AlphaFold3

AlphaFold3 is available on Rockfish as a two-stage workflow that combines CPU-based preprocessing and GPU-based inference. This guide will walk you through accessing the module, running jobs, and preparing batch or interactive sessions.

Accessing AlphaFold3

To access AlphaFold3, load the module:

module load alphafold/3

After loading, the wrapper script run_alphafold3.sh becomes available. This script runs AlphaFold3 (run_alphafold.py) in two stages:

CPU Stage – Data preparation (auto-submitted to a CPU node)
GPU Stage – Model inference (runs on a GPU node)

Note

Recommended GPU Partition: Use ica100 for best compatibility with AlphaFold3’s requirements.

Basic Usage

run_alphafold3.sh [OPTIONS] [ALPHAFOLD_EXTRA_OPTS]

Options

Required Parameters

--input_dir=: Path to directory containing input file(s)
--output_dir=: Path to directory where results will be saved (must already exist)
--models_dir=: Path to AF3 models → /scratch4/datasets/alphafold3_models
--db_dir=: Path to AF3 databases → /scratch4/datasets/alphafold3
--input_file=: Name of a JSON input file (for single inputs only)

Optional Parameters

--cpu_partition=: Default is parallel; use bigmem for large jobs or shared for small jobs
--cpus=: Number of CPU cores (default: 24)
--account=: Slurm account (required for bigmem partition)
--qos=: QoS flag (required for bigmem partition)

Note

Additional options are forwarded directly to run_alphafold.py.

Examples

Running with a Single Input (JSON)

run_alphafold3.sh \
  --input_dir=/path/to/input \
  --output_dir=/path/to/output \
  --models_dir=/scratch4/datasets/alphafold3_models \
  --db_dir=/scratch4/datasets/alphafold3 \
  --input_file=my_input.json \
  --cpu_partition=parallel \
  --cpus=16

Note

For single input jobs: - Use --input_dir with the full path to the directory. - Use --input_file with the JSON file name.

Running with Multiple Inputs

run_alphafold3.sh \
  --input_dir=/path/to/input \
  --output_dir=/path/to/output \
  --models_dir=/scratch4/datasets/alphafold3_models \
  --db_dir=/scratch4/datasets/alphafold3 \
  --cpu_partition=bigmem \
  --account=myPI_bigmem \
  --qos=qos_bigmem

Note

For multiple input files, --input_file is not needed — use --input_dir with a directory of JSON files.

Data Pipeline Only (CPU Stage Only)

run_alphafold3.sh \
  --input_dir=/path/to/input \
  --output_dir=/path/to/output \
  --models_dir=/scratch4/datasets/alphafold3_models \
  --db_dir=/scratch4/datasets/alphafold3 \
  --norun_inference

Note

This will generate MSAs and templates only, skipping model inference. Output files can be used in later inference runs.

Model Inference Only (GPU Stage Only)

run_alphafold3.sh \
  --input_dir=/path/to/data_jsons \
  --output_dir=/path/to/output \
  --models_dir=/scratch4/datasets/alphafold3_models \
  --db_dir=/scratch4/datasets/alphafold3 \
  --norun_data_pipeline

Note

This mode skips the CPU stage. JSON files must include precomputed MSAs and templates.

SLURM Batch Script Example

Recommended for production workloads:

#!/bin/bash
#SBATCH --job-name=my_job_af3
#SBATCH --partition=ica100
#SBATCH --account=MyAccount_gpu
#SBATCH --qos=qos_gpu
#SBATCH --time=48:00:00
#SBATCH --nodes=1
#SBATCH --gres=gpu:1

module load alphafold/3

run_alphafold3.sh \
  --input_dir=/path/to/input \
  --output_dir=/path/to/output \
  --models_dir=/scratch4/datasets/alphafold3_models \
  --db_dir=/scratch4/datasets/alphafold3 \
  --cpu_partition=shared \
  --cpus=16

Note

Submit to a GPU partition unless you use --norun_inference, which runs only the CPU portion.

Running Interactively

To launch AlphaFold3 interactively on a GPU node:

interact -p ica100 -q qos_gpu -a MyAccount_gpu -t 02:00:00 -g 1

Once on the GPU node:

module load alphafold/3

run_alphafold3.sh \
  --input_dir=/path/to/input \
  --output_dir=/path/to/output \
  --models_dir=/scratch4/datasets/alphafold3_models \
  --db_dir=/scratch4/datasets/alphafold3 \
  --cpu_partition=shared \
  --cpus=12

Tip

Make sure to monitor your job’s resource usage and ensure it runs on a GPU node when required.