AlphaFold is a Deep learning system developed by DeepMind that predicts a protein's three-dimensional structure from its amino acid sequence. Predicting how a linear chain of amino acids folds into the complex three-dimensional shape that determines a protein's biological function had been an open problem in structural biology for roughly fifty years, known as the protein folding problem, since experimentally determining structures through methods like X-ray crystallography or cryo-electron microscopy is slow and expensive.
CASP14 and the 2020 breakthrough
AlphaFold's second version, AlphaFold2, achieved a result at the 2020 Critical Assessment of protein Structure Prediction competition, known as CASP14, that organizers and independent structural biologists described as effectively solving the protein folding problem for the majority of proteins. CASP is a biennial blind competition in which research teams predict the structures of proteins whose real structures have been experimentally determined but not yet published, allowing organizers to score predictions against unpublished ground truth. AlphaFold2 achieved a median accuracy score competitive with the resolution of experimental methods for most target proteins, a level of accuracy that had not been approached by any prior computational method and that CASP's own assessors and outside biologists called a transformative result for the field.
Architecture
AlphaFold2 used a Transformer (architecture)-based architecture with a component called Evoformer, which processed both the amino acid sequence and evolutionary information drawn from multiple sequence alignments, related protein sequences across different species, alongside geometric constraints, to iteratively refine a predicted structure. DeepMind published the AlphaFold2 model architecture and code in 2021, and released predicted structures for nearly all proteins with known sequences in the UniProt database through the freely accessible AlphaFold Protein Structure Database (AlphaFold DB), a joint effort with the European Bioinformatics Institute, ultimately covering over 200 million protein structures.
Impact and the Nobel Prize
AlphaFold's open release of both code and predicted structures was widely credited with accelerating research across structural biology, drug discovery, and enzyme design, since researchers who previously needed months of experimental work or lacked access to structural determination facilities could obtain a high-confidence predicted structure in minutes. In 2024, DeepMind co-founder Demis Hassabis and AlphaFold lead researcher John Jumper shared half of the Nobel Prize in Chemistry for the development of AlphaFold, with the other half awarded to David Baker for separate computational protein design work; the award was widely covered as a landmark recognition of AI's direct contribution to a Nobel-level scientific achievement. DeepMind subsequently released AlphaFold3 in 2024, extending prediction capability beyond single proteins to interactions between proteins, DNA, RNA, and small-molecule ligands, of particular relevance to drug discovery.
Reception and limitations
AlphaFold is broadly considered one of the most significant applied successes of deep learning outside of language and vision, frequently cited alongside AlphaGo as evidence of DeepMind's research strategy of applying Reinforcement learning and deep learning methods to well-posed scientific problems with clear evaluation criteria. Researchers noted limitations, including lower reliability for intrinsically disordered proteins that lack a single stable structure and for predicting how proteins change shape dynamically or interact within a live cell, areas that remained active research directions after AlphaFold3's release.