Alphafold3

AlphaFold 3: Biomolecular Structure Prediction

Overview

AlphaFold 3 predicts the structure of mixed biomolecular systems in a single unified model:

• Proteins (including post-translational modifications)
• RNA and DNA (including modified bases)
• Small molecule ligands (via CCD codes or SMILES)
• Ions and cofactors
• Covalently modified residues

This makes AF3 the tool of choice when your system contains anything beyond a bare protein. For pure protein structure prediction, AlphaFold 2 / ColabFold remains widely used and well-benchmarked.

⚠️ License and Access

AF3 is non-commercial only. Key constraints:

• Source code: CC-BY-NC-SA 4.0 — non-commercial use only
• Model weights: separate terms of use — must apply directly from Google; commercial use requires a separate agreement
• Outputs: subject to output terms of use

For commercial use, check with Google DeepMind directly.

Access Options

Option 1: Web Server (fastest, no install)

https://alphafoldserver.com — free, up to 20 jobs/day, no installation required. Best for:

• Exploratory work
• Single predictions
• Proteins + limited ligand set

The web server uses a slightly simplified JSON format (alphafoldserver dialect) and has a more limited set of ligands and covalent modifications than the local install.

Option 2: Local Install (full capabilities)

Requires:

• Linux (Ubuntu 22.04 recommended)
• NVIDIA GPU with Compute Capability ≥ 8.0 (A100 or H100 80GB)
• ≥ 64 GB RAM
• ~1 TB disk for databases (SSD recommended)
• Model weights from Google (apply here)

# Clone and build Docker image
git clone https://github.com/google-deepmind/alphafold3.git && cd alphafold3
docker build -t alphafold3 -f docker/Dockerfile .

# Download databases (~600 GB download)
bash fetch_databases.sh /data/af3_databases

# Run prediction
docker run -it \
  --volume $HOME/af_input:/root/af_input \
  --volume $HOME/af_output:/root/af_output \
  --volume /data/af3_models:/root/models \
  --volume /data/af3_databases:/root/public_databases \
  --gpus all \
  alphafold3 \
  python run_alphafold.py \
    --json_path=/root/af_input/input.json \
    --model_dir=/root/models \
    --output_dir=/root/af_output

Two-stage pipeline:

• --run_data_pipeline=true (default) — MSA and template search, CPU-only, run separately on a compute node
• --run_inference=true (default) — structure prediction, requires GPU

Input JSON Format

Unlike AF2 (FASTA input), AF3 uses a structured JSON file specifying every molecular entity. Each entity gets a chain ID.

Minimal example: single protein

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

Protein–ligand complex

{
  "name": "kinase_atp",
  "modelSeeds": [1],
  "sequences": [
    {
      "protein": {
        "id": "A",
        "sequence": "MKTAYIAKQRQISFVKSHFSRQLEERLGLIEVQAPILSRVGD"
      }
    },
    {
      "ligand": {
        "id": "B",
        "ccdCodes": ["ATP"]
      }
    },
    {
      "ligand": {
        "id": "C",
        "ccdCodes": ["MG"]
      }
    }
  ],
  "dialect": "alphafold3",
  "version": 1
}

Protein–DNA complex

{
  "name": "transcription_factor",
  "modelSeeds": [1],
  "sequences": [
    {
      "protein": {
        "id": "A",
        "sequence": "MKTAYIAKQRQISFVK"
      }
    },
    {
      "dna": {
        "id": ["B", "C"],
        "sequence": "GACCTCTGAGGT"
      }
    }
  ],
  "dialect": "alphafold3",
  "version": 1
}

Custom ligand via SMILES

{
  "ligand": {
    "id": "B",
    "smiles": "CC(=O)Nc1ccc(O)cc1"
  }
}

Note: backslashes in SMILES must be escaped as \\ in JSON.

Homomeric complex (multiple copies of the same chain)

{
  "protein": {
    "id": ["A", "B", "C"],
    "sequence": "MKTAYIAKQRQISFVK"
  }
}

Post-translational modifications

{
  "protein": {
    "id": "A",
    "sequence": "PVLSCGEWQL",
    "modifications": [
      {"ptmType": "SEP", "ptmPosition": 3},
      {"ptmType": "TPO", "ptmPosition": 7}
    ]
  }
}

See references/input-format.md for the full entity type reference, covalent bonds, custom MSA, and template specification.

Building Input JSON Programmatically

python scripts/build_input.py \
  --name my_complex \
  --protein "MKTAYIAKQRQISFVK" \
  --ligand-ccd ATP MG \
  --seeds 1 2 3 \
  --output input.json

Output Format

AF3 outputs mmCIF files (not PDB) plus JSON confidence files.

/
├── _model.cif                    ← best prediction (top ranking_score)
├── _confidences.json             ← full confidence arrays for best model
├── _summary_confidences.json     ← summary metrics for best model
├── _ranking_scores.csv           ← all predictions ranked
├── seed-1_sample-0/                        ← all individual predictions
│   ├── *_model.cif
│   ├── *_confidences.json
│   └── *_summary_confidences.json
└── seed-1_sample-1/  ...

Key confidence metrics

| Metric | Type | Range | Interpretation | |---|---|---|---| | atom_plddts | per-atom array | 0–100 | Local confidence; >70 = reliable | | pae | N×N matrix | 0–∞ Å | Relative position error; low = confident | | ptm | scalar | 0–1 | Global fold confidence; >0.5 = plausible | | iptm | scalar | 0–1 | Interface confidence; >0.8 = high, <0.6 = failed | | chain_pair_pae_min | M×M matrix | 0–∞ Å | Inter-chain interaction confidence | | contact_probs | N×N matrix | 0–1 | Probability of contact (<8 Å) | | ranking_score | scalar | | 0.8×iptm + 0.2×ptm + 0.5×disorder − 100×has_clash |

Note: AF3 pLDDT is per-atom (not per-residue like AF2). For proteins, mean per-residue pLDDT can be derived by averaging atoms in each residue.

See references/outputs.md for parsing mmCIF and confidence JSON in Python.

AF3 vs AF2 Comparison

| | AlphaFold 2 | AlphaFold 3 | |---|---|---| | Input format | FASTA | JSON | | Output format | PDB | mmCIF | | Proteins | ✓ | ✓ | | Protein complexes | ✓ (Multimer) | ✓ | | RNA/DNA | ✗ | ✓ | | Small molecule ligands | ✗ | ✓ | | PTMs / modified bases | ✗ | ✓ | | License | Apache 2.0 | CC-BY-NC-SA 4.0 | | Model weights | CC BY 4.0 | Non-commercial only | | Web server | AFDB (precomputed) | alphafoldserver.com |

Resources

• Web server: https://alphafoldserver.com
• GitHub: https://github.com/google-deepmind/alphafold3
• Model weights request: https://forms.gle/svvpY4u2jsHEwWYS6
• Paper: Abramson et al., Nature 2024 — https://doi.org/10.1038/s41586-024-07487-w

References

• references/input-format.md — full JSON schema: RNA, DNA, ligands, PTMs, covalent bonds, custom MSA, templates
• references/outputs.md — parsing mmCIF structures, confidence JSON, embeddings, contact probabilities

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: naity
• Source: naity/FM4Life
• License: MIT

Install and usage instructions live in the source repository linked above.