Protenix

Protenix is ByteDance's open-source reproduction of AlphaFold 3 for high-accuracy biomolecular structure prediction. The Pairformer plus diffusion architecture predicts all-atom 3D structures for complexes mixing proteins, DNA, RNA, small-molecule ligands, and ions, with support for post-translational modifications, MSAs, structural templates, and soft pocket or contact constraints.

The full stack ships under Apache 2.0: training and inference code, pretrained weights, data and MSA pipelines, and a browser-based web server. Protenix was the first open-source model to match or beat AlphaFold 3 on head-to-head benchmarks, and the family includes lightweight Protenix-Mini variants plus PXDesign for binder design. The honest caveats: full-accuracy runs want a 24GB+ GPU and the MSA pipeline adds real latency, and as a reproduction it inherits AlphaFold 3's blind spots on highly flexible or disordered regions.