Timeline of a breakthrough

DeepMind’s AlphaGo program defeats legendary Go player Lee Sedol in a challenge match in Seoul. This watershed moment demonstrated that DeepMind’s AI techniques were potentially advanced enough to be applied to scientific challenges including the “protein-folding problem”. Shortly after, DeepMind establishes a small team to begin work on protein structure prediction.

AlphaFold’s performance is benchmarked in the 13th Critical Assessment of Protein Structure Prediction (CASP13), placing first in the rankings (under entry A7D). The methods are subsequently published in the scientific journal Nature. The team is expanded, and work begins on an innovative new system.

AlphaFold2 wins CASP14 by a huge margin and is recognised as a solution to the 50-year-old “protein-folding problem” by the organisers of CASP after predicting structures down to atomic accuracy with a median error (RMSD_95) of less than 1 Angstrom – 3 times more accurate than the next best system and comparable to experimental methods.

John Jumper and Demis Hassabis each give 30-minute presentations about the ideas, architecture, and publication plans of the AlphaFold system to CASP14 attendees, confirming DeepMind’s commitment to provide broad access to our work.

Nature publishes AlphaFold’s detailed methodology in the paper “Highly accurate protein structure prediction with AlphaFold” and DeepMind open sources the code along with 60 pages of supplemental information detailing every aspect of the system.

A week later, Nature publishes a second DeepMind paper containing the structure predictions of the entire human proteome, doubling the number of high confidence structures known. In close collaboration with the European Bioinformatics Institute at the European Molecular Biology Laboratory (EMBL-EBI), DeepMind launches the AlphaFold Protein Structure Database to give the scientific community free and open access to the human proteome along with another 20 model organisms – over 350,000 structures in total.

DeepMind publishes a further paper on biorxiv, “Protein complex prediction with AlphaFold-Multimer”, which properly accounts for multi-chain proteins, and demonstrates superior performance in predicting complexes compared to existing approaches, including vanilla AlphaFold2.

DeepMind updates the AlphaFold2 source code to account for multi-chain protein complexes – providing a significant improvement in accuracy for predicting protein interactions.

DeepMind adds more than 400k protein structures to the AlphaFold Protein Structure Database with EMBL-EBI. This release included structures for most proteins in the added predictions for most of the manually-curated UniProt entries in UniProtKB/SwissProt, more than doubling the size of the database.

DeepMind adds 27 new proteomes (190k+ proteins) to the AlphaFold Protein Structure Database with EMBL-EBI, 17 of which represent Neglected Tropical Diseases that continue to devastate the lives of more than 1 billion people globally. Over 300,000 researchers worldwide have made use of the database to date.

DeepMind expands the AlphaFold Protein Structure Database from nearly 1 million to over 200 million structures, including predictions for most proteins in UniProt.