Protein Folding
Protein folding is the process by which a chain of amino acids, synthesized as a linear sequence, spontaneously adopts the specific three-dimensional shape that lets the protein function, and protein structure prediction is the computational problem of determining that shape from the sequence alone. Shape matters because function follows form: enzymes, antibodies, receptors and structural proteins work through the geometry of surfaces, pockets and interfaces, and a misfolded protein can lose function or aggregate, as in diseases such as prion disorders.
The foundational idea, associated with Christian Anfinsen's experiments on ribonuclease in the 1950s and early 1960s, is that the sequence contains the information needed to fold. Anfinsen showed that a protein chemically unfolded in the laboratory could refold into its active form once the denaturing conditions were removed, implying that the native structure is the thermodynamically favored state determined by the amino-acid sequence; he shared the 1972 Nobel Prize in Chemistry for this work. That principle, often called Anfinsen's dogma, is what made prediction from sequence conceivable in the first place, with the important qualification that in living cells some proteins need chaperones or partners, and some segments stay disordered.
Cyrus Levinthal pointed out in 1969 that a chain cannot find its shape by trying every conformation. Even with a crude model of a few possible orientations per residue, the number of conformations for a typical protein is astronomically large, vastly beyond what could be sampled in the milliseconds to seconds that real proteins take to fold. The resolution, usually called the Levinthal paradox, is that folding is not a random search but a directed descent on an energy landscape shaped like a funnel, with local interactions guiding the chain quickly toward the native state. The paradox explains why brute-force simulation was hopeless and why the problem was long regarded as one of biology's grand challenges.
The community yardstick is CASP, the Critical Assessment of protein Structure Prediction, a blind competition begun in 1994 in which groups predict structures that experimentalists have solved but not yet published. Results are scored with measures such as the global distance test (GDT_TS), where a score around 90 out of 100 is generally taken to be comparable to the uncertainty of experimental methods. Progress was gradual for two decades until DeepMind's first AlphaFold stood out at CASP13 in 2018, and AlphaFold2 then performed at CASP14, whose results were announced in late 2020, with a median accuracy high enough that the organizers described the central structure-prediction problem as largely solved for single-chain proteins. Jumper and colleagues published the method in Nature in 2021; the system uses an attention-based network over multiple sequence alignments and residue-pair representations, exploiting evolutionary coevolution signals, and an independent system, RoseTTAFold from David Baker's group (Baek et al., 2021), reached related accuracy.
In July 2021 DeepMind and the European Bioinformatics Institute launched the AlphaFold Protein Structure Database, and it was later expanded to hundreds of millions of predicted structures covering most known protein sequences. Researchers can retrieve a model for a protein of interest in seconds rather than waiting for months of crystallography or cryo-electron microscopy. In October 2024 the Nobel Prize in Chemistry was awarded with one half to David Baker for computational protein design and the other half jointly to Demis Hassabis and John Jumper for protein structure prediction, recognizing both reading and writing of the protein sequence-structure code.
A common misconception is that AlphaFold simulates the folding process or solved the protein-folding problem entirely. It predicts a static structure, typically the most likely one, and says little by itself about the pathway or kinetics, flexible and disordered regions, the effect of single mutations on stability, protein dynamics, or how ligands and other molecules bind, although later versions such as AlphaFold 3 (Abramson et al., 2024) extended prediction to complexes with nucleic acids and small molecules. Low-confidence regions in a prediction should not be treated as structure, and a predicted model is a hypothesis to be checked against experiment, especially before using it for drug design.
Read protein folding as the problem of how one-dimensional sequence determines three-dimensional function: Anfinsen showed the sequence suffices, Levinthal showed the search must be guided, CASP kept the field honest, and AlphaFold2 turned that insight into a practical prediction tool whose outputs remain static, confidence-labeled models rather than a movie of folding.