How to Use RoseTTAFold3

Open-source all-atom foundation model for structure prediction and generative design.

Key run controls include MSA Mode, Diffusion Steps, Diffusion Batch Size, and Number Recycles, making protocol choices explicit and reproducible for collaborative projects.

The 1.0 results page exposes analysis sections like Results Overview, predicted Local Distance Difference Test (pLDDT), and Chain Pair PAE (Summary), output artifacts such as rank_1.cif and scores.json, and interactive views including protein viewer molstar and select prediction for rapid model prioritization and export-ready reporting.

RosettaFold-3 (RF3) is a unified biomolecular foundation model developed using the AtomWorks framework. It is designed to bridge the gap between open-source models and AlphaFold 3, offering state-of-the-art accuracy in predicting arbitrary biomolecular complexes, including proteins, nucleic acids, small molecules, and mixed L/D peptides. RF3 distinguishes itself with improved handling of chirality through geometric conditioning and the ability to accept arbitrary atom-level constraints for tasks like docking and conformer templating.

Core capabilities include Built on AtomWorks, a modular framework for rapid prototyping and high-quality data processing., Explicit geometric features for chirality ensure 88% accuracy on ligand chiral centers., and Arbitrary atom-level conditioning allows user-specified distance constraints and ligand conformers., which directly shape how outputs should be interpreted for this method.

In practice, interpretation proceeds through inputs (Input Sequences, Input Molecules, and Residue Modifications), result sections (Results Overview, predicted Local Distance Difference Test (pLDDT), and Chain Pair PAE (Summary)), files (rank_1.cif and scores.json), and visual components (protein viewer molstar and select prediction), which makes RoseTTAFold3 outputs easier to triage and act on across large job batches.

• Study-fit inputs: RoseTTAFold3 accepts Input Sequences, Input Molecules, Residue Modifications, and Custom MSA, reducing preprocessing friction and preserving experimental context.
• Protocol control: Researchers can tune MSA Mode, Diffusion Steps, Diffusion Batch Size, and Number Recycles to match assay constraints, confidence thresholds, and downstream validation plans.
• Readable evidence: Results are presented through Results Overview, predicted Local Distance Difference Test (pLDDT), and Chain Pair PAE (Summary), rank_1.cif and scores.json, and protein viewer molstar and select prediction, improving cross-run comparison and scientific communication.
• Faster iteration: Managed execution on Neurosnap removes infrastructure overhead so teams can focus on structure prediction and confidence-guided model ranking rather than deployment and environment maintenance.