Alterlab Rfdiffusion

RFdiffusion (de-novo backbone generation)

Overview

RFdiffusion (Watson et al., Nature 2023; RosettaCommons/RFdiffusion) is a diffusion model that generates protein backbones — new 3D structures, not sequences. It supports unconditional generation, motif scaffolding (build a fold around a fixed functional motif), binder design (generate a backbone that binds a target surface), and symmetric assemblies. It is the structure-generation step that starts the de-novo design pipeline; alterlab-proteinmpnn then designs sequences for the backbone and alterlab-alphafold validates them.

When to Use This Skill

Use this skill when the user wants to:

• Generate a novel protein backbone from scratch (unconditional).
• Scaffold a functional motif (e.g. a binding loop / catalytic geometry) into a new fold.
• Design a binder backbone against a given target protein surface / hotspots.
• Build symmetric oligomers (cyclic/dihedral) as backbones.

Does NOT Trigger

Core Capabilities

1. Unconditional generation

bashCopy code
# RosettaCommons/RFdiffusion — run from the repo root. RFD points at the repo's
# scripts directory, where run_inference.py lives; it is driven by a Hydra config.
RFD=./scripts
python "$RFD"/run_inference.py \
  'contigmap.contigs=[100-100]' \
  inference.output_prefix=out/uncond \
  inference.num_designs=10

contigmap.contigs specifies what to build (here, a 100-residue monomer). Quote the contig string — the brackets and / are shell metacharacters. Outputs backbone PDBs with no sequence, plus a .trb metadata file recording which output residues came from the motif.

2. Motif scaffolding

Fix a functional motif (residues from an input PDB) and let RFdiffusion build a supporting fold around it — the way to transplant a binding/catalytic geometry into a new, stable scaffold. The contig string mixes generated lengths with fixed ranges named by chain+index:

bashCopy code
python "$RFD"/run_inference.py inference.input_pdb=motif.pdb \
  'contigmap.contigs=[5-15/A10-25/30-40]' contigmap.length=55-55 \
  inference.output_prefix=out/scaffold inference.num_designs=10

Here 5-15 and 30-40 are generated segments resampled per design, A10-25 is the fixed motif, and /0 (with the trailing space) would start a new chain. Only the residues named in the contig are shown to the model — other chains and residues in the input PDB are ignored, so you do not need to pre-trim the file.

3. Binder design

Provide a target structure and hotspot residues; RFdiffusion generates binder backbones docked against that surface. Hotspots are the target residues the binder must contact:

bashCopy code
python "$RFD"/run_inference.py inference.input_pdb=target.pdb \
  'contigmap.contigs=[B1-100/0 100-100]' 'ppi.hotspot_res=[A30,A33,A34]' \
  inference.output_prefix=out/binder inference.num_designs=10

Follow with sequence design (alterlab-proteinmpnn) and an interface validation refold (alterlab-alphafold, read ipTM).

4. Which RFdiffusion (2026-09)

Three generations coexist; they are separate codebases, not upgrades of one CLI.

RFD3 replaces the Hydra/contig interface with a JSON or YAML input specification, so RFdiffusion-1 command lines do not carry over. Reach for it when the design target involves non-protein atoms or per-atom constraints; stay on RFdiffusion 1 for ordinary backbone generation, where its recipes and published benchmarks are what the field has replicated.

5. The full design → fold → score loop

1. Generate backbones here (RFdiffusion).
2. Design sequences with alterlab-proteinmpnn (or alterlab-ligandmpnn if a ligand is present).
3. Score by refolding with alterlab-alphafold and keeping only self-consistent designs.

GPU-heavy — dispatch generation and the fold sweep via alterlab-remote-compute.

Resources

• references/rfdiffusion_usage.md — install/pinning, contig grammar, motif/binder/symmetry configs, and loop integration. Loaded on demand.

Part of the AlterLab Academic Skills suite.