AlphaFold 2

AlphaFold 2 won CASP14 in November 2020 with a median score of 92.4 GDT_TS — within experimental accuracy for two-thirds of targets — and DeepMind published the full method in Nature on July 15, 2021 while open-sourcing the code. Its Evoformer architecture jointly reasons over multiple sequence alignments and pairwise residue features with 48 attention blocks, followed by an SE(3)-equivariant structure module that outputs atomic coordinates directly. With EMBL-EBI, DeepMind built the AlphaFold Protein Structure Database, now holding over 200 million predicted structures used by millions of researchers in 190+ countries.

The code and model parameters are openly available (Apache 2.0), and the paper is among the most-cited in modern science. The honest limitations: AlphaFold 2 needs deep multiple sequence alignments so it struggles on orphan proteins, it predicts a single static conformation rather than dynamics, and it does not model protein-ligand or multi-chain complexes — gaps that AlphaFold3, RoseTTAFold, and Boltz-1 were built to address.