Use RosettaFold All-Atom Online
Commercially Available RosettaFold All-Atom No-Code Web Server
RoseTTAFold All-Atom: Predicting Protein, Nucleic Acid, Small Molecule and Metal Complexes
RoseTTAFold All-Atom (RFAA) is a deep network that models full biological assemblies containing proteins, nucleic acids, small molecules, metals and covalent modifications, given the sequences of the polymers and the bonded geometry of the small molecules and modifications. It was developed by Rohith Krishna, Jue Wang, Woody Ahern and colleagues in David Baker's group at the University of Washington Institute for Protein Design, with collaborators at the University of Sheffield and Seoul National University.
AlphaFold2 and RoseTTAFold transformed protein structure prediction, but they cannot model covalent modifications or interactions with small molecules and other non-protein partners that carry out key biological functions. A common workaround is to predict the protein first and then add non-protein components with classical docking, but systematically evaluating and optimizing such multi-step procedures is not straightforward.
RFAA combines a sequence-based description of proteins and nucleic acids with an atom-bond graph for small molecules and modifications, so a single model can predict all atoms of an assembly. It also outputs per-residue and pairwise confidence estimates (pLDDT and PAE) so users can identify higher-quality predictions.
Property | Detail |
|---|---|
Components modeled | Proteins, DNA, RNA, small molecules, metals, covalent modifications |
Inputs | Polymer sequences plus atomic bond geometry of small molecules and modifications |
Architecture basis | RoseTTAFold2 with 46 added element-type tokens and atom-bond graph inputs |
Training data | 121,800 protein-small molecule structures, 112,546 protein-metal complexes and 12,689 covalently modified protein structures from the PDB |
Confidence outputs | pLDDT and PAE, including an interaction PAE between chains |
How RoseTTAFold All-Atom Represents Small Molecules and Modifications
Graph inputs: Small molecules are atom-bond graphs. The element type of each atom enters the 1D track, bond types (single, double, triple, aromatic) the 2D track, and chirality the 3D track.
Permutation invariance: Atoms carry no sequence-position encoding, so predictions do not depend on atom order.
Disconnected start: The system begins as a disconnected gas of residues, bases and free atoms that the network transforms into a plausible assembly.
All-atom loss: An all-atom version of the Frame Aligned Point Error loss guides training.
Atomization: Short stretches of three to five residues are randomly given as atoms during training, helping the model learn general atomic interactions.
RFAA Accuracy for Ligand Docking, Covalent Modifications and Generalization
RFAA was enrolled in the blind CAMEO ligand docking evaluation. Using interaction PAE below 10 as a confidence filter, 43% of CAMEO targets were predicted with high confidence, and 77% of those had ligand RMSD under 2 Angstrom.
Benchmark | Result |
|---|---|
CAMEO, cases modeled by both servers | RFAA 32% success (ligand RMSD under 2 Angstrom) vs 8% for the AutoDock Vina server |
Docking benchmark with privileged protein input | RFAA 42% vs DiffDock 38%; AutoDock Vina 52% when given the bound protein and pocket residues |
Covalent modifications (931 recent PDB entries) | 46% accurate (modification RMSD under 2.5 Angstrom) |
Glycans | Median RMSD of 3.2 Angstrom; chains up to seven monosaccharides |
Protein structure vs AlphaFold2 | Median GDT 85 vs 86 |
Novel protein clusters vs seen clusters | 41% vs 23% success rate |
The authors also show predictions for assemblies with three or more components, such as a fatty acid decarboxylase with heme and a lipid substrate, and a DNA polymerase with bound DNA, a nucleotide analog and a magnesium ion.
What is Tamarind Bio?
Tamarind Bio is a no-code bioinformatics platform built to give life scientists and researchers access to powerful computational tools. Many cutting-edge machine learning models are hard to deploy and use. Tamarind provides an intuitive, web-based environment that removes the complexity of high-performance computing, software dependencies and command-line interfaces.
The platform is designed for biologists, chemists and other researchers who may not have a background in programming or cloud infrastructure but want to run models on their own data. Key features include:
A user-friendly graphical interface for setting up and launching experiments
A robust API for integration into existing research pipelines
An automated system for managing and scaling computational resources
Tamarind treats information and data security as a top priority, as detailed in its Trust Center and Terms of Service.
Accelerating Discovery with RoseTTAFold All-Atom on Tamarind Bio
Protein-ligand docking: Predict the protein and ligand pose together from sequence and ligand chemistry, allowing a flexible target.
Multi-component complexes: Model proteins with several non-protein partners, such as cofactors, metal ions and nucleic acids, in one pass.
Covalent modifications: Predict covalently attached cofactors, drugs and glycans.
Confidence-based triage: Use interaction PAE and pLDDT to decide which predictions to trust before experiments.
How to Use RoseTTAFold All-Atom on Tamarind Bio
Open the tool: Log in at tamarind.bio and open RoseTTAFold All-Atom.
Enter polymer chains: Provide sequences for protein, DNA or RNA chains, as the tool allows.
Add small molecules and metals: Describe each ligand by its chemical structure so its atom-bond graph can be built, if exposed.
Specify covalent links: For modified residues or attached ligands, define the bonded atoms, if exposed.
Run: Launch the prediction.
Download outputs: Retrieve predicted coordinates and confidence metrics.
Check confidence: Review pLDDT and the interaction PAE; the paper treats interaction PAE below 10 as high confidence.
Use downstream: Carry confident models into analysis, docking refinement or design.
Things to Keep in Mind
Not all predictions are accurate; the most common failure is a ligand in the correct pocket but wrong orientation.
Accuracy is higher for proteins and ligands similar to the training data (for ligands dissimilar to training, 20% vs 16% success in the paper's comparison).
Physics-based docking with the bound protein structure given outperformed RFAA on that easier task.
The authors identify the small number of protein-small molecule training structures as the main limit on accuracy.
Source: Krishna R, Wang J, Ahern W, Sturmfels P, Venkatesh P, Kalvet I, Lee GR, Morey-Burrows FS, Anishchenko I, Humphreys IR, McHugh R, Vafeados D, Li X, Sutherland GA, Hitchcock A, Hunter CN, Baek M, DiMaio F, Baker D. Generalized Biomolecular Modeling and Design with RoseTTAFold All-Atom. bioRxiv preprint, posted October 9, 2023 (not peer reviewed in this version). DOI: 10.1101/2023.10.09.561603.