AlphaFold 3
DeepMind trained a system to predict the structure not of a single protein but of a complex: protein with DNA, with a drug, with ions, all in one model.
Why it matters
Prediction moved from the shape of one molecule to how molecules hold on to each other, which is what the chemistry of life consists of.
AlphaFold 2 gave the shape of a protein. But a cell works through interactions: a protein latches onto DNA, onto a small molecule, onto another protein. AlphaFold 3 replaced the specialised geometric module with a diffusion module that generates atomic coordinates directly, which let it describe every type of molecule the same way. For protein-drug interactions the accuracy beat specialised docking methods. The model was not released openly but through a web server with restrictions, which started a separate argument about reproducibility.