OpenFold: Trainable PyTorch Reproduction of AlphaFold2 New

OpenFold (Ahdritz et al., 2022; Columbia AQLab) is a faithful but fully trainable PyTorch reproduction of DeepMind's AlphaFold2. It is designed to be memory-efficient and GPU-friendly, enabling researchers to train or fine-tune the full structure-prediction pipeline on their own data rather than only running inference with pretrained weights.

The repository ships training and inference code, and the installation script downloads DeepMind's pretrained parameters into the resources directory. Licensing: while AlphaFold's (and by extension OpenFold's) source code is under the permissive Apache License 2.0, DeepMind's pretrained parameters are distributed under CC BY 4.0 - this replaced the original, more restrictive CC BY-NC 4.0 license as of January 2022, so the parameters may now be used commercially under attribution.

Why it matters for physics AI: OpenFold democratized high-accuracy protein structure modeling by making a trainable, open implementation available, and it underpins large-scale datasets such as OpenProteinSet, accelerating structural-biology ML research beyond the original AlphaFold2 release.