Textbook of AI

AlphaFold is the protein-structure-prediction system developed by DeepMind. The first version (Senior et al., 2018, published 2020) won the CASP13 biennial blind-prediction competition by a substantial margin. The dramatically more powerful AlphaFold 2 (Jumper et al., 2020, published 2021) won CASP14 with results that scientists in the field declared as "essentially having solved" the protein-folding problem.

Protein folding, predicting 3D structure from amino-acid sequence, had been one of the great open problems of molecular biology since the 1960s. CASP14 GDT-TS scores around 92.4 (out of 100) on hard targets approached experimental accuracy. The result transformed structural biology overnight.

AlphaFold 2's architecture combines: An Evoformer stack, Transformer-style attention over multiple sequence alignments and pair representations of residues, allowing the model to extract co-evolutionary signals. A structure module, equivariant attention over residue frames, iteratively refining 3D coordinates. End-to-end training on the Protein Data Bank.

DeepMind released the AlphaFold Protein Structure Database in 2021, providing predicted structures for over 200 million proteins (nearly every known protein sequence). The database is now a central resource of structural biology and has been used in tens of thousands of subsequent biomedical research papers.

Demis Hassabis and John Jumper shared the 2024 Nobel Prize in Chemistry with David Baker for the AlphaFold work. Successor systems, AlphaFold 3 (2024) extends the framework to protein-ligand and protein-nucleic-acid complexes; ESMFold and others use evolutionary scale models more directly, continue to extend the capabilities.