AlphaFold3

Description

AlphaFold is an artificial intelligence (AI) program developed by DeepMind, a subsidiary of Alphabet, which performs predictions of protein structure.

The software provided on this workspace is an inference pipeline for the AlphaFold3 model. To utilize the model, you will also need to request access to the model parameters.

The recommended iBridges client for Yoda and iRODS is also preinstalled, to transfer data to or from the workspace.

Creating an alphafold instance in VRE

Preparation

Data parameters

The data parameters need to be requested by your own group by filling in this form. According to the alphafold documentation, you will get a response within 2-3 business days.

Download the model parameter to your local computer.

Creating a workspace and storage

Storage

The AlphaFold catalog item requires a storage unit. If no storage is attached to the workspace, creation will fail. See the Getting started page for more info about how and why to create a storage volume.

  • Make sure to choose the capacity of 1.5 TB for the storage.
  • On the workspace, the shell variable $ALPHAFOLD_STORAGE will be set to /data/<your_storage_name>/alphafold3. You can use this variable to specify parameters when running AlphaFold.

Create a workspace

In the Research Cloud portal click the ‘Create a new workspace’ button and follow the steps in the wizzard.

See the workspace creation manual page for more guidance.

  • When selecting the hardware for you workspace, choose at least an A10 1 GPU.
  • Make sure to select the storage unit you have created in the ‘Storage’ step.

It will take 15-30 minutes for the new workspace to be configured.

Usage

After creation is finished, there will be a new workspace that you can access on the commandline, via ssh.

After logging in, follow these steps:

  1. Verify configuration (optional)
  2. Upload model parameters
  3. Create input files
  4. Run AlphaFold
  5. Results

Verify configuration (optional)

Inside the workspace, there should be two components:

  • Docker and alphafold docker image. To verify, use this command:

    $ docker images

    You should see an image called alphafold.

  • Public (genetic) databases. To check that these are available, use this command:

    $ ls /data/<permanent_storage_name>/alphafold3/

    There should be a folder named “public_databases” with genetic database files (*.fasta) within the folder.

Upload model parameters

Note: the model parameters need to be uploaded only a single time, after which all users can utilize it. To check if they have already been uploaded, use: ls $ALPHAFOLD_STORAGE/model_parameters/. You should see a .bin file.

  1. Upload the model parameters (acquired earlier) from your local computer to the workspace.
    • You can use scp or other tools to transfer the file (probably called af3.bin.zst) to the workspace. See our data transfer manual.
  2. The uploaded file is a compressed archive containing a directory with model parameters. You now need to decompress it:
    • unzstd af3.bin.zst
  3. Place the decompressed model parameters in a directory on your permanent storage so all users on the machine can utilize it:
    • mv af3.bin $ALPHAFOLD_STORAGE/model_parameters/

Create input files

You need to create one or more input files for the AlphaFold model to work on. Input files are in the JSON format and describe the proteins that you want to predict. See an example input file, and the detailed documentation for input files.

Of course, you can also upload pre-existing input files with scp, or download them to the workspace (e.g. with git clone if they are in a public git repository, or with curl if they are on the web).

Run AlphaFold

Run using convenience script

If you placed your model parameters in $ALPHAFOLD_STORAGE/model_parameters, and your input files in $ALPHAFOLD_STORAGE/af_input/, you can simply use the following command to run alphafold:

run_alphafold

This convenience script will automatically detect the output directory and genetic databases directory on your storage, and pass them on to alphafold.

(Optional) You can also override the following locations by setting environment variables:

  • ALPHAFOLD_MODEL: path to the directory containing the model parameters
  • ALPHAFOLD_INPUT: path to the input directory, or a specific .json input file
  • ALPHAFOLD_OUTPUT: path to the output directory
  • ALPHAFOLD_DBS: path to the public genetic databases directory

For instance:

$ ALPHAFOLD_MODEL=/path/to/model/parameters/dir ALPHAFOLD_INPUT=~/input.json run_alphafold

Don’t worry if you see warning messages of the following sort:

Unable to initialize backend 'rocm': module 'jaxlib.xla_extension' has no attribute 'GpuAllocatorConfig'
Unable to initialize backend 'tpu': INTERNAL: Failed to open libtpu.so: libtpu.so: cannot open shared object file: No such file or directory

These are informational warnings, but if inference continues succesfully, you can ignore these.

Run manually

You can also invoke the alphafold3 command directly, using Docker. This allows you to customize the locatino of your input files, output files, model parameters, and genetic databases live.

For this example, we will assume that all of these are located on your permanent storage volume. You can use the shell variable $ALPHAFOLD_STORAGE as a base directory. Hence, you can use this command:

$ docker run -it \
    --volume $ALPHAFOLD_STORAGE/af_input:/root/af_input \
    --volume $ALPHAFOLD_STORAGE/af_output:/root/af_output \
    --volume $ALPHAFOLD_STORAGE/model_parameters:/root/models \
    --volume $ALPHAFOLD_STORAGE/gendb:/root/public_databases \
    --gpus all \
    alphafold3 \
    python run_alphafold.py \
    --input_dir=/root/af_input/ \
    --model_dir=/root/models \
    --output_dir=/root/af_output

Results

After running this command, your output files will appear in the following directory:

/data/af_output/<job_name>

…where is the name specified in your input file. There will be multiple output directories if you specified used multiple input files.

Source

[1] “AlphaFold”. Deepmind. Archived from the original on 19 January 2021. Retrieved 30 November 2020.

[2] https://en.wikipedia.org/wiki/AlphaFold