.chunk_size to None in the config. If a value is specified, OpenFold will attempt to dynamically tune it, considering the chunk size specified in the config as a minimum. This tuning process automatically ensures consistently fast runtimes regardless of input sequence length, but it also introduces some runtime variability, which may be undesirable for certain users. It is also recommended to disable this feature for very long chains (see below). To do so, set the tune_chunk_size option in the config to False.

For large-scale batch inference, we offer an optional tracing mode, which massively improves runtimes at the cost of a lengthy model compilation process. To enable it, add --trace_model to the inference command.

To get a speedup during inference, enable FlashAttention in the config. Note that it appears to work best for sequences with < 1000 residues.

Input FASTA files containing multiple sequences are treated as complexes. In this case, the inference script runs AlphaFold-Gap, a hack proposed here, using the specified stock AlphaFold/OpenFold parameters (NOT AlphaFold-Multimer). To run inference with AlphaFold-Multimer, use the (experimental) multimer branch instead.

To minimize memory usage during inference on long sequences, consider the following changes:

  • As noted in the AlphaFold-Multimer paper, the AlphaFold/OpenFold template stack is a major memory bottleneck for inference on long sequences. OpenFold supports two mutually exclusive inference modes to address this issue. One, average_templates in the template section of the config, is similar to the solution offered by AlphaFold-Multimer, which is simply to average individual template representations. Our version is modified slightly to accommodate weights trained using the standard template algorithm. Using said weights, we notice no significant difference in performance between our averaged template embeddings and the standard ones. The second, offload_templates, temporarily offloads individual template embeddings into CPU memory. The former is an approximation while the latter is slightly slower; both are memory-efficient and allow the model to utilize arbitrarily many templates across sequence lengths. Both are disabled by default, and it is up to the user to determine which best suits their needs, if either.
  • Inference-time low-memory attention (LMA) can be enabled in the model config. This setting trades off speed for vastly improved memory usage. By default, LMA is run with query and key chunk sizes of 1024 and 4096, respectively. These represent a favorable tradeoff in most memory-constrained cases. Powerusers can choose to tweak these settings in openfold/model/primitives.py. For more information on the LMA algorithm, see the aforementioned Staats & Rabe preprint.
  • Disable tune_chunk_size for long sequences. Past a certain point, it only wastes time.
  • As a last resort, consider enabling offload_inference. This enables more extensive CPU offloading at various bottlenecks throughout the model.
  • Disable FlashAttention, which seems unstable on long sequences.

Using the most conservative settings, we were able to run inference on a 4600-residue complex with a single A100. Compared to AlphaFold's own memory offloading mode, ours is considerably faster; the same complex takes the more efficent AlphaFold-Multimer more than double the time. Use the long_sequence_inference config option to enable all of these interventions at once. The run_pretrained_openfold.py script can enable this config option with the --long_sequence_inference command line option

Training

To train the model, you will first need to precompute protein alignments.

You have two options. You can use the same procedure DeepMind used by running the following:

python3 scripts/precompute_alignments.py mmcif_dir/ alignment_dir/ \
    --uniref90_database_path data/uniref90/uniref90.fasta \
    --mgnify_database_path data/mgnify/mgy_clusters_2018_12.fa \
    --pdb70_database_path data/pdb70/pdb70 \
    --uniclust30_database_path data/uniclust30/uniclust30_2018_08/uniclust30_2018_08 \
    --bfd_database_path data/bfd/bfd_metaclust_clu_complete_id30_c90_final_seq.sorted_opt \
    --cpus_per_task 16 \
    --jackhmmer_binary_path lib/conda/envs/openfold_venv/bin/jackhmmer \
    --hhblits_binary_path lib/conda/envs/openfold_venv/bin/hhblits \
    --hhsearch_binary_path lib/conda/envs/openfold_venv/bin/hhsearch \
    --kalign_binary_path lib/conda/envs/openfold_venv/bin/kalign

As noted before, you can skip the binary_path arguments if these binaries are at /usr/bin. Expect this step to take a very long time, even for small numbers of proteins.

Alternatively, you can generate MSAs with the ColabFold pipeline (and templates with HHsearch) with:

python3 scripts/precompute_alignments_mmseqs.py input.fasta \
    data/mmseqs_dbs \
    uniref30_2103_db \
    alignment_dir \
    ~/MMseqs2/build/bin/mmseqs \
    /usr/bin/hhsearch \
    --env_db colabfold_envdb_202108_db
    --pdb70 data/pdb70/pdb70

where input.fasta is a FASTA file containing one or more query sequences. To generate an input FASTA from a directory of mmCIF and/or ProteinNet .core files, we provide scripts/data_dir_to_fasta.py.

Next, generate a cache of certain datapoints in the template mmCIF files:

python3 scripts/generate_mmcif_cache.py \
    mmcif_dir/ \
    mmcif_cache.json \
    --no_workers 16

This cache is used to pre-filter templates.

Next, generate a separate chain-level cache with data used for training-time data filtering:

python3 scripts/generate_chain_data_cache.py \
    mmcif_dir/ \
    chain_data_cache.json \
    --cluster_file clusters-by-entity-40.txt \
    --no_workers 16

where the cluster_file argument is a file of chain clusters, one cluster per line (e.g. PDB40).

Optionally, download an AlphaFold-style validation set from CAMEO using scripts/download_cameo.py. Use the resulting FASTA files to generate validation alignments and then specify the validation set's location using the --val_... family of training script flags.

Finally, call the training script:

python3 train_openfold.py mmcif_dir/ alignment_dir/ template_mmcif_dir/ output_dir/ \
    2021-10-10 \ 
    --template_release_dates_cache_path mmcif_cache.json \ 
    --precision bf16 \
    --gpus 8 --replace_sampler_ddp=True \
    --seed 4242022 \ # in multi-gpu settings, the seed must be specified
    --deepspeed_config_path deepspeed_config.json \
    --checkpoint_every_epoch \
    --resume_from_ckpt ckpt_dir/ \
    --train_chain_data_cache_path chain_data_cache.json \
    --obsolete_pdbs_file_path obsolete.dat

where --template_release_dates_cache_path is a path to the mmCIF cache. Note that template_mmcif_dir can be the same as mmcif_dir which contains training targets. A suitable DeepSpeed configuration file can be generated with scripts/build_deepspeed_config.py. The training script is written with PyTorch Lightning and supports the full range of training options that entails, including multi-node distributed training, validation, and so on. For more information, consult PyTorch Lightning documentation and the --help flag of the training script.

Note that, despite its variable name, mmcif_dir can also contain PDB files or even ProteinNet .core files.

To emulate the AlphaFold training procedure, which uses a self-distillation set subject to special preprocessing steps, use the family of --distillation flags.

In cases where it may be burdensome to create separate files for each chain's alignments, alignment directories can be consolidated using the scripts in scripts/alignment_db_scripts/. First, run create_alignment_db.py to consolidate an alignment directory into a pair of database and index files. Once all alignment directories (or shards of a single alignment directory) have been compiled, unify the indices with unify_alignment_db_indices.py. The resulting index, super.index, can be passed to the training script flags containing the phrase alignment_index. In this scenario, the alignment_dir flags instead represent the directory containing the compiled alignment databases. Both the training and distillation datasets can be compiled in this way. Anecdotally, this can speed up training in I/O-bottlenecked environments.

Testing

To run unit tests, use

scripts/run_unit_tests.sh

The script is a thin wrapper around Python's unittest suite, and recognizes unittest arguments. E.g., to run a specific test verbosely:

scripts/run_unit_tests.sh -v tests.test_model

Certain tests require that AlphaFold (v2.0.1) be installed in the same Python environment. These run components of AlphaFold and OpenFold side by side and ensure that output activations are adequately similar. For most modules, we target a maximum pointwise difference of 1e-4.

Building and using the docker container

Building the docker image

Openfold can be built as a docker container using the included dockerfile. To build it, run the following command from the root of this repository:

docker build -t openfold .

Running the docker container

The built container contains both run_pretrained_openfold.py and train_openfold.py as well as all necessary software dependencies. It does not contain the model parameters, sequence, or structural databases. These should be downloaded to the host machine following the instructions in the Usage section above.

The docker container installs all conda components to the base conda environment in /opt/conda, and installs openfold itself in /opt/openfold,

Before running the docker container, you can verify that your docker installation is able to properly communicate with your GPU by running the following command:

docker run --rm --gpus all nvidia/cuda:11.0-base nvidia-smi

Note the --gpus all option passed to docker run. This option is necessary in order for the container to use the GPUs on the host machine.

To run the inference code under docker, you can use a command like the one below. In this example, parameters and sequences from the alphafold dataset are being used and are located at /mnt/alphafold_database on the host machine, and the input files are located in the current working directory. You can adjust the volume mount locations as needed to reflect the locations of your data.

docker run \
--gpus all \
-v $PWD/:/data \
-v /mnt/alphafold_database/:/database \
-ti openfold:latest \
python3 /opt/openfold/run_pretrained_openfold.py \
/data/fasta_dir \
/database/pdb_mmcif/mmcif_files/ \
--uniref90_database_path /database/uniref90/uniref90.fasta \
--mgnify_database_path /database/mgnify/mgy_clusters_2018_12.fa \
--pdb70_database_path /database/pdb70/pdb70 \
--uniclust30_database_path /database/uniclust30/uniclust30_2018_08/uniclust30_2018_08 \
--output_dir /data \
--bfd_database_path /database/bfd/bfd_metaclust_clu_complete_id30_c90_final_seq.sorted_opt \
--model_device cuda:0 \
--jackhmmer_binary_path /opt/conda/bin/jackhmmer \
--hhblits_binary_path /opt/conda/bin/hhblits \
--hhsearch_binary_path /opt/conda/bin/hhsearch \
--kalign_binary_path /opt/conda/bin/kalign \
--openfold_checkpoint_path /database/openfold_params/finetuning_ptm_2.pt

Copyright notice

While AlphaFold's and, by extension, OpenFold's source code is licensed under the permissive Apache Licence, Version 2.0, DeepMind's pretrained parameters fall under the CC BY 4.0 license, a copy of which is downloaded to openfold/resources/params by the installation script. Note that the latter replaces the original, more restrictive CC BY-NC 4.0 license as of January 2022.

Contributing

If you encounter problems using OpenFold, feel free to create an issue! We also welcome pull requests from the community.

Citing this work

Please cite our paper:

@article {Ahdritz2022.11.20.517210,
	author = {Ahdritz, Gustaf and Bouatta, Nazim and Kadyan, Sachin and Xia, Qinghui and Gerecke, William and O{\textquoteright}Donnell, Timothy J and Berenberg, Daniel and Fisk, Ian and Zanichelli, Niccolò and Zhang, Bo and Nowaczynski, Arkadiusz and Wang, Bei and Stepniewska-Dziubinska, Marta M and Zhang, Shang and Ojewole, Adegoke and Guney, Murat Efe and Biderman, Stella and Watkins, Andrew M and Ra, Stephen and Lorenzo, Pablo Ribalta and Nivon, Lucas and Weitzner, Brian and Ban, Yih-En Andrew and Sorger, Peter K and Mostaque, Emad and Zhang, Zhao and Bonneau, Richard and AlQuraishi, Mohammed},
	title = {OpenFold: Retraining AlphaFold2 yields new insights into its learning mechanisms and capacity for generalization},
	elocation-id = {2022.11.20.517210},
	year = {2022},
	doi = {10.1101/2022.11.20.517210},
	publisher = {Cold Spring Harbor Laboratory},
	abstract = {AlphaFold2 revolutionized structural biology with the ability to predict protein structures with exceptionally high accuracy. Its implementation, however, lacks the code and data required to train new models. These are necessary to (i) tackle new tasks, like protein-ligand complex structure prediction, (ii) investigate the process by which the model learns, which remains poorly understood, and (iii) assess the model{\textquoteright}s generalization capacity to unseen regions of fold space. Here we report OpenFold, a fast, memory-efficient, and trainable implementation of AlphaFold2, and OpenProteinSet, the largest public database of protein multiple sequence alignments. We use OpenProteinSet to train OpenFold from scratch, fully matching the accuracy of AlphaFold2. Having established parity, we assess OpenFold{\textquoteright}s capacity to generalize across fold space by retraining it using carefully designed datasets. We find that OpenFold is remarkably robust at generalizing despite extreme reductions in training set size and diversity, including near-complete elisions of classes of secondary structure elements. By analyzing intermediate structures produced by OpenFold during training, we also gain surprising insights into the manner in which the model learns to fold proteins, discovering that spatial dimensions are learned sequentially. Taken together, our studies demonstrate the power and utility of OpenFold, which we believe will prove to be a crucial new resource for the protein modeling community.},
	URL = {https://www.biorxiv.org/content/10.1101/2022.11.20.517210},
	eprint = {https://www.biorxiv.org/content/early/2022/11/22/2022.11.20.517210.full.pdf},
	journal = {bioRxiv}
}

Any work that cites OpenFold should also cite AlphaFold.

About

Trainable, memory-efficient, and GPU-friendly PyTorch reproduction of AlphaFold 2

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages

  • Python 91.6%
  • Jupyter Notebook 3.8%
  • Shell 3.5%
  • Cuda 0.7%
  • C++ 0.2%
  • Dockerfile 0.2%