Foldseek

Foldseek enables fast and sensitive comparisons of large sets of protein structures. It describes structures with a structural alphabet (3Di) and searches and clusters them with MMseqs2-based algorithms. On Roihu it is provided as a GPU-accelerated build for the GH200 GPU nodes.

Foldseek

License

Free to use and open source under GNU GPLv3.

Available

Roihu-GPU: 10-941cd33, via the bio-apps module.

Foldseek runs on the Roihu GPU (GH200) nodes. Log in to roihu-gpu.csc.fi and load the modules there so you get the GPU build.

Usage

Foldseek is part of the bio-apps collection on Roihu. Load the bio-apps module tree and then the Foldseek module:

module load bio-apps/v202603
module load foldseek/10-941cd33

Search a set of query structures (PDB or mmCIF files) against a set of target structures:

foldseek easy-search queries/ targets/ results.m8 tmp

Cluster a set of structures:

foldseek easy-cluster structures/ clusters tmp -c 0.9

Databases

CSC provides shared, prebuilt Foldseek databases on Roihu, and the foldseek module sets the environment variable $FOLDSEEK_DB to their location (under /dataset/project_2020345/foldseek). Give the path to a database as the search target:

Database	Source (`foldseek databases` name)	Padded copy for GPU search
`afdb`	Alphafold/UniProt	`afdb_pad`
`afdb50`	Alphafold/UniProt50	`afdb50_pad`
`afdb_proteome`	Alphafold/Proteome	`afdb_proteome_pad`
`afdb_swissprot`	Alphafold/Swiss-Prot	`afdb_swissprot_pad`
`esmatlas30`	ESMAtlas30	`esmatlas30_pad`
`pdb`	PDB	`pdb_pad`
`cath50`	CATH50	`cath50_pad`
`bfmd`	BFMD	`bfmd_pad`
`bfvd`	BFVD	`bfvd_pad`

foldseek easy-search queries/ $FOLDSEEK_DB/pdb results.m8 tmp

The ProstT5 model is also provided, so you can search with amino acid sequences (FASTA) instead of structures:

foldseek easy-search queries.fasta $FOLDSEEK_DB/pdb results.m8 tmp --prostt5-model $FOLDSEEK_DB/prostt5

To use another database, foldseek databases downloads prebuilt databases. Run it without arguments to list all available databases. Download large databases to your project's /scratch:

foldseek databases <name> /scratch/<project>/<database> tmp

foldseek databases only downloads and unpacks prebuilt files, so you can run it on the login node.

GPU-accelerated search

Add --gpu 1 to run a search on the GPU. A GPU search needs a padded target database. Use the padded copies of the shared databases (names ending in _pad):

foldseek easy-search queries/ $FOLDSEEK_DB/afdb_swissprot_pad results.m8 tmp --gpu 1 --prefilter-mode 1

GPU searches must run on a GPU node. Use a batch job on a GPU partition (see the example batch script) for searches against the large databases. For short tests, you can also start an interactive session with sinteractive on roihu-gpu.csc.fi, but interactive sessions will get a GPU slice with only 12 GiB of GPU memory once slicing is configured, so use a small database such as afdb_swissprot_pad there.

For your own databases, make a padded copy with makepaddedseqdb:

foldseek makepaddedseqdb mydb mydb_pad

Example batch script

#!/bin/bash
#SBATCH --job-name=foldseek
#SBATCH --account=<project>
#SBATCH --output=output_%j.txt
#SBATCH --error=errors_%j.txt
#SBATCH --partition=gpumedium
#SBATCH --time=04:00:00
#SBATCH --nodes=1
#SBATCH --ntasks=1
#SBATCH --cpus-per-task=72
#SBATCH --gres=gpu:gh200:1

module load bio-apps/v202603
module load foldseek/10-941cd33

foldseek easy-search queries/ $FOLDSEEK_DB/afdb_swissprot_pad results.m8 tmp --gpu 1 --prefilter-mode 1 \
  --threads $SLURM_CPUS_PER_TASK

Replace <project> with your CSC project (for example project_2001234). See creating a batch job script for Roihu for more about batch jobs, and the GPU partitions for the available GPU resources.

Support

CSC Service Desk