AlphaFold and the End of the Protein Folding Problem

The 2021 Nature paper describes a system that folds physical and biological priors into an attention-based network rather than simulating physics. Its input is not just the target sequence but a multiple sequence alignment of evolutionary relatives: residues that mutate together across species tend to touch in the folded structure, a signal the network learns to exploit. Two representations — one over the alignment, one over pairs of residues — exchange information through repeated attention blocks, then a structure module places every atom in 3D and the whole output is fed back through the network several times to refine itself. Every residue comes with a confidence score, so the model tells you where not to trust it. DeepMind open-sourced the code and, with EMBL-EBI, built a public database: by July 2022 it had grown from under a million predicted structures to more than 200 million, against about 190,000 experimentally solved structures in the Protein Data Bank, and had been used by more than half a million researchers. The 2024 Nobel Prize in Chemistry went to Demis Hassabis and John Jumper for the prediction work and to David Baker for computational protein design. The honest caveats still stand: a prediction is a single static shape, not the dynamics or complexes that biology runs on, and confidence drops on proteins with few evolutionary relatives.