Skip to main content

​
General Questions

AlphaFold 3 is a deep learning system for predicting the 3D structures of biomolecular complexes, including proteins, nucleic acids, ligands, and post-translational modifications. It represents a significant advancement over AlphaFold 2, extending structure prediction beyond proteins to diverse biomolecular interactions.The model was published in Nature: Accurate structure prediction of biomolecular interactions with AlphaFold 3.
Key differences include:
  • Broader scope: AlphaFold 3 can model proteins, RNA, DNA, ligands, ions, and post-translational modifications
  • Complex interactions: Predicts interactions between different biomolecule types
  • Improved accuracy: Enhanced performance on protein-ligand and protein-nucleic acid interactions
  • Covalent modifications: Supports modeling of covalent bonds and modifications
  • New architecture: Uses an updated neural network architecture optimized for diverse biomolecules
No, AlphaFold 3 is licensed for non-commercial use only under the CC-BY-NC-SA 4.0 license.
The AlphaFold 3 source code and model parameters are subject to non-commercial licensing terms. Review the terms of use before use.
For commercial applications, please contact the AlphaFold team at [email protected].
Yes, AlphaFold 3 is available at alphafoldserver.com for non-commercial use.However, the server has limitations:
  • More limited set of ligands
  • Fewer covalent modifications supported
  • No ability to customize model parameters
The local installation provides full functionality and flexibility.

​
Model Access and Installation

To request access to the model parameters:
  1. Complete the access request form
  2. Wait for approval from Google DeepMind (typically 2-3 business days)
  3. You will receive a download link via email
You may only use AlphaFold 3 model parameters if received directly from Google. Access is granted at Google DeepMind’s sole discretion.
Minimum requirements:
  • GPU: NVIDIA GPU with at least 16GB VRAM (e.g., V100, A100, RTX 4090)
  • RAM: 32GB+ system RAM
  • Disk: Several TB for databases (500GB+ for minimal setup)
  • CPU: Multi-core processor for data pipeline
Recommended:
  • GPU: A100 80GB or H100 for large complexes
  • RAM: 64GB+ system RAM
  • Disk: 3-5TB SSD for full databases
The data pipeline (genetic searches) can run on CPU-only machines without a GPU.
While Docker is the officially supported method, technically you can run AlphaFold 3 without Docker by:
  1. Manually installing all dependencies
  2. Setting up the Python environment
  3. Configuring paths for databases and models
However, this is not recommended because:
  • Complex dependency management
  • Potential version conflicts
  • Lack of official support
Docker ensures a consistent, reproducible environment.
Installation time varies:
  • Docker build: 30-60 minutes (depending on network speed)
  • Database download: Several hours to days (depending on which databases and network speed)
  • Model parameters: 15-30 minutes (depending on network speed)
Total setup time: Typically 1-3 days for a complete installation.

​
Running Predictions

Prediction time depends on several factors:Data pipeline (genetic searches):
  • Simple proteins: 1-4 hours
  • Complex proteins: 4-24 hours
  • Very large complexes: 24-48+ hours
Inference (GPU):
  • Small proteins (<200 residues): 5-15 minutes
  • Medium proteins (200-500 residues): 15-45 minutes
  • Large complexes (>500 residues): 45 minutes - 2 hours
You can run data pipeline and inference separately using --run_data_pipeline and --run_inference flags.
Yes, with caveats:
  • Data pipeline: Can run multiple in parallel if you have enough CPU and RAM
  • Inference: Limited by GPU memory - typically 1-2 predictions per GPU depending on complex size
For maximum efficiency:
  1. Run data pipeline on CPU-only machines in parallel
  2. Run inference sequentially or on multiple GPUs
A model seed controls the random initialization of the neural network. Using multiple seeds generates diverse predictions.Recommendations:
  • Standard use: 5 seeds ("modelSeeds": [1, 2, 3, 4, 5])
  • Quick testing: 1 seed
  • High confidence needed: 10+ seeds
The model ranks all predictions, with the best one typically selected.
Data pipeline: Cannot be easily resumed - must restart if interruptedInference: Cannot be resumed - must rerun if interrupted
Save your data pipeline outputs to avoid rerunning expensive genetic searches.
Best practices:
  • Run data pipeline separately with --run_data_pipeline=true --run_inference=false
  • Save MSA and template outputs
  • Run inference later with pre-computed data

​
Input and Output

AlphaFold 3 supports:
  • Proteins: Any amino acid sequence
  • RNA: Nucleotide sequences
  • DNA: Nucleotide sequences
  • Ligands: Small molecules (via SMILES or CCD codes)
  • Ions: Common ions (via CCD codes)
  • Post-translational modifications: Glycans, phosphorylation, etc.
  • Covalent bonds: Disulfide bridges, peptide bonds to ligands, etc.
See the Input Format documentation for details.
AlphaFold 3 provides several confidence metrics:pLDDT (per-residue confidence):
  • 90: Very high confidence
  • 70-90: Generally reliable
  • 50-70: Low confidence
  • <50: Should not be interpreted
Ranking score:
  • Higher scores indicate better overall model quality
  • Compares different model seeds
  • Score of -99 or lower indicates severe problems
Always inspect structures visually and validate predictions experimentally when possible.
For each prediction:
  • CIF files: 3D structure coordinates (mmCIF format)
  • JSON files: Confidence scores and metadata
  • Ranking file: Comparison of different model seeds
  • Data pipeline outputs: MSAs, templates (if run)
See the Output Format documentation for details.
AlphaFold 3 is primarily a structure prediction tool, not a design tool. However:Possible uses:
  • Validate designed sequences by predicting their structures
  • Screen multiple design candidates
  • Predict effects of mutations
Not designed for:
  • Inverse folding (sequence design from structure)
  • Optimizing sequences for specific functions
Consider tools specifically designed for protein design (e.g., ProteinMPNN, RFdiffusion).

​
Ligands and Small Molecules

Ligands can be specified in two ways:1. Using SMILES:
{
  "ligand": {
    "id": ["L"],
    "smiles": "CC(C)CC1=CC=C(C=C1)C(C)C(=O)O"
  }
}
2. Using CCD codes:
{
  "ligand": {
    "id": ["L"],
    "ccdCodes": ["ATP"]
  }
}
See the Input Format documentation for complete examples.
For non-standard ligands:
  1. Use SMILES notation: Define the molecule with a SMILES string
  2. Include stereochemistry: Ensure stereochemistry is explicit in SMILES
  3. Specify covalent bonds: Use bondedAtomPairs if needed
See the Community Tools page for tools like JAAG that help generate ligand inputs.
Glycan modeling requires:
  1. Correct CCD codes: Use SNFG nomenclature
  2. Bonded atom pairs: Specify glycosidic bonds
  3. Proper JSON structure: Follow the glycan format
See the paper Modeling glycans with AlphaFold 3 for:
  • Step-by-step tutorials
  • Ready-to-run scripts
  • Comprehensive CCD tables
  • Discussion of limitations
Tools like JAAG automate glycan input generation.

​
Troubleshooting

Common causes:
  1. Input errors: Check sequences and ligand definitions
  2. Hardware issues: See Known Issues for GPU-specific problems
  3. Inherent uncertainty: Some targets are genuinely difficult to predict
  4. Insufficient data: Limited homologs in genetic databases
  5. Complex interactions: Novel or unusual binding modes
See the Troubleshooting page for detailed solutions.
Support resources:
  1. Documentation: Check the official documentation
  2. Known Issues: Review the Known Issues page
  3. GitHub: Search the issues tracker
  4. Community: See Community Tools for community resources
  5. Email: Contact [email protected]
When reporting issues, include:
  • Hardware configuration
  • Error messages
  • Input JSON
  • Steps to reproduce

​
Citation and Publication

See the Citation page for the complete citation information.
Any publication using AlphaFold 3 source code, model parameters, or outputs must cite the AlphaFold 3 paper.
Yes, you can share predictions, but:
  • Must cite the AlphaFold 3 paper
  • Must acknowledge the non-commercial license
  • Cannot use for commercial purposes
  • Must include appropriate disclaimers about prediction confidence
Predictions are for theoretical modeling only and not validated for clinical use.

Build docs developers (and LLMs) love

Get started for freeTalk to us