OpenFold provides a GPU-friendly, memory-efficient PyTorch reproduction of DeepMind’s AlphaFold 2, the landmark deep learning model for predicting protein 3D structures from amino acid sequences. Unlike the original, OpenFold is fully trainable, enabling researchers to fine-tune, retrain, and study the learned mechanisms of protein folding. It reproduces AlphaFold’s multi-stage architecture including evoformer blocks, structure modules, and template/search encoders.