Alphafold compute setup
=======================

Overview
--------

Alphafold requires a customised compute environment to run. The machine
needs a GPU, and access to a 2.2 Tb reference data store.

This document is designed to provide details on the compute environment
required for Alphafold operation, and the Galaxy job destination
settings to run the wrapper. We strongly suggest reading this entire document
to ensure that your setup is compatible with the hardware that you are
deploying to.

For full details on Alphafold requirements, see
https://github.com/deepmind/alphafold.

HARDWARE
~~~~~~~~

The machine is recommended to have the following specs: - 12 cores - 80
Gb RAM - 2.5 Tb storage - A fast Nvidia GPU.

As a minimum, the Nvidia GPU must have 8Gb RAM. It also requires
**unified memory** to be switched on. Unified memory is usually enabled
by default, but some HPC systems will turn it off so the GPU can be
shared between multiple jobs concurrently.

ENVIRONMENT
~~~~~~~~~~~

This wrapper runs Alphafold as a singularity container. The following
software are needed:

-  `Singularity <https://sylabs.io/guides/3.0/user-guide/installation.html>`_
-  `NVIDIA Container
   Toolkit <https://docs.nvidia.com/datacenter/cloud-native/container-toolkit/install-guide.html>`_

As Alphafold uses an Nvidia GPU, the NVIDIA Container Toolkit is needed.
This makes the GPU available inside the running singularity container.

To check that everything has been set up correctly, run the following

::

   singularity run --nv docker://nvidia/cuda:11.0-base nvidia-smi

If you can see something similar to this output (details depend on your
GPU), it has been set up correctly.

::

   +-----------------------------------------------------------------------------+
   | NVIDIA-SMI 470.57.02    Driver Version: 470.57.02    CUDA Version: 11.4     |
   |-------------------------------+----------------------+----------------------+
   | GPU  Name        Persistence-M| Bus-Id        Disp.A | Volatile Uncorr. ECC |
   | Fan  Temp  Perf  Pwr:Usage/Cap|         Memory-Usage | GPU-Util  Compute M. |
   |                               |                      |               MIG M. |
   |===============================+======================+======================|
   |   0  Tesla T4            Off  | 00000000:00:05.0 Off |                    0 |
   | N/A   49C    P0    28W /  70W |      0MiB / 15109MiB |      0%      Default |
   |                               |                      |                  N/A |
   +-------------------------------+----------------------+----------------------+

   +-----------------------------------------------------------------------------+
   | Processes:                                                                  |
   |  GPU   GI   CI        PID   Type   Process name                  GPU Memory |
   |        ID   ID                                                   Usage      |
   |=============================================================================|
   |  No running processes found                                                 |
   +-----------------------------------------------------------------------------+

REFERENCE DATA
~~~~~~~~~~~~~~

Alphafold needs reference data to run. The wrapper expects this data to
be present at ``/data/alphafold_databases``. A custom path will be read from
the ``ALPHAFOLD_DB`` environment variable, if set.

To download the AlphaFold reference DBs:

::

   # Set your AlphaFold DB path
   ALPHAFOLD_DB=/data/alphafold_databases

   # Set your target AlphaFold version
   ALPHAFOLD_VERSION=  # e.g. 2.1.2

   # Download repo
   wget https://github.com/deepmind/alphafold/releases/tag/v${ALPHAFOLD_VERSION}.tar.gz
   tar xzf v${ALPHAFOLD_VERSION}.tar.gz

   # Ensure dirs
   mkdir -p $ALPHAFOLD_DB

   # Download
   bash alphafold*/scripts/download_all_data.sh $ALPHAFOLD_DB

You will most likely want to run this as a background job, as it will take a
very long time (7+ days in Australia).

This will install the reference data to your ``$ALPHAFOLD_DB``.
To check this has worked, ensure the final folder structure is as
follows:

::

   # NOTE: this structure will change between minor AlphaFold versions
   # The tree shown below was updated for v2.3.1

   data/alphafold_databases
   ├── bfd
   │   ├── bfd_metaclust_clu_complete_id30_c90_final_seq.sorted_opt_a3m.ffdata
   │   ├── bfd_metaclust_clu_complete_id30_c90_final_seq.sorted_opt_a3m.ffindex
   │   ├── bfd_metaclust_clu_complete_id30_c90_final_seq.sorted_opt_cs219.ffdata
   │   ├── bfd_metaclust_clu_complete_id30_c90_final_seq.sorted_opt_cs219.ffindex
   │   ├── bfd_metaclust_clu_complete_id30_c90_final_seq.sorted_opt_hhm.ffdata
   │   └── bfd_metaclust_clu_complete_id30_c90_final_seq.sorted_opt_hhm.ffindex
   ├── mgnify
   │   └── mgy_clusters_2022_05.fa
   ├── params
   │   ├── LICENSE
   │   ├── params_model_1.npz
   │   ├── params_model_1_multimer_v3.npz
   │   ├── params_model_1_ptm.npz
   │   ├── params_model_2.npz
   │   ├── params_model_2_multimer_v3.npz
   │   ├── params_model_2_ptm.npz
   │   ├── params_model_3.npz
   │   ├── params_model_3_multimer_v3.npz
   │   ├── params_model_3_ptm.npz
   │   ├── params_model_4.npz
   │   ├── params_model_4_multimer_v3.npz
   │   ├── params_model_4_ptm.npz
   │   ├── params_model_5.npz
   │   ├── params_model_5_multimer_v3.npz
   │   └── params_model_5_ptm.npz
   ├── pdb70
   │   ├── md5sum
   │   ├── pdb70_a3m.ffdata
   │   ├── pdb70_a3m.ffindex
   │   ├── pdb70_clu.tsv
   │   ├── pdb70_cs219.ffdata
   │   ├── pdb70_cs219.ffindex
   │   ├── pdb70_hhm.ffdata
   │   ├── pdb70_hhm.ffindex
   │   └── pdb_filter.dat
   ├── pdb_mmcif
   │   ├── mmcif_files
   │   └── obsolete.dat
   ├── pdb_seqres
   │   └── pdb_seqres.txt
   ├── uniprot
   │   └── uniprot.fasta
   ├── uniref30
   │   ├── UniRef30_2021_03.md5sums
   │   ├── UniRef30_2021_03_a3m.ffdata
   │   ├── UniRef30_2021_03_a3m.ffindex
   │   ├── UniRef30_2021_03_cs219.ffdata
   │   ├── UniRef30_2021_03_cs219.ffindex
   │   ├── UniRef30_2021_03_hhm.ffdata
   │   └── UniRef30_2021_03_hhm.ffindex
   └── uniref90
      └── uniref90.fasta

In more recent releases of the AlphaFold tool, you will need to download an
additional file to allow the ``reduced_dbs`` option:

::

   bash scripts/download_small_bfd.sh $ALPHAFOLD_DB_ROOT

The ``$ALPHAFOLD_DB_ROOT`` directory should now contain this additional file:

::

   data/alphafold_databases
   ├── small_bfd
   │   └── bfd-first_non_consensus_sequences.fasta


**Upgrading database versions**

When upgrading to a new minor version of AlphaFold, you will most likely have to
upgrade the reference database. This can be a pain, due to the size of the
databases and the obscurity around what has changed. The simplest way to do
this is simply create a new directory and download the DBs from scratch.
However, you can save a considerable amount of time by downloading only the
components that have changed.

If you wish to continue hosting prior versions of the tool, you must maintain
the reference DBs for each version. The ``ALPHAFOLD_DB`` environment variable
must then be set respectively for each tool version in your job conf (on Galaxy
AU this is currently `configured with TPV<https://github.com/usegalaxy-au/infrastructure/blob/master/files/galaxy/dynamic_job_rules/production/total_perspective_vortex/tools.yml#L1515-L1554>`_).

To minimize redundancy between DB version, we have symlinked the database
components that are unchanging between versions. In ``v2.1.2 -> v2.3.1`` the BFD
database is the only component that is persistent, but they are by far the
largest on disk.


JOB DESTINATION
~~~~~~~~~~~~~~~

Alphafold needs a custom singularity job destination to run. The
destination needs to be configured for singularity, and some extra
singularity params need to be set as seen below.

Specify the job runner. For example, a local runner

::

   <plugin id="alphafold_runner" type="runner" load="galaxy.jobs.runners.local:LocalJobRunner"/>

Customise the job destination with required singularity settings. The
settings below are mandatory, but you may include other settings as
needed.

::

   <destination id="alphafold" runner="alphafold_runner">
       <param id="dependency_resolution">'none'</param>
       <param id="singularity_enabled">true</param>
       <param id="singularity_run_extra_arguments">--nv</param>
       <param id="singularity_volumes">"$job_directory:ro,$tool_directory:ro,$job_directory/outputs:rw,$working_directory:rw,/data/alphafold_databases:/data:ro"</param>
   </destination>

CUSTOM PARAMETERS
~~~~~~~~~~~~~~~~~

A few parameters can be customized with the use of environment variables set in the job destination:

- ``ALPHAFOLD_DB``: path to the reference database root (default ``/data``)
- ``ALPHAFOLD_USE_GPU [True/False]``: set to ``False`` to disable GPU dependency (defaults to ``True``)
- ``ALPHAFOLD_AA_LENGTH_MIN``: minimum accepted sequence length (default ``0``)
- ``ALPHAFOLD_AA_LENGTH_MAX``: maximum accepted sequence length (default ``0`` - no validation)

Closing
~~~~~~~

If you are experiencing technical issues, feel free to write to
help@genome.edu.au. We may be able to provide advice on setting up
Alphafold on your compute environment.