What it does AlphaFold 3 predicts 3D structures of biomolecular interactions—proteins, nucleic acids, ligands, and more. This repo contains the full inference pipeline: Dockerized setup, data preprocessing (genetic search, template lookup), and the GPU-powered neural network run. Feed it a JSON file with sequences, get back predicted coordinates.

The interesting bit The code is open; the brains are not. Model parameters require a Google-approved application with a 2–3 business day turnaround, and usage is bound by separate terms of use. It’s a peculiar hybrid: enough transparency to reproduce the method, enough control to keep DeepMind’s lawyers employed.

Key highlights

• Docker-first deployment with separate CPU data pipeline and GPU inference stages
• Supports proteins, DNA, RNA, ligands, and post-translational modifications (broader than AlphaFold 2)
• Non-commercial web interface available at alphafoldserver.com with reduced chemical scope
• CC-BY-NC-SA 4.0 license on code; custom terms on weights
• Published in Nature 2024; citation required for any disclosed findings

Caveats

• Not for clinical use; explicitly disclaimed for medical decisions
• “Not an officially supported Google product”
• Known issues documented but not enumerated in the README itself

Verdict Structural biologists and method developers should apply for access; commercial drug hunters are explicitly not invited to the party. If you just want to fold a protein casually, the web server is less paperwork.