15 July 2021

RoseTTAFold reproduces protein structure prediction and releases the code

Baek and colleagues at the Institute for Protein Design published RoseTTAFold in Science on 15 July 2021, having rebuilt protein structure prediction from the ideas DeepMind had described at CASP14 eight months earlier without releasing. Its accuracy approached AlphaFold 2 without matching it, and it ran in about ten minutes on a gaming computer.

Biology Capability threshold Independently validated

Background

DeepMind presented AlphaFold 2 at CASP14 in November 2020 and the result was not in doubt: the structure prediction problem, open since Anfinsen’s Nobel in 1972 and assessed biennially since 1994, had been largely solved.

What DeepMind released was a ten-minute presentation. No code, no weights, no paper. For eight months the field knew the problem was solvable and had no way to solve it.

What happened

Baek and colleagues at the Institute for Protein Design published RoseTTAFold in Science on 15 July 2021, built from the ideas described at the conference.

Its architecture is a three-track network: the sequence, the map of distances between amino acids, and the three-dimensional coordinates, each processed in its own track with information passing between them, so that the network reasons about all three at once.

The paper is explicit that it did not match AlphaFold 2 on the CASP14 targets. It came close, it beat everything else, and it says plainly that DeepMind’s model was better.

What it had instead was availability. The code went on GitHub, a web server opened, and the Institute reports over 4,500 proteins submitted in the first month. A structure takes about ten minutes on a single gaming computer.

What followed

AlphaFold 2’s own paper appeared in Nature the same week. The two events are close enough that people have drawn a line between them, and the connection is not documented, so this entry notes the timing and leaves it there.

Baek went on to build RoseTTAFoldNA for nucleic acids, and the same network was later fine-tuned into RFdiffusion, which designs proteins rather than predicting them. David Baker shared the 2024 Nobel Prize in Chemistry for that line of work.

Why it mattered

DeepMind had won CASP14 in November 2020 and published nothing usable, so for eight months the field could see that the problem was solved and could not solve it. The Baker Lab released working code, and AlphaFold 2's own paper appeared in the same week.

Sources

Accurate prediction of protein structures and interactions using a three-track neural network. science.org. Primary source
RoseTTAFold: Accurate protein structure prediction accessible to all. ipd.uw.edu. Official
Accurate prediction of protein structures and interactions using a 3-track neural network. ncbi.nlm.nih.gov. Primary source

Cite this page

AI Achievements. (2021). RoseTTAFold reproduces protein structure prediction and releases the code. Retrieved 2026-08-29, from https://achievements.ai/milestone/rosettafold-protein-structure-prediction

@misc{achievements_rosettafold_protein_structure_prediction,
  title  = {RoseTTAFold reproduces protein structure prediction and releases the code},
  author = {{AI Achievements}},
  year   = {2021},
  url    = {https://achievements.ai/milestone/rosettafold-protein-structure-prediction}
}