Proteins: reading nature's shapes and writing new ones

A protein is a tiny machine inside every living thing. It begins as a long chain of small parts called amino acids, strung together like beads on a thread. That thread folds up into one exact 3D shape, and the shape is what lets the protein do its job.

Scientists could easily read the order of the beads, but for 50 years they could not work out what shape the thread would fold into. A chain can fold in so many ways that trying them all would take longer than the universe has existed. This puzzle was called the protein folding problem.

David Baker went the other way. Instead of predicting the shape of a natural protein, he designs a shape he wants, then finds a brand-new chain that folds into it. These are proteins that have never existed in nature.

A protein is a chain of amino acids drawn from an alphabet of 20. The order of those amino acids is written in a gene, and once the chain is made it folds, usually on its own, into a particular 3D structure. That structure decides the function. In 1972 Christian Anfinsen argued that the sequence alone should hold all the information needed to specify the fold, an idea now known as Anfinsen's dogma.

If the sequence sets the structure, you ought to be able to compute the structure from the sequence. The trouble is scale. In 1969 Cyrus Levinthal pointed out that a typical chain has an astronomical number of possible shapes, so many that sampling them one at a time would take longer than the age of the universe. Yet real proteins fold in fractions of a second. Predicting the final shape from sequence became the protein folding problem, and progress was scored every two years at a blind contest called CASP.

David Baker attacked the mirror-image problem. His program Rosetta searches for the lowest-energy fold of a sequence, and he realised the same machinery could run backward: choose a structure, then design a sequence to fit it. In 2003 his team built Top7, a 93-amino-acid protein with a fold unlike anything in nature, and its measured X-ray structure matched the design closely. Baker released Rosetta openly, and his group has since designed proteins that serve as vaccines, medicines, sensors, and nanomaterials.

The scientific premise is Anfinsen's thermodynamic hypothesis: for most small proteins the native fold is the global minimum of free energy, so the amino acid sequence encodes the structure. That is what makes prediction well posed in principle. It also dissolves Levinthal's paradox in practice, because predicting the endpoint of folding does not require simulating the kinetic pathway through the conformational landscape. The two halves of the 2024 prize sit on this premise from opposite directions: read structure from sequence, or write a sequence for a chosen structure.

AlphaFold2 infers by learning from evolution. It takes a multiple sequence alignment of related proteins plus a residue-residue pair representation and passes them through the Evoformer, a stack of attention blocks that lets the alignment and the pair map exchange information repeatedly. Coevolving residues, ones that mutate together across species, signal physical contact in the fold. A final structure module then places every atom directly, and the whole network is trained end to end from sequence to coordinates. At CASP14 in November 2020 this reached a median backbone accuracy around 2.1 angstroms and topped 90 on the global distance test for most targets, the level organisers treated as comparable to experiment.

David Baker's Rosetta works the energy landscape directly, scoring candidate structures with a physically grounded function and sampling toward the lowest-energy fold. Inverting that search turns prediction into design: fix a target backbone, then search sequence space for amino acids that make that backbone the energy minimum. The 2003 protein Top7, 93 residues with a topology absent from nature, validated the idea when its crystal structure matched the model to about 1.2 angstroms. Baker open-sourced Rosetta, and the field has since folded deep learning into design through tools such as RoseTTAFold, yielding designed binders, enzymes, nanomaterials, and SKYCovione, the first de novo protein approved as a human medicine.