Every cell depends on proteins, but scientists have seen the shape of only a tiny fraction of them. AlphaFold showed that a sequence of chemical letters can be enough to predict a protein’s three-dimensional form with near-experimental accuracy.

Proteins are essential to life, and understanding their structure can help explain how they work. But only around 100,000 unique protein structures have been determined, a small fraction of the billions of known protein sequences.

The obstacle is time: determining a single protein structure can require months to years of painstaking effort. Predicting the three-dimensional structure from its amino acid sequence has therefore remained an important open problem for more than 50 years.

AlphaFold addresses this gap by predicting protein structures with atomic accuracy even when no similar structure is known. AlphaFold predicts the three-dimensional coordinates of all heavy atoms in a protein from its primary amino acid sequence together with aligned sequences from related proteins.

AlphaFold uses the protein’s amino acid sequence and aligned sequences of related proteins as inputs, then directly predicts the three-dimensional coordinates of all its heavy atoms. The test was deliberately blind. The assessment used recently solved structures that had not been publicly disclosed, so the prediction system could not simply look up the answers.

AlphaFold was entered into this assessment as a completely different model from the earlier AlphaFold system, and it predicted structures to near-experimental accuracy in a majority of cases. For protein researchers and patients awaiting better biological predictions, the striking result is that predicted shapes nearly overlap experimentally measured ones—even for a zinc-binding region and a chain 2,180 building blocks long—making reliable structure prediction look practical rather than merely theoretical.

AlphaFold’s median backbone error was about one carbon-atom width, compared with nearly twice that for the next best method. Its full-atom predictions were also considerably more accurate than the best alternative. The system could also handle very long proteins, including one with 2,180 residues and no known structural relative.

It provided a separate reliability estimate for each residue, helping indicate where a prediction could be trusted. The result was not limited to the blind assessment. High accuracy and reliability also carried over to a collection of recent protein structures, including structures representing new folds.

When the backbone prediction was accurate, the smaller side chains were also accurate. The system’s confidence measure reliably predicted the accuracy of the corresponding part of the structure. But AlphaFold is not equally reliable everywhere.

Accuracy decreases substantially when the alignment contains fewer than around 30 related sequences, while adding more than around 100 produces only small gains. It is also weaker when a protein’s shape depends mainly on contacts with other protein chains, such as a bridging part of a larger complex.

It is often accurate for proteins made from intertwined copies of the same chain. That distinction matters because AlphaFold is already helping experimental scientists replace missing structures and interpret maps made with cryogenic electron microscopy.

Its predictions can be produced in minutes to hours of computing, depending on protein length. This opens the possibility of predicting protein structures at proteome scale and beyond; a companion paper demonstrates applying AlphaFold to the entire human proteome.

AlphaFold turns a problem that could take months or years of laboratory work into a prediction that can guide experiments much sooner, while also indicating which parts of that prediction are trustworthy.