AlphaFold is a family of artificial intelligence systems developed by Google DeepMind for predicting the three-dimensional structures of proteins and, in later versions, biomolecular complexes. It uses deep learning to infer molecular geometry from sequence information and other inputs. AlphaFold 2 established substantially improved accuracy in protein structure prediction, while AlphaFold 3 extended the approach to complexes containing proteins, nucleic acids, and other molecular components. Its outputs are computational predictions rather than experimentally determined structures. (ebi.ac.uk)
Scientific background
A protein consists of a sequence of amino acids, but its biological activity depends strongly on its three-dimensional arrangement. Experimental techniques such as X-ray crystallography, cryo-electron microscopy, and nuclear magnetic resonance provide structural information. Nevertheless, experimentally characterized structures cover only a fraction of known protein sequences, creating demand for accurate computational models. (nature.com)
Protein structure prediction is related to, but distinct from, explaining protein folding. A prediction system can estimate a folded structure without reproducing the physical sequence of events through which a protein reaches that state. AlphaFold therefore does not simulate the complete folding pathway or establish every aspect of folding dynamics. Its principal achievement concerns structural inference from biological information. (ebi.ac.uk)
Development and milestones
The first AlphaFold system achieved leading results at the 2018 Critical Assessment of Protein Structure Prediction assessment, known as CASP13. It used neural networks to predict geometric relationships between amino acid residues, which guided construction of three-dimensional models. CASP evaluates predictions against experimentally determined structures that are not publicly available to participating predictors. (nature.com)
AlphaFold 2 introduced a substantially redesigned architecture and demonstrated a major improvement at CASP14 in 2020. Its results were competitive with experimental structures in a majority of assessed cases. The detailed methodology appeared in Nature on July 15, 2021, accompanied by release of the software code. AlphaFold-Multimer subsequently adapted the approach to predicting assemblies containing multiple protein chains. (nature.com)
AlphaFold 3, described in Nature on May 8, 2024, was developed by Google DeepMind and Isomorphic Labs. In October 2024, Demis Hassabis and John Jumper jointly received half of the Nobel Prize in Chemistry for protein structure prediction; the other half went to David Baker for computational protein design. (nature.com)
How AlphaFold 2 works
AlphaFold 2 combines machine learning with representations tailored to molecular geometry. Its training data include experimentally determined structures from the Protein Data Bank and extensive sequence information. For a target sequence, its processing pipeline searches sequence databases to construct a multiple sequence alignment, or MSA, of related proteins. Patterns across these sequences supply information relevant to structural constraints. Previously determined structures can also provide templates. (nature.com)
The central network component, called the Evoformer, exchanges information between sequence-alignment representations and representations of residue pairs. It uses attention mechanisms to process these relationships. A structure module then generates a three-dimensional backbone and places amino acid side chains. Rather than producing coordinates only once, the system “recycles” intermediate representations and the predicted structure through the network to refine its output. These computational iterations are not a physical folding trajectory. (nature.com)
AlphaFold 3 and molecular complexes
AlphaFold 3 predicts joint structures containing proteins, nucleic acids, small molecules, ions, and modified residues. Its supported interaction categories include protein–DNA, protein–RNA, and protein–ligand complexes, as well as antibody–antigen interactions. The original study reported improved performance over selected specialized methods on benchmarks spanning several of these categories. Such results describe performance on particular evaluation sets, not guaranteed accuracy for every complex. (nature.com)
Its architecture replaces AlphaFold 2’s main Evoformer with a Pairformer and uses a diffusion model to generate atomic coordinates by progressively transforming noisy configurations into predicted structures. This supports a unified approach across different molecular types, rather than restricting coordinate generation to protein chains. (ebi.ac.uk)
Confidence and interpretation
AlphaFold predictions include confidence estimates. For AlphaFold 2, predicted local distance difference test scores, or pLDDT, express confidence in local structure on a scale from 0 to 100. High scores indicate greater expected local accuracy; low scores may identify uncertain or disordered regions. A high local score does not, by itself, establish that two protein domains or chains have the correct relative orientation. (alphafold.ebi.ac.uk)
Predicted aligned error, or PAE, is presented as a matrix describing expected positional error after alignment on another part of the model. It helps distinguish confidence within individual domains from confidence in their relative arrangement. These measures concern predicted structural accuracy, not direct experimental confirmation of biological function. (ebi.ac.uk)
Database, access, and limitations
DeepMind and EMBL-EBI launched the AlphaFold Protein Structure Database on July 22, 2021, initially providing more than 350,000 predictions. An expansion announced on July 28, 2022 brought its coverage to over 200 million predicted structures. This resource makes precomputed models available without requiring users to run the prediction software themselves. (ebi.ac.uk)
Access conditions differ between implementations. AlphaFold 2’s code was publicly released, while AlphaFold 3’s published model-parameter terms restrict the supplied parameters and associated outputs to specified non-commercial uses. Code availability and permission to use trained parameters are therefore separate questions. (deepmind.google)
AlphaFold does not comprehensively describe conformational dynamics, folding pathways, or environmental effects. A predicted structure can assist experimental interpretation and research planning, but uncertain regions, missing molecular context, and alternative conformations remain important limitations. Structural predictions can accelerate experimental investigation without replacing verification of consequential molecular details. (ebi.ac.uk)