RoseTTAFold All-Atom
B+UW Baker Lab
Structure PredictionComplex PredictionSmall Molecule Bindingopen
Updated 1 month agoNextIn take
RoseTTAFold-AA is betting that the future is unified all-atom models, and they're probably right — the question is whether the Baker Lab or the Boltz team gets there first with production-grade accuracy.
Specifications
| Architecture | Three-track neural network extended to all-atom representation including small molecules and ions |
| Parameters | ~100M |
| Training Data | PDB structures including protein-ligand and protein-nucleic acid complexes |
| License | BSD license. Fully open weights and code from the Baker Lab. |
| Hardware | 1x A100 40GB for most predictions |
| Inference Cost | Self-hosted — ~$0.02/prediction on cloud GPU |
| API Available | No |
| Weights Available | Yes |
Benchmark Performance
| Benchmark | Score |
|---|---|
| CASP15 | 0.79 GDT-TS (avg) |
| PoseBusters | 52% valid |
When to Use This
- +Predictions involving non-standard molecules (cofactors, ions, modified residues)
- +When you need an all-atom open-source alternative to AF3
When NOT to Use This
- −Pure protein structure prediction — Boltz-1 is more accurate
- −High-accuracy ligand docking — Boltz-2 or AF3 are better
Production Readiness
research
Known Users
- Baker Lab collaborators
- Academic structural biology groups
Grade Rationale
B+
The most ambitious open model — attempting to handle proteins, ligands, nucleic acids, and ions in one framework. The all-atom approach is the right long-term architecture, but accuracy still lags specialized models on each individual task.
Sources
Update History
2026-02-15Initial entry