RoseTTAFold3 (RF3)

RF3 is an all-atom biomolecular structure prediction network from the Institute for Protein Design at the University of Washington that predicts the structures of arbitrary complexes containing proteins, DNA, RNA, small molecules, ions, and covalent modifications. It is the successor to RoseTTAFold All-Atom and adopts the diffusion-based design established by AlphaFold 3, with the explicit goal of closing the accuracy gap between that closed-source system and the open-source alternatives available to the community.

The model is inseparable from AtomWorks, the data framework it was built on. Preparing Protein Data Bank structures for training — resolving leaving groups, bond orders, charges, covalent geometries, alternate occupancies, ligands at symmetry centers — has historically consumed months of engineering per project and locked that expertise inside a handful of labs. AtomWorks turns those steps into composable transforms over a shared atom-level representation built on Biotite's AtomArray, so that more than 80% of code is reused across the networks trained with it, including RF3, RoseTTAFold All-Atom, ProteinMPNN, and LigandMPNN. Rebuilding LigandMPNN's pipeline replaced over 2,000 lines of bespoke code with a 100-line declarative specification.

Both packages were released together in August 2025 under a permissive BSD 3-Clause license, covering the framework, the training and curation code, and the RF3 checkpoints — an unusually complete open release for a model of this class. The curated training corpus itself is not distributed; it is reproduced by running AtomWorks over public structural databases.

#Key Features

• Arbitrary biomolecular complexes: A single network predicts proteins, protein-protein interfaces, protein-ligand complexes, nucleic acids, and covalently modified residues, taking sequences, SMILES, CCD codes, and SDF/CIF ligand definitions as input.
• Learned stereochemistry: Chirality is encoded as the sign of the angles around each chiral center, with the gradient of the ideal-angle error supplied at every denoising step. RF3 gets 88% of ligand chiral centers correct without inference-time guidance, against 84% for AlphaFold 3 and 76% for Boltz-2.
• Atom-level conditioning: Users can specify pairwise distances between arbitrary atoms to encode experimental restraints, dock against a known receptor, or fold a protein around a fixed ligand conformer. Templating a rigid ligand lifts median protein-ligand interface accuracy from 0.821 to 0.882, and holo distance restraints raise it to 0.890.
• Disorder-aware training: A distillation set generated with Rosetta supplies extended backbones for disordered regions, so the network models unresolved segments as extended rather than hallucinating secondary structure.
• Permissive code and weight release: Framework, data-curation code, training and inference code, and weights are distributed under BSD 3-Clause, with checkpoints for both a 09/2021 and a 01/2024 training cutoff to support leak-free benchmarking.

#Technical Details

RF3 pairs a Pairformer trunk with a diffusion module and a separately trained confidence head. Inputs include the polymer sequence, protein and RNA multiple sequence alignments, atom-level templates, per-residue RDKit reference conformers, and chiral features. Training proceeded in three stages: 384-token crops, then 768-token crops with an added polymer-nonpolymer bond loss, then the same configuration with the date cutoff extended to January 2024, using an exponential moving average of weights with 0.999 decay. Beyond the PDB, training draws on new distillation sets, including 27K protein-nucleic acid complexes and 10K RNA structures. On a low-homology test set of PDB entries released after January 2024, the 01/2024 checkpoint reaches median interface lDDT of 0.607 for protein-protein, 0.798 for protein-ligand, 0.523 for protein-DNA, and 0.772 for RNA-only predictions. On a de-leaked antibody-antigen set, 33% of RF3 predictions reach DockQ > 0.23, compared with 44% for AlphaFold 3, 28% for Chai-1, and 22% for Boltz-2. Across categories, accuracy sits between closed-source AlphaFold 3 and the open-source Boltz models.

#Applications

RF3 serves the routine structural biology workload: modeling protein complexes, docking small molecules into binding sites, predicting nucleic acid assemblies, and generating antibody-antigen models for therapeutic triage. Its chirality handling makes it particularly suited to mixed L/D and macrocyclic peptides, a therapeutic class where diffusion models have historically required inference-time guidance to respect stereochemistry. The atom-level conditioning interface lets structural biologists fold a target around a crystallographic ligand pose or impose crosslinking and NMR restraints. AtomWorks itself is the broader deliverable for method developers, supporting structure prediction, generative design, and fixed-backbone sequence design within one pipeline; RF3 ships from the same RosettaCommons Foundry repository as RFdiffusion3.

#Impact

RF3 raises the floor for open-source structure prediction, giving academic and commercial groups a permissively licensed all-atom predictor whose code and weights are fully inspectable, and whose training set can be regenerated from public databases with the released curation code. Its more durable contribution may be AtomWorks: by making PDB curation a shared, tested, documented library rather than per-project tribal knowledge, it removes the largest practical barrier to training new biomolecular models. The work is a preprint and has not completed peer review, and RF3 still trails AlphaFold 3 on most interface categories — the gap is narrowed, not closed.