What is protein folding?

How it works

Proteins are chains assembled, in general, from 20 kinds of amino acids. Each chain folds into a three-dimensional structure that determines what the protein does. That comes from the Royal Swedish Academy of Sciences' popular-science background to the 2024 Nobel Prize in Chemistry. It says the American scientist Christian Anfinsen concluded in 1961 that the amino acid sequence alone governs the final shape.

In 1969, Cyrus Levinthal pointed out a paradox. A chain of just 100 amino acids could in theory take at least 10 to the power of 47 shapes. That is far too many to find by folding at random. Yet in a cell, folding takes a few milliseconds.

The protein folding problem

The European Bioinformatics Institute (EMBL-EBI) describes two linked challenges: understanding the process by which a chain folds, and predicting the final structure from the sequence. AI tools such as AlphaFold address the second.

Lab methods are slow. Determining one structure can take months to years, researchers at DeepMind (now Google DeepMind) wrote in a July 2021 Nature paper. They said about 100,000 unique structures had been determined, against billions of known sequences. The paper presented the system now called AlphaFold2 as the first computational method to regularly predict structures with atomic accuracy. It cited its results in CASP14, a blind prediction test held in 2020. Demis Hassabis and John Jumper of Google DeepMind shared half of the 2024 Nobel Prize in Chemistry for protein structure prediction.

Where things stand in 2026

Google DeepMind and EMBL-EBI developed the AlphaFold Protein Structure Database. As of October 2026, its home page says it gives open access to more than 260 million predicted structures. Background text lower on the same page puts the latest release at over 200 million entries.

Later models cover more than proteins. AlphaFold 3, described in Nature in May 2024, predicts the structure of complexes that combine proteins with nucleic acids (DNA and RNA), small molecules and ions.

Limits remain. EMBL-EBI's training material says AlphaFold2 was designed to predict static snapshots. So by default, it says, the model does not capture the shape changes proteins go through as they work. Out of the box, it is also not sensitive to mutations that change a single amino acid. EMBL-EBI adds that predicting the final shape does not necessarily require understanding the folding process, the first of the two challenges.