AI Tool Releases Database of One Billion Predicted Protein Structures

Researchers at the Chan Zuckerberg Initiative Biohub released a database containing more than one billion predicted protein structures along with billions of additional protein sequences. The database is called the ESM Atlas. It was generated using ESMFold2, an artificial-intelligence model developed by the Biohub team.

The ESM Atlas contains more than 800 million additional entries than the AlphaFold Database of predicted protein structures. It also exceeds a previous ESM Atlas by roughly 300 million entries. The predictions include metagenomic sequences drawn from soil, ocean, and other environments that are not present in the AlphaFold database.

Biohub states that ESMFold2 outperforms AlphaFold3 and other current protein-structure prediction systems when determining structures of protein complexes, including antibody-antigen interactions. The model is based on a protein language model released by the same team in 2024 and trained on billions of proteins from across the tree of life.

ESMFold2 is fully open source.

The atlas is described in a preprint released today. Alex Rives, Biohub science head, said the atlas shows the totality of protein biology and especially the parts that are most unknown. Other researchers noted the open-source status of the model while observing that multiple competing protein-prediction systems continue to advance.