Joint structure prediction for proteins, nucleic acids and ligands (AlphaFold 3)
A diffusion-based successor to AlphaFold predicts complexes spanning proteins, DNA, RNA, small molecules and ions in a single model, with large accuracy gains over the specialised docking tools used for each case.
- Model
- AlphaFold 3
- Field
- Biology
- Date
- 2024-05-08
- Human collaborators
- John Jumper, Max Jaderberg
Sources
Original work
Announcement
Challenge
What was found
AlphaFold 2 predicted the fold of a protein chain; drug discovery mostly needs the structure of that protein bound to something else. AlphaFold 3 replaces the structure module with a diffusion model operating directly on atom coordinates, which lets one network cover protein–ligand, protein–nucleic-acid and antibody–antigen complexes. Reported accuracy substantially exceeds specialised docking tools on protein–ligand benchmarks. Published in Nature.
Novelty check
Docking and complex prediction had decades of specialised tools (AutoDock, RoseTTAFold All-Atom, and AlphaFold-Multimer for protein assemblies). The contribution is one model spanning molecule types at higher accuracy, not a new biological result; the structures it predicts are natural ones.
Caveats and known objections
At publication neither code nor weights were released, only a rate-limited web server, which drew an open letter from over a thousand researchers about reproducibility; weights for academic use followed in November 2024. Accuracy on antibody–antigen complexes and some nucleic-acid targets is far below the headline figures, and the model can produce confidently wrong hallucinated structure in disordered regions. An independent benchmark of 2,600 protein-ligand systems released after the training cutoff found accuracy falls on complexes unlike the training data and concluded that co-folding methods including this one largely memorise ligand poses. Predictions are hypotheses, not measurements.
Independent checks
Škrinjar, Eberhardt, Studer, Tauriello, Schwede and Durairaj, University of Basel and SIB (Runs N' Poses benchmark): On 2,600 high-resolution protein-ligand systems released after the training cutoff, accuracy of AlphaFold 3 and three similar co-folding methods declines on complexes unlike the training data; the authors conclude that these methods largely memorise ligand poses. · link ↗
Disagree with these grades?
Bring a citation: a grade moves on evidence, not on argument.
Or on GitHub: submit a check challenge the grade send a correction or send a pull request
Flag this for triage
Signals order the review queue and nothing else. They are never published, and they never move a grade: that takes a citation.
History of this entry
- CorrectedCited the peer-reviewed version of the Runs N' Poses benchmark, published in Nature Structural & Molecular Biology in May 2026 with the same conclusion. No grade moved. source ↗
- ChallengedCited an independent benchmark finding that protein-ligand accuracy falls on complexes unlike the training data. No grade moved. source ↗
- AddedEntered the registry graded Peer reviewed and AI-led.
Entries are never deleted. A grade that does not hold up is downgraded on the record, with the reason beside it.
Community discussion
Graded peer reviewed for verification and ai-led for autonomy. What these mean.
Cite this entry
whataifound.org. (2024). Joint structure prediction for proteins, nucleic acids and ligands (AlphaFold 3). whataifound.org: A Registry of AI Scientific and Mathematical Discoveries. https://whataifound.org/finding/2024-05-08-alphafold3BibTeX
@misc{whataifound-googledeepmind-2024-alphafold3,
title = {Joint structure prediction for proteins, nucleic acids and ligands (AlphaFold 3)},
author = {{whataifound.org}},
year = {2024},
howpublished = {whataifound.org: A Registry of AI Scientific and Mathematical Discoveries},
note = {Result by Google DeepMind / Isomorphic Labs. Verification: Peer reviewed. Autonomy: AI-led.},
url = {https://whataifound.org/finding/2024-05-08-alphafold3}
}