alphafold3-pytorch

Why we included this project

AlphaFold 3 came out as a Nature paper in 2024, and this repository is a PyTorch port of its architecture that reads like it was written to be followed alongside the paper. It covers the full training path from parsing Protein Data Bank files down to molecule and atomic inputs, so it works both as a map of how the method fits together and as a base to build your own experiments on. The pieces that matter for diffusion-based structure prediction are all here, from relative positional encoding to smooth LDDT loss and weighted rigid alignment, and a separate Lightning + Hydra fork exists for anyone running larger training jobs. Because DeepMind keeps the official AlphaFold 3 weights and full server restricted, an open implementation like this is often the most practical way for academic groups to actually run these ideas on their own hardware.