AlphaFold: How AI Solved Biology's 50-Year Problem

Proteins are the machines of biology. Chains of amino acids fold into precise three-dimensional shapes, and those shapes determine what the proteins do, whether they catalyse a reaction, form a membrane, carry oxygen, or signal between cells. For 50 years, determining a protein's structure from its amino acid sequence was one of biology's hardest open problems. X-ray crystallography and cryo-electron microscopy could measure structures, but were slow, expensive, and often impossible for certain proteins. The gap between known sequences and known structures was growing.

The CASP competition (Critical Assessment of protein Structure Prediction) had run biannually since 1994, benchmarking computational methods for predicting protein structure. DeepMind entered CASP13 in 2018 with AlphaFold 1 and placed first. At CASP14 in November 2020, AlphaFold 2 achieved something that stunned the structural biology community: it predicted protein structures with accuracy comparable to experimental methods, a Global Distance Test score of over 90, where previous best methods scored around 70.

AlphaFold 2's architecture combined multiple sequence alignments (to find evolutionarily related proteins), a novel "Evoformer" transformer block that jointly processed sequence and pairwise residue relationships, and a structure module that used the learned representations to predict atomic coordinates. The training signal was experimental structures from the Protein Data Bank. It was, at its core, a transformer applied to biology with exceptional care for the domain's geometry.

In 2021, DeepMind released AlphaFold 2's code and weights openly, and partnered with the European Bioinformatics Institute to release predicted structures for 200 million proteins, essentially every known protein in every organism sequenced. The database has been used to accelerate drug discovery, understand disease mechanisms, and design new enzymes. John Jumper and Demis Hassabis received the 2024 Nobel Prize in Chemistry for AlphaFold. The problem that had stumped structural biology for 50 years was solved by a team that had been working on games four years earlier.