AlphaFold 3 predicts how proteins interact with DNA, RNA and drugs
Google DeepMind and Isomorphic Labs published AlphaFold 3 in Nature on 8 May 2024. Where AlphaFold 2 predicted the shape of a protein, this predicts the joint structure of complexes containing proteins, nucleic acids, small molecules and ions, using a diffusion architecture that produces atom coordinates directly.
Background
AlphaFold 2 predicted the folded shape of a protein from its sequence, and that was the problem the field had been stuck on. But a protein alone does very little. It works by binding: to DNA, to RNA, to a drug molecule, to another protein.
Those interactions were handled by separate tools, each built for one kind of pairing, none of them as good at its job as AlphaFold 2 was at its own. RoseTTAFoldNA had extended the approach to nucleic acids in 2022.
What happened
AlphaFold 3 was published in Nature on 8 May 2024 by Google DeepMind and Isomorphic Labs. It predicts the joint structure of a complex containing proteins, nucleic acids, small molecules, ions and chemically modified residues, from a list of what is present.
The architecture changed with it. AlphaFold 2 reasoned about relationships between amino acids; AlphaFold 3 uses a diffusion model that produces atom coordinates directly, which is what allows one model to handle molecules that are not proteins at all.
The paper reports better accuracy than specialised docking tools for protein and drug binding, better than nucleic-acid-specific predictors for protein and DNA or RNA, and better than the earlier AlphaFold-Multimer for antibody and antigen.
What followed
The paper appeared without its code or its model weights. Researchers objected at once, and the objection was not about competitive advantage: a structure prediction nobody else can run is not something anyone else can check.
Access was through a web server, free for non-commercial use, which lets a scientist get an answer and not examine how it was reached.
In November the code was released for academic use, and Nature published an addendum in December recording it. A journal amending a paper because its methods had become available afterwards is an unusual thing to have to do, and it is the part of this entry most likely to matter in ten years.
Why it mattered
The paper appeared without its code or its weights, and the objection from other researchers was immediate: a result nobody can reproduce is not a result. Nature published an addendum in December recording that the inference code had since been released, which is an unusual thing for a journal to have to do and a useful precedent.
Sources
- Accurate structure prediction of biomolecular interactions with AlphaFold 3. nature.com. Primary source
- Addendum: Accurate structure prediction of biomolecular interactions with AlphaFold 3. nature.com. Primary source
- AlphaFold 3 predicts the structure and interactions of all of life's molecules. isomorphiclabs.com. Official
Cite this page
AI Achievements. (2024). AlphaFold 3 predicts how proteins interact with DNA, RNA and drugs. Retrieved 2026-08-29, from https://achievements.ai/milestone/alphafold-3-biomolecular-interactions
@misc{achievements_alphafold_3_biomolecular_interactions,
title = {AlphaFold 3 predicts how proteins interact with DNA, RNA and drugs},
author = {{AI Achievements}},
year = {2024},
url = {https://achievements.ai/milestone/alphafold-3-biomolecular-interactions}
}