AlphaFold 3 (Google DeepMind and Isomorphic Labs)
2024-05-08 · Publisher : Google DeepMind · Domain : Artificial Intelligence
Google DeepMind and Isomorphic Labs publish AlphaFold 3 in Nature. It predicts the structure and interactions of nearly all of life's molecules: proteins, DNA, RNA and ligands. A free server opens it to researchers.
What changes
- Beyond proteins: AlphaFold 3 also models DNA, RNA, ligands and chemical modifications.
- Accuracy at least 50% better than existing methods for interactions between proteins and other molecules, doubled in some categories.
- The free AlphaFold Server lets researchers use it without machine learning expertise, for non-commercial use.
- Code and weights were released in November 2024 for academic research, after criticism of the lack of code at publication.
- Isomorphic Labs uses it for drug design with pharmaceutical partners.
On May 8, 2024, Google DeepMind and Isomorphic Labs, its drug discovery subsidiary, published AlphaFold 3 in Nature. Whereas AlphaFold 2 predicted the shape of proteins, this version models the structure and interactions of most of life's molecules. ### What happened - AlphaFold 3 predicts the structure of complexes combining proteins, DNA, RNA, small molecules (ligands) and chemical modifications. - Its architecture relies on a diffusion module, close to image generation techniques, that directly assembles atomic coordinates. - Google DeepMind simultaneously launched AlphaFold Server, a free web interface for non-commercial research. - In November 2024, the model's code and weights were released for academic use. ### Key figures | Item | Value (Google DeepMind claim) | |---|---| | Interactions between proteins and other molecules | at least 50% more accurate than existing methods | | Some interaction categories | accuracy doubled | | PoseBusters benchmark (ligand docking) | 50% more accurate than traditional methods | | AlphaFold database (from AlphaFold 2) | 200 million protein structures | ### Why it is a turning point Drugs work by binding to proteins, often in contact with DNA or RNA. Predicting these assemblies in minutes rather than months of crystallography speeds up both basic research and the design of therapeutic molecules. A few months later, in October 2024, Demis Hassabis and John Jumper received the Nobel Prize in Chemistry for protein structure prediction. ### Caveats At publication, the paper was not accompanied by code, contrary to Nature's usual practice, which prompted an open letter from researchers. The server caps daily predictions and excludes commercial use. Like any predictive model, AlphaFold 3 can produce plausible but wrong structures, especially for disordered regions. ### Key takeaways AlphaFold 3 extends AI-assisted structural biology from proteins to molecular interactions as a whole, with a direct outlet in drug discovery.
Links
Tags : IA, Science, Biologie, Google DeepMind, AlphaFold