Skip to content
Switch branches/tags



This package provides an implementation of the inference pipeline of AlphaFold v2.0. This is a completely new model that was entered in CASP14 and published in Nature. For simplicity, we refer to this model as AlphaFold throughout the rest of this document.

Any publication that discloses findings arising from using this source code or the model parameters should cite the AlphaFold paper. Please also refer to the Supplementary Information for a detailed description of the method.

You can use a slightly simplified version of AlphaFold with this Colab notebook or community-supported versions (see below).

CASP14 predictions

First time setup

The following steps are required in order to run AlphaFold:

  1. Install Docker.

  2. Download genetic databases (see below).

  3. Download model parameters (see below).

  4. Check that AlphaFold will be able to use a GPU by running:

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

    The output of this command should show a list of your GPUs. If it doesn't, check if you followed all steps correctly when setting up the NVIDIA Container Toolkit or take a look at the following NVIDIA Docker issue.

Genetic databases

This step requires aria2c to be installed on your machine.

AlphaFold needs multiple genetic (sequence) databases to run:

We provide a script scripts/ that can be used to download and set up all of these databases:

  • Default:

    scripts/ <DOWNLOAD_DIR>

    will download the full databases.

  • With reduced_dbs:

    scripts/ <DOWNLOAD_DIR> reduced_dbs

    will download a reduced version of the databases to be used with the reduced_dbs preset.

We don't provide exactly the versions used in CASP14 -- see the note on reproducibility. Some of the databases are mirrored for speed, see mirrored databases.

📒 Note: The total download size for the full databases is around 415 GB and the total size when unzipped is 2.2 TB. Please make sure you have a large enough hard drive space, bandwidth and time to download. We recommend using an SSD for better genetic search performance.

This script will also download the model parameter files. Once the script has finished, you should have the following directory structure:

$DOWNLOAD_DIR/                             # Total: ~ 2.2 TB (download: 438 GB)
    bfd/                                   # ~ 1.7 TB (download: 271.6 GB)
        # 6 files.
    mgnify/                                # ~ 64 GB (download: 32.9 GB)
    params/                                # ~ 3.5 GB (download: 3.5 GB)
        # 5 CASP14 models,
        # 5 pTM models,
        # LICENSE,
        # = 11 files.
    pdb70/                                 # ~ 56 GB (download: 19.5 GB)
        # 9 files.
    pdb_mmcif/                             # ~ 206 GB (download: 46 GB)
            # About 180,000 .cif files.
    small_bfd/                             # ~ 17 GB (download: 9.6 GB)
    uniclust30/                            # ~ 86 GB (download: 24.9 GB)
            # 13 files.
    uniref90/                              # ~ 58 GB (download: 29.7 GB)

bfd/ is only downloaded if you download the full databasees, and small_bfd/ is only downloaded if you download the reduced databases.

Model parameters

While the AlphaFold code is licensed under the Apache 2.0 License, the AlphaFold parameters are made available for non-commercial use only under the terms of the CC BY-NC 4.0 license. Please see the Disclaimer below for more detail.

The AlphaFold parameters are available from, and are downloaded as part of the scripts/ script. This script will download parameters for:

  • 5 models which were used during CASP14, and were extensively validated for structure prediction quality (see Jumper et al. 2021, Suppl. Methods 1.12 for details).
  • 5 pTM models, which were fine-tuned to produce pTM (predicted TM-score) and predicted aligned error values alongside their structure predictions (see Jumper et al. 2021, Suppl. Methods 1.9.7 for details).

Running AlphaFold

The simplest way to run AlphaFold is using the provided Docker script. This was tested on Google Cloud with a machine using the nvidia-gpu-cloud-image with 12 vCPUs, 85 GB of RAM, a 100 GB boot disk, the databases on an additional 3 TB disk, and an A100 GPU.

  1. Clone this repository and cd into it.

    git clone
  2. Modify DOWNLOAD_DIR in docker/ to be the path to the directory containing the downloaded databases.

  3. Build the Docker image:

    docker build -f docker/Dockerfile -t alphafold .
  4. Install the dependencies. Note: You may optionally wish to create a Python Virtual Environment to prevent conflicts with your system's Python environment.

    pip3 install -r docker/requirements.txt
  5. Run pointing to a FASTA file containing the protein sequence for which you wish to predict the structure. If you are predicting the structure of a protein that is already in PDB and you wish to avoid using it as a template, then max_template_date must be set to be before the release date of the structure. For example, for the T1050 CASP14 target:

    python3 docker/ --fasta_paths=T1050.fasta --max_template_date=2020-05-14

    By default, Alphafold will attempt to use all visible GPU devices. To use a subset, specify a comma-separated list of GPU UUID(s) or index(es) using the --gpu_devices flag. See GPU enumeration for more details.

  6. You can control AlphaFold speed / quality tradeoff by adding --preset=reduced_dbs, --preset=full_dbs or --preset=casp14 to the run command. We provide the following presets:

    • reduced_dbs: This preset is optimized for speed and lower hardware requirements. It runs with a reduced version of the BFD database and with no ensembling. It requires 8 CPU cores (vCPUs), 8 GB of RAM, and 600 GB of disk space.
    • full_dbs: The model in this preset is 8 times faster than the casp14 preset with a very minor quality drop (-0.1 average GDT drop on CASP14 domains). It runs with all genetic databases and with no ensembling.
    • casp14: This preset uses the same settings as were used in CASP14. It runs with all genetic databases and with 8 ensemblings.

    Running the command above with the casp14 preset would look like this:

    python3 docker/ --fasta_paths=T1050.fasta --max_template_date=2020-05-14 --preset=casp14

AlphaFold output

The outputs will be in a subfolder of output_dir in They include the computed MSAs, unrelaxed structures, relaxed structures, ranked structures, raw model outputs, prediction metadata, and section timings. The output_dir directory will have the following structure:


The contents of each output file are as follows:

  • features.pkl – A pickle file containing the input feature NumPy arrays used by the models to produce the structures.

  • unrelaxed_model_*.pdb – A PDB format text file containing the predicted structure, exactly as outputted by the model.

  • relaxed_model_*.pdb – A PDB format text file containing the predicted structure, after performing an Amber relaxation procedure on the unrelaxed structure prediction (see Jumper et al. 2021, Suppl. Methods 1.8.6 for details).

  • ranked_*.pdb – A PDB format text file containing the relaxed predicted structures, after reordering by model confidence. Here ranked_0.pdb should contain the prediction with the highest confidence, and ranked_4.pdb the prediction with the lowest confidence. To rank model confidence, we use predicted LDDT (pLDDT) scores (see Jumper et al. 2021, Suppl. Methods 1.9.6 for details).

  • ranking_debug.json – A JSON format text file containing the pLDDT values used to perform the model ranking, and a mapping back to the original model names.

  • timings.json – A JSON format text file containing the times taken to run each section of the AlphaFold pipeline.

  • msas/ - A directory containing the files describing the various genetic tool hits that were used to construct the input MSA.

  • result_model_*.pkl – A pickle file containing a nested dictionary of the various NumPy arrays directly produced by the model. In addition to the output of the structure module, this includes auxiliary outputs such as:

    • Distograms (distogram/logits contains a NumPy array of shape [N_res, N_res, N_bins] and distogram/bin_edges contains the definition of the bins).
    • Per-residue pLDDT scores (plddt contains a NumPy array of shape [N_res] with the range of possible values from 0 to 100, where 100 means most confident). This can serve to identify sequence regions predicted with high confidence or as an overall per-target confidence score when averaged across residues.
    • Present only if using pTM models: predicted TM-score (ptm field contains a scalar). As a predictor of a global superposition metric, this score is designed to also assess whether the model is confident in the overall domain packing.
    • Present only if using pTM models: predicted pairwise aligned errors (predicted_aligned_error contains a NumPy array of shape [N_res, N_res] with the range of possible values from 0 to max_predicted_aligned_error, where 0 means most confident). This can serve for a visualisation of domain packing confidence within the structure.

The pLDDT confidence measure is stored in the B-factor field of the output PDB files (although unlike a B-factor, higher pLDDT is better, so care must be taken when using for tasks such as molecular replacement).

This code has been tested to match mean top-1 accuracy on a CASP14 test set with pLDDT ranking over 5 model predictions (some CASP targets were run with earlier versions of AlphaFold and some had manual interventions; see our forthcoming publication for details). Some targets such as T1064 may also have high individual run variance over random seeds.

Inferencing many proteins

The provided inference script is optimized for predicting the structure of a single protein, and it will compile the neural network to be specialized to exactly the size of the sequence, MSA, and templates. For large proteins, the compile time is a negligible fraction of the runtime, but it may become more significant for small proteins or if the multi-sequence alignments are already precomputed. In the bulk inference case, it may make sense to use our make_fixed_size function to pad the inputs to a uniform size, thereby reducing the number of compilations required.

We do not provide a bulk inference script, but it should be straightforward to develop on top of the RunModel.predict method with a parallel system for precomputing multi-sequence alignments. Alternatively, this script can be run repeatedly with only moderate overhead.

Note on reproducibility

AlphaFold's output for a small number of proteins has high inter-run variance, and may be affected by changes in the input data. The CASP14 target T1064 is a notable example; the large number of SARS-CoV-2-related sequences recently deposited changes its MSA significantly. This variability is somewhat mitigated by the model selection process; running 5 models and taking the most confident.

To reproduce the results of our CASP14 system as closely as possible you must use the same database versions we used in CASP. These may not match the default versions downloaded by our scripts.

For genetics:

For templates:

An alternative for templates is to use the latest PDB and PDB70, but pass the flag --max_template_date=2020-05-14, which restricts templates only to structures that were available at the start of CASP14.

Citing this work

If you use the code or data in this package, please cite:

  author  = {Jumper, John and Evans, Richard and Pritzel, Alexander and Green, Tim and Figurnov, Michael and Ronneberger, Olaf and Tunyasuvunakool, Kathryn and Bates, Russ and {\v{Z}}{\'\i}dek, Augustin and Potapenko, Anna and Bridgland, Alex and Meyer, Clemens and Kohl, Simon A A and Ballard, Andrew J and Cowie, Andrew and Romera-Paredes, Bernardino and Nikolov, Stanislav and Jain, Rishub and Adler, Jonas and Back, Trevor and Petersen, Stig and Reiman, David and Clancy, Ellen and Zielinski, Michal and Steinegger, Martin and Pacholska, Michalina and Berghammer, Tamas and Bodenstein, Sebastian and Silver, David and Vinyals, Oriol and Senior, Andrew W and Kavukcuoglu, Koray and Kohli, Pushmeet and Hassabis, Demis},
  journal = {Nature},
  title   = {Highly accurate protein structure prediction with {AlphaFold}},
  year    = {2021},
  volume  = {596},
  number  = {7873},
  pages   = {583--589},
  doi     = {10.1038/s41586-021-03819-2}

Community contributions

Colab notebooks provided by the community (please note that these notebooks may vary from our full AlphaFold system and we did not validate their accuracy):


AlphaFold communicates with and/or references the following separate libraries and packages:

We thank all their contributors and maintainers!

License and Disclaimer

This is not an officially supported Google product.

Copyright 2021 DeepMind Technologies Limited.

AlphaFold Code License

Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with the License. You may obtain a copy of the License at

Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License.

Model Parameters License

The AlphaFold parameters are made available for non-commercial use only, under the terms of the Creative Commons Attribution-NonCommercial 4.0 International (CC BY-NC 4.0) license. You can find details at:

Third-party software

Use of the third-party software, libraries or code referred to in the Acknowledgements section above may be governed by separate terms and conditions or license provisions. Your use of the third-party software, libraries or code is subject to any such terms and you should check that you can comply with any applicable restrictions or terms and conditions before use.

Mirrored Databases

The following databases have been mirrored by DeepMind, and are available with reference to the following: