In July 2021 Science published “Accurate prediction of protein structures and interactions using a three-track neural network” by Minkyung Baek, David Baker, and colleagues at the University of Washington Institute for Protein Design. It appeared within days of DeepMind’s AlphaFold 2 paper and gave the open research community a high-performing structure predictor of its own.
The method, RoseTTAFold, uses a “three-track” architecture in which information at three levels is processed and exchanged simultaneously: the one-dimensional amino acid sequence, the two-dimensional map of distances between residue pairs, and the three-dimensional atomic coordinates. Letting these tracks inform each other lets the network reason about sequence, contacts, and geometry together rather than in separate stages.
RoseTTAFold produced structure predictions with accuracy approaching AlphaFold 2’s while remaining computationally efficient enough to run on a single graphics card. The paper also showed the network could generate models of protein-protein complexes directly from sequence, and the authors used its predictions to help solve real X-ray crystallography and cryo-electron microscopy structures. Code and a public server were released openly.
For a general reader, RoseTTAFold matters because it proved the AlphaFold breakthrough was reproducible and not a one-lab fluke, and because the Baker lab’s open release seeded a wave of follow-on tools for designing entirely new proteins. David Baker shared the 2024 Nobel Prize in Chemistry for this line of work.