AlphaFold 3 is an inference pipeline for protein structure prediction developed by Google DeepMind. The repository contains the complete source code necessary to run AlphaFold 3 predictions, written primarily in Python. Access to the model parameters themselves requires a separate request through a Google form, with responses typically provided within two to three business days. The model parameters are subject to specific terms of use and may only be used if received directly from Google.
The inference pipeline operates in two main stages that can be controlled independently. The data pipeline stage handles genetic and template searches and runs on CPU only, making it time-consuming but executable on machines without GPU access. The inference stage requires GPU resources and performs the actual structure prediction. Users can toggle these stages on or off using command-line flags when running the main prediction script.
AlphaFold 3 accepts input in JSON format and produces structured output documenting predicted structures and associated confidence metrics. The repository includes comprehensive documentation covering installation procedures, input specifications, output formats, and known performance characteristics. Users encountering issues are directed to check the known issues documentation before creating new issue reports.
The codebase builds on several established scientific and computational libraries. Key dependencies include JAX for numerical computation, Haiku for neural network implementation, RDKit for molecular chemistry operations, and HMMER Suite for sequence analysis. The pipeline also integrates DSSP for secondary structure assignment and uses various bioinformatics databases including mirrored versions of BFD, PDB, MGnify, UniProt, and UniRef90, along with nucleotide and RNA databases.
Community engagement around the repository shows active maintenance and responsiveness. Across 586 tracked issues and pull requests, the median response latency is 5.1 hours with a mean of 25.6 hours. The most common issue category is questions, with 395 labeled items, followed by setup-related issues with 77 items and enhancement requests with 52 items. Augustin Žídek serves as the engineering lead and is the most active contributor with 1861 tracked events, followed by joshabramson with 278 events and jsspencer with 57 events.
The repository maintains connections to related projects through overlapping contributors, linking to tracel-ai/burn, jwohlwend/boltz, and huggingface/transformers. This indicates broader integration within the machine learning and bioinformatics ecosystem.
Publications using AlphaFold 3 must cite the primary paper "Accurate structure prediction of biomolecular interactions with AlphaFold 3" published in Nature. The source code is licensed under Apache License 2.0, while model parameters are governed by separate terms of use. The software is explicitly not intended for clinical use and outputs should be interpreted as theoretical predictions with varying confidence levels. AlphaFold 3 is also available through alphafoldserver.com for non-commercial use, though with a more limited set of supported ligands and covalent modifications compared to the full inference pipeline.
