AlphaFold
By Google DeepMind
AlphaFold is a deep learning system developed by Google DeepMind that predicts a protein's three-dimensional structure from its amino acid sequence with accuracy approaching experimental methods, a problem known as the "protein folding…
Definition
AlphaFold is a deep learning system developed by Google DeepMind that predicts a protein's three-dimensional structure from its amino acid sequence with accuracy approaching experimental methods, a problem known as the "protein folding problem" that had challenged biology for decades.
Overview
Proteins are chains of amino acids that fold into complex 3D shapes, and a protein's function is largely determined by that shape. Determining structure experimentally — via methods like X-ray crystallography or cryo-electron microscopy — is slow and expensive, often taking months or years per protein. AlphaFold was built to predict structure computationally instead, directly from the amino acid sequence alone. The breakthrough version, AlphaFold2, was entered into CASP14 (the 2020 Critical Assessment of Structure Prediction, biology's standard benchmark competition) and achieved a median accuracy score that DeepMind and independent assessors described as competitive with experimental methods — a result widely regarded as a landmark achievement in computational biology. Architecturally, AlphaFold2 combines a Transformer-based system with structure-specific components, including an "Evoformer" module that processes evolutionary relationships across similar protein sequences (multiple sequence alignments) and geometric modules that reason directly about 3D atomic coordinates. In 2021, DeepMind and the European Bioinformatics Institute released the AlphaFold Protein Structure Database, providing predicted structures for nearly all catalogued proteins known to science — over 200 million structures — freely available to researchers worldwide. This dramatically accelerated research in areas such as drug discovery, enzyme design, and understanding disease mechanisms, since researchers no longer needed to wait for or fund experimental structure determination for many proteins of interest. AlphaFold's success, along with related deep learning research like AlphaGo and AlphaZero, contributed to the 2024 Nobel Prize in Chemistry being awarded to DeepMind researchers Demis Hassabis and John Jumper (shared with David Baker), recognizing computational protein structure prediction and design as a transformative scientific tool.
Key Concepts
- Predicts 3D protein structure directly from amino acid sequence
- Achieved near-experimental accuracy at CASP14 in 2020 with AlphaFold2
- Combines Transformer-based Evoformer modules with 3D geometric reasoning
- Uses multiple sequence alignments to leverage evolutionary information
- Underpins the freely available AlphaFold Protein Structure Database (200M+ structures)
- Recognized by the 2024 Nobel Prize in Chemistry