AlphaFold 3 for Mac

Run AlphaFold 3 protein structure prediction natively on Apple Silicon Macs. The model inference layer is rewritten in Apple's MLX framework while the data pipeline and output format remain fully compatible with the original. No NVIDIA GPU or Linux required.

Features

Predict and visualize in the browser

Submit jobs, track real-time progress, and explore results -- all from a local web interface. The 3D viewer (Mol*) lets you rotate, zoom, and inspect the predicted structure interactively.

Build complex inputs visually

Define multi-chain complexes with proteins, nucleic acids, ligands, and ions. Paste sequences, look up PDB entries, add post-translational modifications per residue, or upload existing input files. The entity builder validates input in real time.

Guide docking with restraints

Specify distance, contact, or repulsive restraints between chains to steer the diffusion process during structure generation. After prediction, a dedicated satisfaction panel reports which restraints were met and which were violated.

Analyze confidence at every level

Results include per-residue confidence (pLDDT), predicted aligned error (PAE), global fold confidence (pTM), and interface confidence (ipTM) for multi-chain complexes. Multi-sample ranking helps you pick the best prediction.

More highlights

• Native Apple Silicon -- M1 through M4 (Max and Ultra) with unified memory
• CLI -- single-command predictions from the terminal
• MSA caching -- content-addressed cache skips redundant HMMER searches
• Sequence-only mode -- run without genetic databases when they are unavailable
• REST API -- programmatic access for automation and pipelines

Supported Hardware

Requires an Apple Silicon Mac with a Max or Ultra chip (M1 through M4). Minimum 36 GB unified memory recommended. Larger proteins and multi-chain complexes require more RAM -- as a rough guide, a single-chain protein of ~500 residues fits comfortably in 64 GB, while complexes with thousands of residues benefit from 128 GB or more.

Quick Start

1. Install

git clone https://github.com/omrikais/alphafold3-mac.git
cd alphafold3-mac
./scripts/install.sh

The interactive installer sets up Python, MLX, HMMER, the web UI, and optionally downloads genetic databases (~500 GB). See the full Installation guide for details.

2. Obtain model weights

Request access to the AlphaFold 3 model parameters from Google DeepMind via this form. Place the downloaded af3.bin.zst in the weights directory configured during installation (default ~/.alphafold3/weights/model/).

3. Run a prediction

Web interface:

./scripts/start.sh
# Open http://127.0.0.1:8642

CLI:

source .venv/bin/activate
PYTHONPATH=src python3 run_alphafold_mlx.py \
  --input examples/desi1_monomer.json \
  --output_dir output/my_prediction

Documentation

Browse the full documentation at omrikais.github.io/alphafold3-mac, or read the Markdown sources directly in the docs/ directory.

Key pages:

• Quickstart
• Input Format
• Output Format
• Web Interface
• Restraint-Guided Docking
• CLI Reference
• API Reference
• Performance Tuning
• Troubleshooting

Architecture

Web UI + REST API          Next.js 15 + FastAPI
        |
Data Pipeline (unchanged) HMMER / MSA / Templates
        |
Model Inference (MLX)      Evoformer -> Diffusion -> Confidence
        |
Post-processing            mmCIF output, confidence scores

The original src/alphafold3/ data pipeline is preserved. Model inference lives in src/alphafold3_mlx/ and runs entirely on Apple GPU via MLX.

Citing This Work

Any publication that discloses findings arising from using this source code, the model parameters, or outputs produced by those should cite:

Abramson, J. et al. "Accurate structure prediction of biomolecular interactions with AlphaFold 3." Nature 630, 493--500 (2024). doi:10.1038/s41586-024-07487-w

BibTeX

@article{Abramson2024,
  author  = {Abramson, Josh and Adler, Jonas and Dunger, Jack and others},
  title   = {Accurate structure prediction of biomolecular interactions
             with {AlphaFold} 3},
  journal = {Nature},
  year    = {2024},
  volume  = {630},
  number  = {8016},
  pages   = {493--500},
  doi     = {10.1038/s41586-024-07487-w}
}

License

The AlphaFold 3 source code is licensed under CC-BY-NC-SA 4.0. Model parameters are subject to the AlphaFold 3 Model Parameters Terms of Use.

Based on AlphaFold 3 by Google DeepMind. This is not an officially supported Google product.