Definition
Protein folding is the process by which a newly made chain of amino acids collapses out of a floppy, disordered string into one specific three-dimensional shape. That shape is not decoration: it is the protein's function, because an enzyme's reaction pocket, an antibody's binding surface, and a channel's pore are all just features of the folded geometry. Get the fold and you have a working molecule; lose it and you usually have inert or harmful junk.
The reason the term shows up constantly in AI is a related but distinct idea called the protein folding problem: given only the amino acid sequence, predict the folded 3D structure by computation, without growing a crystal or running a microscope. That prediction problem sat largely unsolved for roughly 50 years and was cracked in practical form by deep learning - Google DeepMind's AlphaFold2 - at a 2020 contest called CASP14. So "protein folding" in an AI context almost always means this prediction task, not the underlying biophysics, even though the two share a name.
How It Works
A protein starts life as a linear sequence drawn from 20 possible amino acids, strung together in the order written in a gene. Chemically that string is like beads on a wire; biologically it is useless until it folds. Folding is driven mostly by the hydrophobic effect: some amino acids are water-repelling and some are water-loving, and in the watery inside of a cell the water-repelling ones are pushed to bury themselves in the core, dragging the chain into a compact ball. Hydrogen bonds then lock in local patterns - the corkscrew of an alpha helix, the pleated strands of a beta sheet - and the whole thing settles into the single arrangement with the lowest free energy, its native state. Chaperone proteins stand by to catch chains that start folding wrongly and give them another attempt.
Why predicting the fold is so hard: Levinthal's paradox
The obvious way to predict a structure by computer is to try every possible shape and keep the most stable one. Cyrus Levinthal pointed out in 1969 why that is hopeless, and the point is durable arithmetic that does not age. Take a modest 100-amino-acid protein and pretend, generously, that each residue can sit in just three orientations. The number of whole-chain shapes is then 3^100, which is about 5 x 10^47. Even sampling a trillion (10^12) of those shapes every second, an exhaustive search would take on the order of 10^35 seconds - more than 10^17 times the current age of the universe (about 14 billion years). Yet real proteins fold on their own in microseconds to seconds. This is Levinthal's paradox, framed formally by Zwanzig, Szabo, and Bagchi in a 1992 paper: a random search over all configurations "can take an enormously long time," yet "proteins can fold in seconds or less."
The resolution is that folding is not a blind search. The energy landscape is shaped like a funnel, so a chain is nudged downhill toward the native state rather than wandering flat ground. For a computer predicting structure, though, the paradox is the whole difficulty: you cannot brute-force the answer, and for decades physics-based simulation was too slow and too inaccurate to fold most proteins reliably.
AlphaFold2 sidestepped the search entirely. Instead of simulating physics, it learns from the roughly 200,000 structures humans have solved experimentally, and it leans hard on evolution: for a target sequence it gathers many related sequences from other organisms into a multiple sequence alignment, because pairs of amino acids that mutate in lockstep across species tend to be touching in the folded structure. A transformer-style neural network reads those evolutionary correlations and predicts the map of which residues are close together, from which the 3D coordinates are reconstructed. Small proteins that fold in a single cooperative step ("two-state" folders) and larger ones that fold through intermediates are handled the same way, because the model predicts the final structure rather than the folding pathway.
Real-World Applications
The clearest real deployment is the AlphaFold Protein Structure Database, run by Google DeepMind and EMBL-EBI. In July 2022 it expanded to over 200 million predicted structures - almost every protein catalogued in the UniProt sequence database, spanning nearly every sequenced organism - all free to download. Before this, a structural biologist might spend months or years solving one structure by X-ray crystallography or cryo-electron microscopy; now the predicted starting point is a lookup, which changes what an early-stage project can assume it already knows. A freely downloadable database covers the organisms nobody would have funded a structure for, which is why it is the example most often reached for in arguments about AI for good — and one of the few where the claim is checkable.
In drug work, a predicted structure is the target that AI drug discovery pipelines dock candidate molecules against: to design a small molecule that blocks a protein, you first need the shape of the pocket it must fit. Structure prediction has been used to model targets that had resisted experimental methods, giving medicinal chemists something concrete to design against. The same structures feed research into diseases of misfolding, where the failure mode is folding itself - beta-amyloid and tau aggregates in Alzheimer's disease, alpha-synuclein in Parkinson's, and the misfolded CFTR channel in cystic fibrosis are all cases where a protein's mischosen shape, not its absence, is the problem, and this connects the term to precision medicine.
Beyond reading natural proteins, groups such as David Baker's laboratory now run folding logic in reverse to design new proteins - specifying a desired shape and searching for a sequence that will fold into it - producing binders, enzymes, and vaccine scaffolds that do not exist in nature. Structure prediction turned this from a rare feat into a routine step, a shift tracked more broadly under AI in Science.
Key Concepts
Structure has four levels. Primary structure is the amino acid sequence; secondary structure is local patterns like alpha helices and beta sheets; tertiary structure is the full 3D fold of one chain; quaternary structure is how several folded chains assemble into a complex. "Predicting structure" usually means getting the tertiary structure right.
GDT is how folding predictions are scored. CASP, the Critical Assessment of protein Structure Prediction, is the biennial blind contest that has benchmarked the field since 1994. Its main yardstick is the Global Distance Test (GDT), which runs from 0 to 100 and measures how many of the predicted atom positions land close to the experimentally determined structure. A score around 90 GDT is informally treated as competitive with experiment - close enough that the errors are comparable to the wobble between two lab measurements of the same protein.
CASP14 (2020) is the result the field points to. AlphaFold2 reached a median of 92.4 GDT across all targets, far ahead of every other method, and DeepMind's own report frames the win as reaching "accuracy competitive with experimental structures in a majority of cases." That figure is why the folding prediction problem is described as largely solved, even though the physical question of how a chain finds its fold in real time is not.
Confidence is part of the output. AlphaFold reports a per-residue confidence score alongside each structure, so a user can tell a trustworthy region from a shaky one. This matters because accuracy is not uniform: well-conserved, tightly folded cores score high, while flexible loops and floppy tails score low, and treating a low-confidence region as if it were solved is a common misuse.
Challenges
The headline number hides what structure prediction still does not deliver, and the gaps are specific to this problem rather than generic AI caveats.
- One shape, not the motion. A predicted structure is a single snapshot, but many proteins do their job by flexing between several shapes - a channel opening and closing, an enzyme clamping onto its substrate. Predicting that ensemble of states and the transitions between them is largely unsolved, and it is the part of folding the static prediction never addressed.
- Disordered regions have no fold to predict. A meaningful fraction of the human proteome is intrinsically disordered, staying floppy on purpose. For these stretches there is no single native structure, and a model trained to output one confidently returns low-confidence coordinates that should not be over-read.
- Small sequence changes, unpredictable stability shifts. Predicting how a single point mutation changes a protein's stability or whether it will misfold remains hard, even when the wild-type structure is predicted well - which is exactly the question a disease geneticist or protein engineer most wants answered.
- Dependence on evolutionary depth. Accuracy leans on finding many related sequences to build the alignment. For orphan proteins, fast-evolving viral proteins, and designed sequences with no natural relatives, that evolutionary signal is thin and predictions weaken.
Future Trends
The direction of travel is away from the single-chain snapshot. AlphaFold3, released in 2024, extended prediction from lone proteins to complexes with DNA, RNA, ligands, and ions - the assemblies where biology actually happens - though its model weights are restricted to non-commercial use and granted only on request. In parallel, openly downloadable models such as ESMFold, OpenFold, and RoseTTAFold put comparable capability in labs' own hands without a licence gate. The most active frontier joins prediction to generation: diffusion-based tools like RFdiffusion design brand-new proteins to order, and generative methods aim at the moving target of conformational ensembles that the current, static predictors deliberately leave out.