Alphafold 3

Overview

Alphafold3 is a program developed by Google DeepMind and Isomorphic Labs to efficiently predict protein structures.

Only tagged releases are installed on the CSF as indicated by the module version number.

Restrictions on use

Alphafold3 is licensed for non commercial use only.

To use Alphafold3:
1. Read and understand the information below and then contact us to request access stating you agree to abide by the terms.

2. Obtain the model Parameters from Google DeepMind

Information on specific terms

Any publication that discloses findings arising from using Alphafold3, the model parameters or outputs produced by those should cite the “Accurate structure prediction of biomolecular interactions with AlphaFold 3” paper.

Specific licenses and terms:
Alphafold3 license
Alphafold3 model weights terms of use policy
List of Alphafold3 reference data licenses

Set up procedure

It is good practice to load module files within your jobscript so that you have a full record of how the job was run.

Module load:

module load apps/singularity/alphafold3/3.0.1

Running the application

Please do not run on the login node, jobs should be submitted via batch to a GPU partition.

This installation uses a Singularity/Apptainer container, which is launched with the command:

run_alphafold

All Alphafold3 options are provided as arguments to this command. If you have previously used Alphafold2, there are differences in how jobs are configured and run.

You can view Alphafold3 help information with:

run_alphafold --helpfull

Databases

Alphafold3 uses a set of reference databases. We have downloaded these for the specific version of Alphafold3 and the path can be found in the example sbatch script(s) below.

Batch job submission

Alphafold3 requires an input file in JSON format, an example JSON can be found on the Aphafold3 github. The docker run instructions on the github will not work on the CSF, use the batch job script format provided below.

Create a batch submission script, in this example it will be a file called alphafold3.sbatch:

#!/bin/sh
#
#SBATCH -p gpuL         ## Options: gpuA, gpuL
#SBATCH -G 1            ## Alphafold3 can only use 1 GPU
#SBATCH -n 1            ## One task
#SBATCH -c 12           ## 12 cpu cores
#SBATCH -t 4-00:00:00   ## wallclock limit Days-HH:MM:SS

module purge

module load apps/singularity/alphafold3/3.0.1


## run-specific input and output locations (scratch used as example, modify paths as needed)
INPUT=/scratch/${USER}/alphafold3-test/fold-input.json
OUTPUT=/scratch/${USER}/alphafold3-test/fold-output

## Model weights - obtained individually and saved in your own user space 
WEIGHTS=${HOME}/af3-weights/

## Databases we have pre-downloaded - you should not need to change this
DATABASES=/mnt/data-sets/alphafold3/3.0.1


## execute alphafold3
run_alphafold \
        --json_path=${INPUT} \
        --model_dir=${WEIGHTS} \
        --db_dir=${DATABASES} \
        --output_dir=${OUTPUT} \
        --run_inference=true \
        --run_data_pipeline=true

Submit the jobscript using:

sbatch alphafold3.sbatch

Further info

Github Deepmind/alphafold3

Last modified on October 17, 2025 at 3:37 pm by Martin Wolstencroft