Alphafold

Alphafold 3 inference pipelines can be run on ALICE. Use of the model parameters is subject to these terms of use. Contact the ALICE helpdesk if you would like to use it.

With the right module loaded AF3 can be run on several type of GPU’s.
Check the available partitions GPU’s on ALICE:
Partitions on ALICE

For compute capability 7 use:

At the initialisation fase when AF reads the database most memory is used. AF recommends the use of A100 GPU’s but in our experience a smaller GPU with Compute Capability 7 or a partial GPU (MIG) is sufficient. You can load the appropiate module with the container and the compute environment in your job:

module load alphafold/cc7_3-20250304

Or on Compute Capability 8 (A100, mig,L4) use:

module load alphafold/cc8_3-20250304

Then AF3 is also very I/O intensive, while reading many small files from the database in /zfsstore/databases/AlphaFold3_resources/.

Our storage system can handle this efficiently. If this fase takes a Alternatively the database is compressed in a Zstandard (zstd) compressed SquashFS file system image. Using this image instead of the files directly could be faster. Mount them in your scratch folder:

mkdir $TMPDIR/af_data
squashfuse /zfsstore/databases/AlphaFold3_resources/databases/pdb_2022_09_28_mmcif_files.zstd.sqsh $TMPDIR/af_data/

The image can be unmounted with:

fusermount -u $TMPDIR/af_data

Example of a jobscript (not tested):

#!/bin/bash
#SBATCH --job-name=alphafold
#SBATCH --time=04:00:00
#SBATCH --partition=gpu-short
#SBATCH --nodes=1
#SBATCH --cpus-per-task=8
#SBATCH --mem=120GB
#SBATCH --gres=gpu:1
# note: check ALICE partitions for available gpu's

# load the module
module load alphafold/cc8_3-20250304

# Uncomment the following line(s) if you want to use different values for the number of cores used by hhblits/jackhmmer
#export ALPHAFOLD_HHBLITS_N_CPU=8 # default: 4
#export ALPHAFOLD_JACKHMMER_N_CPU=4 # default: 8

# Run AlphaFold
alphafold --fasta_paths=query.fasta --max_template_date=2020-05-14 --output_dir=output