AlphaFold's database expands to 200 million structures

DeepMind and its partner EMBL’s European Bioinformatics Institute expanded the AlphaFold Protein Structure Database from roughly a million predicted structures to more than 200 million, adding structural predictions for almost every protein catalogued in UniProt, the main reference database used by biologists. Coverage grew from around 10,000 species to close to a million, meaning most pages on UniProt would now carry a predicted three-dimensional structure alongside the protein’s sequence.

The expansion followed the initial 2021 release of AlphaFold’s predictions for the human proteome and a small set of other organisms. The July 2022 release generalised the model’s output across the tree of life, covering plants, bacteria and animals as well as many understudied disease-relevant organisms, and made the full dataset available for bulk download through Google Cloud’s public datasets programme. DeepMind reported the database had already been accessed by more than 500,000 researchers across 190 countries in the year since its first release, ahead of this larger expansion.

AlphaFold’s predictions are computed structures, not experimentally verified ones, and the accompanying confidence scores meant some predictions — particularly for disordered or poorly conserved protein regions — carried substantial uncertainty. Structural biologists nonetheless described the change in practical terms: a task that had taken years of crystallography or cryo-electron microscopy per protein could now return a candidate structure in seconds, to be checked and refined rather than built from scratch. The release is widely cited as one of the clearest cases of an AI system producing a public scientific resource used well beyond the field that built it, predating DeepMind and Demis Hassabis’s 2024 Nobel Prize in Chemistry for the underlying work.