Skip to content

Biology's AI Moment

How AlphaFold Works

11 min+100 XPStep 1/11

Evolution keeps the lab notebook

AlphaFold2's key input isn't chemistry — it's ancestry. A multiple sequence alignment (MSA) lines the target protein up against thousands of relatives gathered across species.

Positions that co-vary — a mutation here tolerated only alongside a compensating mutation there — reveal contacts: those residues are probably touching in 3D. Earlier methods extracted such contacts with statistics; DeepMind's bet was that attention could read far richer patterns out of the same notebook.

insight

The data was there all along — billions of years of mutation-and-selection experiments, logged in genomes. AlphaFold treats evolution itself as a simulator it can learn from.