Chroma
Programmable protein diffusion model steered by exact geometric, symmetry, and energy-potential conditioners and wet-lab validation.
The loop
- Proposes
A polymer-physics-informed graph diffusion model samples 3D protein backbones and all-atom sequences in O(N) time.
- Judged by
Analytic point-group symmetry projectors, distance restraints, AlphaFold refolding consistency, and X-ray crystallography judge candidate proteins.
- Steers
Bayes-rule score conditioning steers reverse-diffusion trajectories directly toward the specified symmetry or substructure constraints.
Headline result
Designed 310 proteins across point-group symmetries, programmable shapes, and protein classes, with crystallographic structures matching design to 1.0-1.4 Angstrom backbone RMSD.
What this does not prove
Conditioning on extreme geometric shapes can reduce natural sequence designability if the score guidance weight is set too high.
How it generalizes the harness
Analytic symmetry constraints and structural refolding oracles steer a non-LLM 3D protein diffusion process.
Classification
| What it proposes | Protein structure |
|---|---|
| Generative model | Polymer-graph SE(3) diffusion model |
| Algorithmic authority | Geometric symmetry projectors, refolding filters, and X-ray crystallography |
| Feedback | Score |
| What gets updated | Population of candidates |
| When it runs | At inference time |
| Validity | Approximate physics (simulation, not lab results) |
| Evidence | Experiments reported |
| Lineage | Original system |