AlphaFold3¶

AlphaFold3 predicts the 3D structure of biomolecular complexes including proteins, DNA, RNA, ligands, and ions. It runs on Hyperion as an optimized Apptainer container.

Loading the module¶

module load AlphaFold/3.0.1

This sets up the container path (AF3_CONTAINER), model weights (AF3_WEIGHTS), reference databases (AF3_DB), and wrapper scripts.

To verify the installation before your first run:

alphafold3-check

Batch job examples¶

#!/bin/bash
#SBATCH --qos=regular
#SBATCH --job-name=alphafold3
#SBATCH --partition=general
#SBATCH --nodes=1
#SBATCH --ntasks=1
#SBATCH --cpus-per-task=48
#SBATCH --mem=100G
#SBATCH --gpus=1
#SBATCH --time=12:00:00
#SBATCH --output=%x.%j.out
#SBATCH --error=%x.%j.err

module load AlphaFold/3.0.1

alphafold3-run input.json results_${SLURM_JOB_ID}/

#!/bin/bash
#SBATCH --qos=regular
#SBATCH --job-name=af3_large
#SBATCH --partition=general
#SBATCH --nodes=1
#SBATCH --ntasks=1
#SBATCH --cpus-per-task=48
#SBATCH --mem=200G
#SBATCH --gpus=1
#SBATCH --constraint=a100
#SBATCH --time=24:00:00
#SBATCH --output=%x.%j.out
#SBATCH --error=%x.%j.err

module load AlphaFold/3.0.1

alphafold3-run input.json results_${SLURM_JOB_ID}/

Input format¶

AlphaFold3 takes a JSON file defining the molecular system. A minimal example for a protein:

{
  "name": "my_prediction",
  "sequences": [
    {
      "protein": {
        "id": ["A"],
        "sequence": "MAFSAEDVLK..."
      }
    }
  ],
  "modelSeeds": [1, 2, 3],
  "dialect": "alphafold3",
  "version": 1
}

Supported entity types in the sequences array:

Use "id": ["A", "B"] to create homodimers (multiple copies of the same chain). Each modelSeeds entry generates an independent structure prediction.

For complex inputs (post-translational modifications, covalent bonds, custom MSAs), see Advanced usage.

GPU and memory selection¶

GPU memory scales roughly quadratically with system size. Count your tokens to choose the right resources:

• Proteins: 1 token per residue
• DNA/RNA: 1 token per nucleotide
• Ligands: 1 token per heavy atom
• Ions: 1 token each

Output files¶

AlphaFold3 produces one structure per seed in mmCIF format, plus confidence metrics:

output_dir/
├── job_name_model_0.cif              # 3D structure (seed 1)
├── job_name_model_1.cif              # 3D structure (seed 2)
├── job_name_confidences.json         # pLDDT scores and PAE matrices
├── job_name_summary_confidences.json # Prediction summary
└── job_name_data.json                # Full prediction data

Open .cif files in PyMOL, ChimeraX, or VMD. Evaluate prediction quality using pLDDT scores: above 90 indicates high confidence, 70-90 is reliable for the backbone, below 50 suggests disorder.

Troubleshooting¶

Out of memory (GPU): Switch to A100 with --constraint=a100, or reduce modelSeeds to fewer entries. GPU memory scales quadratically with token count.

Process killed by SLURM (OOM-kill): This is system RAM, not GPU memory. Increase --mem (try 200G for large systems).

No CUDA GPUs available: Verify --gpus=1 is in your SLURM script and that the module is loaded.

Invalid JSON format: Check that "dialect": "alphafold3" and "version": 1 are present. Validate syntax with python -m json.tool input.json.

See also¶