alphafold2

Why we included this project

Protein structure prediction remains one of the hardest open problems in computational biology, and this project gives researchers a PyTorch take on the AlphaFold2 attention architecture, a more approachable alternative to the official JAX codebase for anyone who wants to train and experiment with it directly. The trunk runs axial self and cross attention over sequences and multiple-sequence alignments, and it can emit distograms, backbone angles, or predicted residue coordinates through equivariant structure modules like SE3 or E(n) transformers. Wrappers accept pretrained protein representations from ESM, MSA, and ProtTrans, so you can feed those straight into the model. Structural biologists and machine learning researchers will find it a readable starting point for experimentation, with support for custom atom configurations and reversible attention. Just keep in mind it is a work in progress, not a drop-in AlphaFold2 replacement, so treat it as a research toolkit rather than a production folding service.