MODEL LEARNING PATH
RFdiffusion
Original RFdiffusion generates protein backbones by denoising residue positions and orientations. Sequence/pair reasoning and an \(\mathrm{SE}(3)\)-equivariant structure update predict a cleaner backbone at each step.
Loading the interactive page…
The original Nature 2023 study, preceded by a 2022 preprint, joined a pretrained structure predictor to generative diffusion and experimentally tested diverse designed proteins. Later catalytic-motif scaffolding methods explicitly build on RFdiffusion. This landmark page is about that original backbone model, not RFdiffusion2 or RFdiffusion3.
Chapters
Specify and perturb a backbone
Design constraints distinguish fixed motifs from movable residues; translations and orientations form the noisy state.
Distinguish fixed design constraints from noisy movable backbone frames.
Build sequence and pair context
Input geometry feeds sequence–pair exchange and geometry-biased tied axial attention.
Trace noisy geometry through sequence–pair exchange and pair attention.
Predict a cleaner backbone
An SE(3)-equivariant update refines residue frames and reconstructs backbone coordinates.
Apply a global frame update and reconstruct a cleaner backbone state.
Advance and feed back
The reverse step changes the noisy state; the previous clean prediction supplies template geometry for the next network call.
Distinguish reverse diffusion from recycling a previous clean prediction.
All exercises in this guide
- Variance-Preserving Forward Noising
- Masked Sum
- Masked Mean
- Stable vector norms, normalization, dot and cross features
- Batched rigid transforms
- Right-handed local frame from three points
- Stable SO(3) exponential and logarithm maps
- SO(3) heat-kernel scores and supplied-variate angle sampling
- Masked pairwise distances and hard distance-bin targets
- RFdiffusion Template Pair Features
- Affine Projection
- Masked Softmax
- Template Pair Cross-Attention
- Template geometry and pair features
- Multi-head attention with masks and additive bias
- Template-axis attention and mean aggregation
- Layer Normalization
- Sequence-Weighted MSA Attention
- MSA column attention
- Masked outer-product mean
- Global-column MSA attention with shared keys and values
- Tied Pair Row Attention
- Tied Pair Column Attention
- Feedforward Block
- ReLU
- RoseTTAFold Pair Block
- RFdiffusion tied axial and sequence-weighted attention
- Global Quaternion Frame Update
- RFdiffusion Full Graph
- RFdiffusion Refinement Graph
- Normalized-quaternion backbone updates
- Masked batched kNN graphs and neighbor gather
- Ragged segment sum and mean
- Typed radius, bond-prefix and sequence-neighbor graph assembly
- Real spherical harmonics from Cartesian polynomial tables
- Real-irrep Clebsch–Gordan paths and channel tensor products
- Origin-safe Bessel channels with a compact polynomial cutoff
- Radial-weighted irrep convolution with receiver aggregation
- Learned irrep query-key attention and tensor values
- Linear pLDDT Head
- pLDDT Confidence Summary
- Gaussian forward corruption and score targets
- DDPM posterior from a clean-coordinate prediction
- RFdiffusion rotational reverse step
- Clean Template Recycling
- Recycling representations and clean-template state
- Gaussian Noising and Score
- Distogram Linear Head
- Learned structural confidence heads and model-aware ranking