AlphaFold
What is AlphaFold?#
AlphaFold is an AI system developed by DeepMind that predicts a protein’s 3D structure from its amino acid sequence. It aims to achieve accuracy competitive with experimental methods.
See https://www.vscentrum.be/alphafold for more information and there you can also find a getting started video recording if you prefer that.
Documentation & extra material#
This chapter focuses specifically on the use of AlphaFold on the HPC-UGent infrastructure. It is intented to augment the existing AlphaFold documentation rather than replace it. It is therefore recommended to first familiarize yourself with AlphaFold. The following resources can be helpful:
- AlphaFold website: https://alphafold.com/
- AlphaFold repository: https://github.com/deepmind/alphafold/tree/main
- AlphaFold FAQ: https://alphafold.com/faq
- VSC webpage about AlphaFold: https://www.vscentrum.be/alphafold
- Introductory course on AlphaFold by VIB: https://elearning.vib.be/courses/alphafold
- "Getting Started with AlphaFold" presentation by Kenneth Hoste (HPC-UGent)
- recording available on YouTube
- slides available here (PDF)
- see also https://www.vscentrum.be/alphafold
Using AlphaFold on HPC-UGent infrastructure#
Several different versions of AlphaFold are installed on both the CPU and GPU HPC-UGent Tier-2 clusters, see the output of module avail AlphaFold.
If you run this command on a GPU cluster, additional CUDA modules will show up:
$ module avail AlphaFold
------------ /apps/gent/RHEL8/cascadelake-volta-ib/modules/all -------------
AlphaFold/2.0.0-fosscuda-2020b
AlphaFold/2.1.1-fosscuda-2020b
AlphaFold/2.1.2-foss-2021a-CUDA-11.3.1
AlphaFold/2.2.2-foss-2021a-CUDA-11.3.1
AlphaFold/2.3.0-foss-2021b-CUDA-11.4.1
AlphaFold/2.3.1-foss-2022a-CUDA-11.7.0
--------------- /apps/gent/RHEL8/cascadelake-ib/modules/all ----------------
AlphaFold/2.0.0-foss-2020b AlphaFold/2.3.1-foss-2022a
AlphaFold/2.1.2-foss-2021a AlphaFold/2.3.4-foss-2022a-ColabFold (D)
AlphaFold/2.2.2-foss-2021a
To use AlphaFold, you should load a particular module, for example:
module load AlphaFold/2.3.1-foss-2022a-CUDA-11.7.0
We strongly advise loading a specific version of an AlphaFold module, so you know exactly which version is being used.
Warning
When using AlphaFold, you should submit jobs to a GPU cluster for better performance, see GPU clusters. Later in this chapter, you will find a comparison between running AlphaFold on CPUs or GPUs.
Multiple revisions of the large database (~2.5TB) that is also required to run AlphaFold have been made available on the HPC-UGent infrastructure in a central location (/arcanine/scratch/gent/apps/AlphaFold), so you do not have to download it yourself.
$ ls /arcanine/scratch/gent/apps/AlphaFold
20210812 20211201 20220701 20230310
The directories located there indicate when the data was downloaded, so that this leaves room for providing updated datasets later.
As of writing this documentation the latest version is 20230310.
Info
The arcanine scratch shared filesystem is powered by fast SSD disks,
which is recommended for the AlphaFold data, because of random access I/O patterns.
See Pre-defined user directories to get more info about the arcanine filesystem.
The AlphaFold installations we provide have been modified a bit to facilitate the usage on HPC-UGent infrastructure.
Setting up the environment#
The location to the AlphaFold data can be specified via the $ALPHAFOLD_DATA_DIR environment variable, so you should define this variable in your AlphaFold job script:
export ALPHAFOLD_DATA_DIR=/arcanine/scratch/gent/apps/AlphaFold/20230310
Use newest version
Do not forget to replace 20230310 with a more up to date version if available.
Running AlphaFold#
AlphaFold provides a script called run_alphafold.py
A symbolic link named alphafold that points to this script is included,
so you can just use alphafold instead of run_alphafold.py or python run_alphafold.py after loading the AlphaFold module.
The run_alphafold.py script has also been slightly modified such that defining the $ALPHAFOLD_DATA_DIR (see above) is sufficient to pick up all the data provided in that location,
so you don't need to use options like --data_dir to specify the location of the data.
Similarly, the script was also tweaked such that the location to commands like hhblits,hhsearch,jackhmmer,kalign are already correctly set, so options like --hhblits_binary_path are not required.
For more information about the script and options see this section in the official README.
READ README
It is strongly advised to read the official README provided by DeepMind before continuing.
Controlling core count for hhblits and jackhmmer#
The Python scripts that are used to run hhblits and jackhmmer have been tweaked so you can control how many cores are used for these tools, rather than hardcoding it to 4 and 8 cores, respectively.
Using the $ALPHAFOLD_HHBLITS_N_CPU environment variable, you can specify how many cores should be used for running hhblits;
the default of 4 cores will be used if $ALPHAFOLD_HHBLITS_N_CPU is not defined.
Likewise for jackhmmer, the core count can be controlled via $ALPHAFOLD_JACKHMMER_N_CPU.
Info
Tweaking this might not yield significant benefits, as we have noticed that these tools may exhibit slower performance when utilizing more than 4/8 cores (though this behavior could vary based on the workload).
CPU/GPU comparison#
The provided timings were obtained by executing the T1050.fasta example, as outlined in the Alphafold README.
For the corresponding jobscripts, they are available here.
Using --db_preset=full_dbs, the following runtime data was collected:
- CPU-only, on doduo, using 24 cores (1 node): 9h 9min
- CPU-only, on doduo, using 96 cores (1 full node): 12h 22min
- GPU on joltik, using 1 V100 GPU + 8 cores: 2h 20min
- GPU on joltik, using 2 V100 GPUs + 16 cores: 2h 16min
This highlights a couple of important attention points:
- Running AlphaFold on GPU is significantly faster than CPU-only (close to 4x faster for this particular example).
- Using more CPU cores may lead to longer runtimes, so be careful with using full nodes when running AlphaFold CPU-only.
- Using multiple GPUs results in barely any speedup (for this particular T1050.fasta example).
With --db_preset=casp14, it is clearly more demanding:
- On doduo, with 24 cores (1 node): still running after 48h...
- On joltik, 1 V100 GPU + 8 cores: 4h 48min
This highlights the difference between CPU and GPU performance even more.
Example scenario#
The following example comes from the official Examples section in the Alphafold README. The run command is slightly different (see above: Running AlphaFold).
Do not forget to set up the environment (see above: Setting up the environment).
Folding a monomer#
Say we have a monomer with the sequence <SEQUENCE>.
Create a file monomer.fasta with the following content:
>sequence_name
<SEQUENCE>
Then run the following command in the same directory:
alphafold
--fasta_paths=monomer.fasta \
--max_template_date=2021-11-01 \
--model_preset=monomer \
--output_dir=.
See AlphaFold output, for information about the outputs.
Info
For more scenarios see the example section in the official README.
Example jobscripts#
The following two example job scripts can be used as a starting point for running AlphaFold.
The main difference between using a GPU or CPU in a job script is what module to load.
For running AlphaFold on GPU, use an AlphaFold module that mentions CUDA (or cuda),
for example AlphaFold/2.3.1-foss-2022a-CUDA-11.7.0.
To run the job scripts you need to create a file named T1050.fasta with the following content:
>T1050 A7LXT1, Bacteroides Ovatus, 779 residues|
MASQSYLFKHLEVSDGLSNNSVNTIYKDRDGFMWFGTTTGLNRYDGYTFKIYQHAENEPGSLPDNYITDIVEMPDGRFWINTARGYVLFDKERDYFITDVTGFMKNLESWGVPEQVFVDREGNTWLSVAGEGCYRYKEGGKRLFFSYTEHSLPEYGVTQMAECSDGILLIYNTGLLVCLDRATLAIKWQSDEIKKYIPGGKTIELSLFVDRDNCIWAYSLMGIWAYDCGTKSWRTDLTGIWSSRPDVIIHAVAQDIEGRIWVGKDYDGIDVLEKETGKVTSLVAHDDNGRSLPHNTIYDLYADRDGVMWVGTYKKGVSYYSESIFKFNMYEWGDITCIEQADEDRLWLGTNDHGILLWNRSTGKAEPFWRDAEGQLPNPVVSMLKSKDGKLWVGTFNGGLYCMNGSQVRSYKEGTGNALASNNVWALVEDDKGRIWIASLGGGLQCLEPLSGTFETYTSNNSALLENNVTSLCWVDDNTLFFGTASQGVGTMDMRTREIKKIQGQSDSMKLSNDAVNHVYKDSRGLVWIATREGLNVYDTRRHMFLDLFPVVEAKGNFIAAITEDQERNMWVSTSRKVIRVTVASDGKGSYLFDSRAYNSEDGLQNCDFNQRSIKTLHNGIIAIGGLYGVNIFAPDHIRYNKMLPNVMFTGLSLFDEAVKVGQSYGGRVLIEKELNDVENVEFDYKQNIFSVSFASDNYNLPEKTQYMYKLEGFNNDWLTLPVGVHNVTFTNLAPGKYVLRVKAINSDGYVGIKEATLGIVVNPPFKLAAALQHHHHHH
Job script for running AlphaFold on GPU#
Job script that runs AlphaFold on GPU using 1 V100 GPU + 8 cores.
Swap to the joltik GPU before submitting it:
module swap cluster/joltik
#!/bin/bash
#PBS -N AlphaFold-gpu-joltik
#PBS -l nodes=1:ppn=8,gpus=1
#PBS -l walltime=10:0:0
module load AlphaFold/2.3.1-foss-2022a-CUDA-11.7.0
export ALPHAFOLD_DATA_DIR=/arcanine/scratch/gent/apps/AlphaFold/20230310
WORKDIR=$VSC_SCRATCH/$PBS_JOBNAME-$PBS_JOBID
mkdir -p $WORKDIR
# download T1050.fasta via via https://www.predictioncenter.org/casp14/target.cgi?target=T1050&view=sequence
cp -a $PBS_O_WORKDIR/T1050.fasta $WORKDIR/
cd $WORKDIR
alphafold --fasta_paths=T1050.fasta --max_template_date=2020-05-14 --db_preset=full_dbs --output_dir=$PWD
echo "Output available in $WORKDIR"
Job script for running AlphaFold CPU-only#
Jobscript that runs AlphaFold on CPU using 24 cores on one node.
#!/bin/bash
#PBS -N AlphaFold-cpu-doduo
#PBS -l nodes=1:ppn=24
#PBS -l walltime=72:0:0
module load AlphaFold/2.3.1-foss-2022a
export ALPHAFOLD_DATA_DIR=/arcanine/scratch/gent/apps/AlphaFold/20230310
WORKDIR=$VSC_SCRATCH/$PBS_JOBNAME-$PBS_JOBID
mkdir -p $WORKDIR
# download T1050.fasta via via https://www.predictioncenter.org/casp14/target.cgi?target=T1050&view=sequence
cp -a $PBS_O_WORKDIR/T1050.fasta $WORKDIR/
cd $WORKDIR
alphafold --fasta_paths=T1050.fasta --max_template_date=2020-05-14 --db_preset=full_dbs --output_dir=$PWD
echo "Output available in $WORKDIR"
In case of problems or questions, don't hesitate to contact use at hpc@ugent.be.