AlphaFold database expands to over 200 million protein structures

What the database holds

The AlphaFold Protein Structure Database is a public collection of predicted protein structures produced with AlphaFold, DeepMind's system for predicting the three-dimensional shape of a protein. Shape matters because a protein's function is closely tied to how it folds: a model of the fold gives researchers a starting point for asking what a protein does and how it might be interfered with.

What happened on 28 July 2022

On that day DeepMind and EMBL's European Bioinformatics Institute (EMBL-EBI) expanded the database from nearly 1 million predicted structures to more than 200 million^[1]. The new set covered nearly all catalogued proteins known to science, rather than a curated selection of especially studied ones^[1]. The release was made in partnership with EMBL-EBI, the bioinformatics institute that hosts and maintains the resource^[1].

Key details of the release

All of the structures were made available for bulk download through Google Cloud Public Datasets, so that groups could take the whole archive rather than query it page by page^[1]. The expanded database includes predicted structures for plants, bacteria, animals and other organisms, not only for humans and a handful of model species^[1]. Because the structures are tied into UniProt, the widely used protein sequence catalogue, most pages on UniProt now come with a predicted structure attached^[1].

DeepMind also reported the reach of the resource: more than 500,000 researchers from 190 countries had already accessed the database by the time of the expansion^[1]. The announcement was written by Demis Hassabis on the DeepMind blog^[1].

Why the scale matters

A database of this size changes how structural information is used. When a structure can be looked up for a protein, a lab does not have to treat the shape as a separate project before it can plan an experiment; the structure becomes an ordinary piece of background information, like the sequence itself. Bulk download matters for the same reason — it lets a whole proteome be analysed at once instead of one entry at a time, and it lets the archive be mirrored or built into other tools.

The breadth of organisms is the other part of the change. Structures for plants, bacteria and other organisms widen the set of questions that can be asked with the same resource, from crop biology to microbial proteins that have never been studied in detail^[1].

What it is not

A predicted structure is a model, not an experimental measurement, and a database entry is a starting point for a researcher rather than a finished result. The value of the expansion lies in coverage and access: hundreds of millions of models, downloadable in bulk, sitting next to the sequence records that people already use^[1].