What it is
AlphaFold 2 (DeepMind, Nature, 2021) predicted single-chain protein structures at useful accuracy; the AlphaFold Database put hundreds of millions of predictions in public hands. AlphaFold 3 (2024) added more complexes and biomolecules. ESMFold (Meta) and RoseTTAFold / RFdiffusion (Baker lab) sit in the same toolbox. This is deployed science software, not a chatbot.
Problem it targets
Enzymes that eat plastic, fix nitrogen more cleanly, or capture carbon are protein-design problems. Knowing a fold is not the same as a stable, cheap, non-toxic industrial enzyme — but it removes a years-long bottleneck.
How it works
The model learned from the Protein Data Bank and evolutionary pairings. Laboratories still express, purify, and assay. RFdiffusion and related tools generate new backbones; wet-lab failure remains common.
Status and players
Google DeepMind and EMBL-EBI (AF Database); Meta ESM; University of Washington Institute for Protein Design (Baker); industrial biotech using the tools on enzymes and drugs.
Risks and limits
Biosecurity (the same tools can sketch harmful proteins — labs and journals now discuss screening). Over-claiming a predicted PETase as a recycling plant. Energy and water of the fermenter still count.
Sources
Jumper et al., Nature, 2021 (AlphaFold 2); AlphaFold Database (EMBL-EBI); Abramson et al., Nature, 2024 (AlphaFold 3); Lin et al., Science, 2023 (ESMFold); Watson et al., Nature, 2023 (RFdiffusion).