Protein structure prediction from an MSA: a residual network predicts inter-residue distances and orientations that Rosetta folds into a 3D model.
No providers recorded yet. Browse all providers
By CASP13 in 2018, deep networks could read a multiple sequence alignment (MSA) and predict how far apart any two residues sit in the folded protein. A distance alone, though, says nothing about which way two residues face each other: two side chains 8 Å apart could point toward each other or away. trRosetta (transform-restrained Rosetta) closes that gap by predicting the full rigid-body relationship between every residue pair (a distance plus five angles) and handing all six to Rosetta as restraints for folding.
The method was developed by Jianyi Yang's group at Nankai University, now the Yang Lab at Shandong University, together with David Baker's lab at the Institute for Protein Design, University of Washington. It was published in PNAS in January 2020. The network is trained once on native PDB structures and then applied unchanged to new proteins, which is what let later work reuse it as a general sequence-to-structure oracle for design.
trRosetta is the direct predecessor of the Baker lab's three-track RoseTTAFold, and its name and inter-residue-geometry recipe carried over to the Yang Lab's RNA model trRosettaRNA, a separate model from a separate paper.
The network turns MSA-derived features (one-hot sequence, position-specific frequencies, entropy, 441 coupling channels from the inverted covariance matrix, and an APC-corrected score: 526 channels in total) into an L×L map. A 1×1 convolution projects this to 64 channels, followed by 61 dilated residual blocks (dilations cycling 1–16, 3×3 filters, ELU) and four softmax heads. Distance is binned at 0.5 Å from 2 to 20 Å; the angles are binned at 15°. Training used 15,051 non-redundant PDB chains (30% identity, X-ray ≤2.5 Å, released before 1 May 2018, each with at least 100 homologs), with MSAs built from Uniclust30 and a roughly 7-billion-sequence metagenomic database. The released predictor averages five networks trained on different 95/5 splits; each network has about 4.6 million parameters and took around nine days to train on one Titan RTX GPU.
On the 31 CASP13 free-modelling domains, trRosetta models averaged a TM-score of 0.625, against 0.587 for DeepMind's CASP13 AlphaFold (A7D) and 0.491 for Zhang-Server. On 131 hard CAMEO targets it averaged 0.621, 8.9% above Robetta. Top-L long-range contact precision on CASP13 FM was 51.9%. The authors stress that these comparisons are retrospective, whereas the CASP and CAMEO entries were blind. In ablations, adding orientations raised the CASP13 TM-score from 0.592 to 0.625.
trRosetta predicts structures for proteins that lack templates, given a deep enough MSA. The authors released the network weights, the training set, and the PyRosetta folding scripts, and the Yang Lab runs a public web server that has since added template detection, MSA input and a retrained network. Because the fixed network also scores single sequences, protein designers use it to check whether a designed sequence encodes its intended fold and to find the positions most sensitive to mutation. The authors also name cryo-EM model fitting as an application. TAPE ships a PyTorch port of the five-network ensemble; this is TAPE's own conversion, not the authors' release.
trRosetta brought orientation prediction and fast restraint-guided Rosetta folding into the post-CASP13 toolkit. Successor pipelines built on it, including the BAKER and Yang-Server entries, placed among the top groups at CASP14. It also turned structure predictors into design engines. Protein "hallucination" (Anishchenko et al., Nature 2021) ran Monte Carlo over sequences to maximize the contrast between trRosetta's predicted geometry and a background model, yielding new proteins whose solved structures matched the hallucinations. trDesign (Norn et al., PNAS 2021) optimized sequences against trRosetta's predicted landscape. Its geometric outputs became inputs to model-quality tools such as DeepAccNet, and TrMRF was built explicitly as its inverse. Within two years, end-to-end networks such as AlphaFold 2 and RoseTTAFold overtook it in accuracy.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.