AlphaFold 3

AlphaFold 2 had predicted the shape of a protein on its own. Biology mostly happens when molecules meet, and on 8 May 2024 DeepMind and its drug-discovery subsidiary Isomorphic Labs published a model that predicted the joint structure of a protein with whatever it bound: another protein, a strand of DNA or RNA, a drug-like small molecule, an ion, a modified residue. Its accuracy on protein–ligand complexes was at least fifty percent better than the specialised tools, and it was the first single method to cover all the classes at once.

The architecture had changed. The evolutionary module was simplified, and the final structure was produced not by a geometric network but by diffusion, the same denoising process that generates images, applied to atomic coordinates. It was released as a web server for non-commercial use, with the code and weights following in November for academic researchers, a more restricted release than AlphaFold 2's that drew protest from the structural-biology community.

Five months later the Nobel committee gave half the chemistry prize to Jumper and Hassabis. AlphaFold 3 is on this timeline as the point at which the models moved from describing biology to being used to design it.