AlphaFold-3

How to use AlphaFold3 on the new chorus GPU nodes.

AlphaFold 3

Reference Data

Reference data is available here: /shared/biodata/alphafold3. This data volume is only available on the Chorus nodes. The reference data is located on the HPC Vast storage system. AlphaFold expects the public data to be mapped to /root/public_databases within the container.

Sbatch Example

Example sbatch script to run AlphaFold3. Save a copy of this and modify it to suit your data.

• --cpus-per-task=8 is the current max; we are working on a method to request more. If you ask for more, your job will stay in the pending (PD) state.
• --gpus=2 is a serving suggestion. Request 2 or 4 to run AlphaFold3.
• Please update the --job-name to something appropriate for your work.
• This script works from the rhino nodes, or maestro
• You need to define BASE, OUTPUT_DIR, and JSON_PATH.
• An example JSON input file can be found here.
• Note that you do not need to load an Apptainer module on the Chorus partition; Apptainer is installed there as an OS-level package.

#!/bin/sh
#SBATCH --job-name="af3"
#SBATCH --partition=chorus
#SBATCH --nodes=1
#SBATCH --cpus-per-task=8
#SBATCH --mem=1024GB
#SBATCH --gpus=2


# Do not change these paths:

# path to model parameters inside the container
export MODEL_PATH=/root/public_databases/parameter_models/
export DOWNLOAD_DIR=/shared/biodata/alphafold3
SIF=/app/software/AlphaFold/containers/alpafold3.sif



# User defined locations, please modify these:

export BASE=$HOME/alphafold3_data
export OUTPUT_DIR=$BASE/output
# example input file can be found at
# https://github.com/google-deepmind/alphafold3?tab=readme-ov-file#installation-and-running-your-first-prediction
export JSON_PATH=$BASE/example_input.json



mkdir -p $OUTPUT_DIR


apptainer exec \
  --nv \
  --bind $DOWNLOAD_DIR:/root/public_databases \
  $SIF \
  python /app/alphafold/run_alphafold.py \
  --json_path=$JSON_PATH \
  --model_dir=$MODEL_PATH \
  --output_dir=$OUTPUT_DIR \
  --db_dir=/root/public_databases

If the above script is saved as script.sh, you can run it with the command sbatch script.sh.