How rosettacommons/rfdiffusion works
RFdiffusion generates protein structures through configurable diffusion inference: a user launches the inference script, which prepares a sampler and conditions, runs repeated denoising with the learned structure model, and writes designs. The README documents unconditional and motif-conditioned generation, symmetry, binder design, fold conditioning, and partial diffusion; these are represented as sampler capabilities, though their detailed implementations are not all shown. The graph also shows the RoseTTAFold model layers and an internal SE(3)-Transformer layer. Links are limited to README statements and source-index references; unsampled implementation details are not asserted.
Main entry points: model_runners.py Sampling runners · Track_module.py Track module · RoseTTAFoldModel.py RoseTTAFold model · Attention_module.py Attention layers
Components 22
Inference orchestration
- Inference script
scripts/run_inference.py - Inference utilities
rfdiffusion/inference/utils.py - Sampling runners
rfdiffusion/inference/model_runners.py
Design conditioning
- Contig mapping
rfdiffusion/contigs.py - Symmetry support
rfdiffusion/inference/symmetry.py - Potential manager
rfdiffusion/potentials/manager.py - Guidance potentials
rfdiffusion/potentials/potentials.py
Structure generation
- Diffusion process
rfdiffusion/diffusion.py - SO(3) schedules
rfdiffusion/igso3.py - Protein geometry
rfdiffusion/kinematics.py - Residue chemistry
rfdiffusion/chemical.py - Design outputs
rfdiffusion/util.py
Learned model
- RoseTTAFold model
rfdiffusion/RoseTTAFoldModel.py - Feature embeddings
rfdiffusion/Embeddings.py - Track module
rfdiffusion/Track_module.py - Attention layers
rfdiffusion/Attention_module.py - SE(3) network
rfdiffusion/SE3_network.py - Tensor utilities
rfdiffusion/util_module.py - SE(3) Transformer
env/SE3Transformer/se3_transformer/model/transformer.py - Fiber features
env/SE3Transformer/se3_transformer/model/fiber.py - Equivariant convolution
env/SE3Transformer/se3_transformer/model/layers/convolution.py
Other components
- Design user
Connections 22
- Design user → Inference script: launches
README.md - Inference script → Inference utilities: uses
scripts/run_inference.py - Inference utilities → Sampling runners: references
rfdiffusion/inference/utils.py - Sampling runners → Contig mapping: uses
rfdiffusion/inference/model_runners.py - Sampling runners → Symmetry support: uses
rfdiffusion/inference/model_runners.py - Sampling runners → Potential manager: configures
rfdiffusion/inference/model_runners.py - Potential manager → Guidance potentials: creates
rfdiffusion/potentials/manager.py - Sampling runners → Diffusion process: uses
rfdiffusion/inference/model_runners.py - Diffusion process → SO(3) schedules: uses
rfdiffusion/diffusion.py - Sampling runners → Protein geometry: uses
rfdiffusion/inference/model_runners.py - Sampling runners → Residue chemistry: uses
rfdiffusion/inference/model_runners.py - Sampling runners → RoseTTAFold model: runs model
rfdiffusion/inference/model_runners.py - Sampling runners → Design outputs: writes designs
rfdiffusion/inference/model_runners.py - RoseTTAFold model → Feature embeddings: uses
rfdiffusion/RoseTTAFoldModel.py - RoseTTAFold model → Track module: uses
rfdiffusion/RoseTTAFoldModel.py - Feature embeddings → Attention layers: uses
rfdiffusion/Embeddings.py - Track module → Attention layers: uses
rfdiffusion/Track_module.py - Track module → SE(3) network: uses
rfdiffusion/Track_module.py - Attention layers → Tensor utilities: uses
rfdiffusion/Attention_module.py - Track module → Tensor utilities: uses
rfdiffusion/Track_module.py - SE(3) Transformer → Fiber features: uses
env/SE3Transformer/se3_transformer/model/transformer.py - SE(3) Transformer → Equivariant convolution: uses
env/SE3Transformer/se3_transformer/model/transformer.py