Loading…
Protein structure and complex prediction from sequence, in a three-track network that reasons over alignments, distances, and 3D coordinates at once.
No providers recorded yet. Browse all providers
A two-track folding network reasons about sequence and about a residue-residue distance map together, then hands the finished distance map to a structure module that builds coordinates. The coordinates never talk back. RoseTTAFold's premise is that they should: it runs a third track, operating directly on 3D backbone coordinates, in parallel with the 1D alignment track and the 2D distance track, information flowing in both directions among all three throughout the network. The network can revise what it believes about a sequence position because of where that residue ended up in space, not only the other way around.
The architecture came out of the Baker Lab at the University of Washington Institute for Protein Design, with crystallographers, cryo-EM groups and enzymologists across a dozen institutions supplying the experimental test cases, and was published in Science in July 2021. Its context was CASP14, where AlphaFold 2 had demonstrated an accuracy jump but had not published its method. Baek and colleagues worked from the five architectural ideas described at the conference, built a two-track network that already beat trRosetta, then extended it to three tracks — and released trained weights and running code roughly two weeks before AlphaFold 2's own code appeared.
Two ways of turning the network's output into a structure ship together. The pyRosetta variant feeds predicted distance and orientation distributions into pyRosetta to build relaxed all-atom models on an 8 GB GPU; the end-to-end variant passes averaged features through a final SE(3)-equivariant layer and emits backbone coordinates directly. Because that path already handles chain breaks, feeding it two or more sequences yields a protein-protein complex.
The 1D track uses Performer-style attention over the multiple sequence alignment, the 2D track attends over the distance map, and the 3D track is built from SE(3)-equivariant transformer layers. GPU memory ruled out training on whole proteins, so the network saw discontinuous crops — two sequence segments totalling 260 residues — with 1D features and 2D distance and orientation predictions averaged across crops at inference. Alignments come from HHblits against UniRef30 and the 272 GB BFD, templates from HHsearch against a pdb100 database, and only the top 1000 alignment sequences are used. After roughly 1.5 hours of sequence and template search, the end-to-end version takes about 10 minutes on an RTX 2080 for proteins under 400 residues, and complexes around 30 minutes on a 24 GB TITAN RTX.
On CASP14 targets the three-track model beat both leading server groups and the second-ranked human group, while remaining short of AlphaFold 2, and on CAMEO's 69 medium and hard targets released between 15 May and 19 June 2021 it outscored every other server evaluated. Applied to 693 domains from human proteins lacking close structural homologs, over a third of the models reached predicted lDDT above 0.8 — corresponding to an average Cα-RMSD of 2.6 Å on CASP14 targets.
The paper's strongest evidence is experimental rescue. Four crystallographic datasets that had resisted molecular replacement with every PDB model available — among them a bacterial surface layer protein and the fungal secreted protein Lrbp — were phased using RoseTTAFold models, where trRosetta models failed on the same data. A prediction of the p101 Gβγ binding domain fit a low-resolution PI3Kγ cryo-EM map to 3.0 Å Cα-RMSD over its β-sheets. The authors also modelled every human GPCR of unknown structure in active and inactive states, and used models of TANGO2, the ADAM33 prodomain and ceramide synthase CERS1 to place pathogenic mutations against predicted active sites.
RoseTTAFold made high-accuracy folding available to any lab with a gaming GPU at a moment when it was otherwise unobtainable, and the Robetta server to labs with none. The code is MIT-licensed; the trained weights carry the Rosetta-DL agreement, non-commercial research only — a split that still governs much of the family. That family is substantial: RFdiffusion fine-tunes this network into a generative denoiser for de novo design, RoseTTAFold2 rebuilds the three-track architecture with AlphaFold 2 training ideas folded in, RoseTTAFold All-Atom extends it to nucleic acids and small molecules, and RF3 carries the lineage into all-atom diffusion. The original's limits are plain in hindsight: it was trained on monomers rather than complexes, and it depends on a slow alignment and template search.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.