Twenty amino acids in a chain fold into a unique three-dimensional ribbon, predicted by attention.
From the chapter: Chapter 17: Applications
Glossary: alphafold, alphafold 3
Transcript
A protein is a chain of amino acids. Twenty letters. Sequences are easy to read.
But proteins do not work as flat strings. They fold into precise three-dimensional shapes. Alpha helices, beta sheets, intricate ribbons. The shape determines the function.
For fifty years, predicting a protein's shape from its sequence was one of the hardest problems in computational biology. Experimental methods, crystallography, cryo-electron microscopy, took months to years per protein. Hundreds of millions of known sequences had no known structure.
In 2020, AlphaFold 2 changed this overnight. A neural network trained on the hundred and seventy thousand structures in the Protein Data Bank, plus sequence data from millions of related proteins.
Inside, attention heads pass information between residues, building up a representation of which pairs are likely to be close in three dimensions. A separate module turns those pair distances into actual coordinates.
Watch a sequence enter. Watch the attention pass information back and forth. Watch the predicted backbone curl into helices and sheets. The resulting ribbon matches the experimental structure to within an angstrom in most cases.
DeepMind released two hundred million predicted structures, essentially the entire known proteome. AlphaFold 3 extended the model to ligands, nucleic acids, and protein complexes. A 2024 Nobel Prize for chemistry recognised the work.
Drug discovery, enzyme design, and structural biology have not been the same since.