Open-AF3

Why we included this project

Structural biologists and drug discovery teams who want to work with AlphaFold3 without depending on DeepMind's closed codebase will find a readable PyTorch reimplementation here. The code walks through the full pipeline in modular pieces: the pairformer trunk, the MSA block, and the diffusion module that denoises raw atomic coordinates into predicted structures, each with short runnable examples. Since it is built on plain PyTorch tensors and installs with pip, it is a practical base for researchers who want to adapt the architecture to their own inputs or trace how the confidence head estimates per-atom and pairwise errors. Keep in mind this is an independent community implementation, not a drop-in trained model, so treat it as a research foundation rather than a production structure prediction service.