How baker-laboratory/rf_diffusion_all_atom works
RFDiffusionAA generates all-atom protein structures by iteratively denoising a model state conditioned on an input structure, ligand and/or preserved protein motif, and design constraints. The principal workflow runs from inference configuration through sampler setup and reverse-diffusion steps to final and trajectory PDBs. README documents ligand-only and ligand-plus-protein-motif designs, reproducible runs, and multiple designs. The graph emphasizes these stages and the potential-guidance layer; unsampled subsystem wiring is omitted where it cannot be established from the excerpts.
Main entry points: model_runners.py Sampler runner · diffusion.py Reverse diffusion · pdbio.py Structure output · run_inference.py Inference entry
Components 17
Inference workflow
- Inference config
config/inference/aa.yaml - Inference entry
run_inference.py - Sampler runner
inference/model_runners.py
Design conditioning
- Target structure
input/7v11.pdb - Contig mapping
contigs.py - All-atom state
aa_model.py - Chemical definitions
chemical.py
Structure generation
- Reverse diffusion
diffusion.py - Geometry features
kinematics.py - Rotation transforms
rotation_conversions.py
Guidance and output
- Potential manager
potentials/manager.py - Contact guidance
potentials/potentials.py - Structure output
pdbio.py - Design structure
- Denoising trajectories
Other components
- Design operator
- Model weights
Connections 16
- Design operator → Inference entry: sets design
- Inference config → Inference entry: configures
- Inference entry → Sampler runner: selects sampler
- Target structure → Sampler runner: conditions sampler
- Sampler runner → Contig mapping: maps design
- Sampler runner → All-atom state: builds state
- All-atom state → Chemical definitions: uses definitions
- Sampler runner → Model weights: loads weights
- Sampler runner → Reverse diffusion: runs denoising
- Sampler runner → Potential manager: initializes guidance
- Potential manager → Contact guidance: instantiates
- Sampler runner → Structure output: writes structures
- Structure output → Design structure: writes final
- Structure output → Denoising trajectories: writes trajectories
- Reverse diffusion → Geometry features: uses geometry
- Reverse diffusion → Rotation transforms: uses rotations