Contribution
One model for the joint structure of proteins, nucleic acids, small molecules, ions and modifications, beating specialist tools in each category tested.
Key insight
Generate atoms directly with diffusion, and drop the geometric machinery. Multi-scale denoising learns local stereochemistry at low noise and global arrangement at high noise, without frames, torsions or equivariant layers.
Key results
- PoseBusters: beats Vina (P = 2.27×10⁻¹³), which gets the pocket, and RoseTTAFold All-Atom (P = 4.45×10⁻²⁵)
- Protein–nucleic acid complexes: better than RoseTTAFold2NA
- Antibody–antigen quality rises with the number of seeds ranked by ipTM
- Cross-distillation from AlphaFold-Multimer “greatly reduced” hallucination (in Extended Data, not read)
Limitations
- 4.4% chirality violations, and clashes in very large complexes
- Static structures. Seeds do not sample the solution ensemble
- Hallucinated regions are low-confidence but not visibly disordered
- Restricted release, and the key comparisons exist only as figures