2. RFdiffusion: Advanced Design

Live Workshop Session

🎥 Live workshop recording — Backbone generation in protein design workflows

Hands-On Exercise

Setup for the activity.
1. Go to the RFdiffusion installation and make a new directory for the activity: mkdir activity.
Generate unconditional monomers.
1. Make new folder for this example: cd activity; mkdir 1_uncond; cd 1_uncond
2. Copy design_unconditional.sh here: cp ../../examples/design_unconditional.sh .
3. Edit the design script by using 1) 'contigmap.contigs=[75-150]', 2) inference.num_designs=5 , 3) ../../scripts/run_inference.py.
4. Run the script: bash design_unconditional.sh
  - What did the script create?
  - Look at the structures. What are kinds of topologies did you get?
Generate monomers that scaffold a motif.
1. Make new folder for this example: cd activity; mkdir 2_motifs; cd 2_motifs
2. Copy design_motifscaffolding.sh here: cp ../../examples/design_motifscaffolding.sh .
3. For this example, we’ll be scaffolding a site from RSV-F protein. Let’s copy that .pdb here: cp ../../examples/input_pdbs/5TPN.pdb .
4. Edit the design script by using 1) inference.input_pdb=5TPN.pdb, 2) 'contigmap.contigs=[10-40/A163-181/10-40]', 3) inference.num_designs=5 , 4) ../../scripts/run_inference.py.
  - The backbone we’re generating will have 10-40 residues (randomly sampled), the the motif residues 163-181 (inclusive) on chain A of the input, then 10-40 residues (randomly sampled).
5. Run the script: bash design_motifscaffolding.sh
  - Look at the structures. How well is the motif scaffolded?
Generate partially diffused structures.
1. Make new folder for this example: cd activity; mkdir 3_partial; cd 3_partial
2. Copy design_partialdiffusion.sh here: cp ../../examples/design_partialdiffusion.sh .
3. For this example, we’ll be partially noising and denoising a 79 residue protein 2KL8. Let’s copy that .pdb here: cp ../../examples/input_pdbs/2KL8.pdb .
4. Edit the design script by using 1) inference.input_pdb=2KL8.pdb, 2) 'contigmap.contigs=[79-79]’, and 3) inference.num_designs=5.
  - Here we’re generating diversity around a particular fold by noising and denoising 10 steps (20% of the full trajectory). We’re adding noise to the entire structure (all 79 residues), but part of the structure can also be held fixed.
5. Run the script: bash design_partialdiffusion.sh
  - Look at the structures. How similar are the outputs to the original structure?
Generate binders with hotspots.
1. Make new folder for this example: cd activity; mkdir 4_hotspot; cd 4_hotspot
2. Copy design_ppi.sh here: cp ../../examples/design_ppi.sh .
3. For this example, we’ll be designing binders to insulin receptor. Let’s copy that .pdb here: cp ../../examples/input_pdbs/insulin_target.pdb .
4. Edit the design script by using 1) inference.input_pdb=insulin_target.pdb, 2) 'contigmap.contigs=[A1-150/0 70-100]’, 3) ’ppi.hotspot_res=[A59,A83,A91]’, 4) inference.num_designs=5, 5) denoiser.noise_scale_ca=0, and 6) denoiser.noise_scale_frame=0.
  - Here, we’re designing binders to insulin receptor. The contig describes the protein we want: residues 1-150 of the A chain of the receptor, a chainbreak (we don’t want to fuse the binder and target!), a 70-100 residue binder to be diffused. We also tell diffusion to target three specific residues, specifically residues 59, 83 and 91 of chain A. Finally, we reduce the noise added during inference to 0 to improve the quality of the designs.
5. Run the script: bash design_ppi.sh
  - Look at the structures. What are the topologies of your binders?
Generate fold-conditioned structures.
1. Make new folder for this example: cd activity; mkdir 5_fold_cond; cd 5_fold_cond
2. Copy design_timbarrel.sh here: cp ../../examples/design_timbarrel.sh .
3. For this example, we’ll be diffusing a TIM barrel by providing course-grained specification of the fold. Let’s copy that fold information here: cp -r ../../examples/tim_barrel_scaffold/ .
  - What do the files in this folder represent?
4. Edit the design script by using inference.num_designs=5.
  - Here, we’re making a TIM barrel by providing course-grained specification of the fold. We specify the output path. We tell RFdiffusion that we want to do scaffoldguided design, and that we are not making a binder to a target (just a monomer). We provide a path to a directory of TIM barrel scaffolds, generated with the helper script helper_scripts/make_secstruc_adj.py. We generate 5 designs, with a reduced noise scale during inference of 0.5. We also sample additional length to increase diversity of the outputs by masking the loops and inserting 0-5 residues into each loop. We also add 0-5 residues to the N and the C-terminus. These allow use to sample additional diversity in our designs.
  - You may need to edit rfdiffusion/inference/model_runners.py like 751 to self.blockadjacency = iu.BlockAdjacency(conf, conf.inference.num_designs)
5. Run the script: bash design_timbarrel.sh
  - Look at the structures. How do these generations compare to the TIM barrel structure used for conditioning (6WVS)?
Generate oligomers of various symmetries.
1. Make new folder for this example: cd activity; mkdir 6_symmetry; cd 6_symmetry
2. Copy design_cyclic_oligos.sh, design_dihedral_oligos.sh, and design_tetrahedral_oligos.sh here: cp ../../examples/design_cyclic_oligos.sh ., cp ../../examples/design_dihedral_oligos.sh ., and cp ../../examples/design_tetrahedral_oligos.sh .
3. Generate cyclic oligomers:
  1. Edit the design_cyclic_oligos.sh script by changing 1) inference.symmetry="C4”, 2) inference.num_designs=5, 3) inference.output_prefix="example_outputs/C4_oligo”, and 4) 'contigmap.contigs=[200-200]’.
    - In this example, we generate 5 designs of C4 symmetric oligomers. For symmetrical diffusion, we need the symmetry config. We also apply an external potential to promote contacts both within (with a relative weight of 1) and between chains (relative weight 0.1). We specify that we want to apply these potentials to all chains, with a guide scale of 2.0 (a sensible starting point). We decay this potential with quadratic form, so that it is applied more strongly initially. We specify a total length of 200 aa, so each chain is 50 residues long.
  2. Run the script: bash design_cyclic_oligos.sh
    - Do the structures have the desired symmetry?
    - What topologies are found in the individual subunits?
4. Generate dihedral oligomers:
  1. Edit the design_dihedral_oligos.sh script by changing inference.num_designs=2.
    - In this example, we generate 2 D2 symmetric oligomers using the symmetry config. We also apply an external potential to promote contacts both within (with a relative weight of 1) and between chains (relative weight 0.1). We specify that we want to apply these potentials to all chains, with a guide scale of 2.0 (a sensible starting point). We decay this potential with quadratic form, so that it is applied more strongly initially. We specify a total length of 320 aa, so each chain is 80 residues long.
  2. Run the script: bash design_dihedral_oligos.sh
    - Do the structures have the desired symmetry?
    - What topologies are found in the individual subunits?
5. Generate tetrahedral oligomers:
  1. Edit the design_tetrahedral_oligos.sh script by changing 1) inference.num_designs=2, 2) 'contigmap.contigs=[720-720]’.
    - In this example, we generate tetrahedral symmetric oligomers. We use the symmetry config, and specify we want a tetrahedral oligomer, with 2 designs generated. We specify the output prefix, and also the potential we want to apply. This external potential promotes contacts both within (with a relative weight of 1) and between chains (relative weight 0.1). We specify that we want to apply these potentials to all chains, with a guide scale of 2.0 (a sensible starting point). We decay this potential with quadratic form, so that it is applied more strongly initially. We specify a total length of 720 aa, so each chain is 60 residues long
  2. Run the script: bash design_tetrahedral_oligos.sh
    - Do the structures have the desired symmetry?
    - What topologies are found in the individual subunits?

Independent Project

(Use your target protein)

Milestone 1: Target Preparation & Initial Exploration

Set up your project directory structure
Download and inspect your assigned target PDB
Identify the binding surface and hotspot residues from the table below
Run 2-3 test designs to verify your setup is working correctly

Milestone 2: Unconditional Binder Generation

Generate at least 10 unconditional binders of length 70-100 towards your target protein.
- Is there a particular epitope on the target that RFdiffusion prefers?
- Is there a particular binder topology that RFdiffusion prefers?

Milestone 3: Hotspot-Guided Binder Design

Each of the targets that were assigned were previously used in binder design experiments (1, 2). These experiments made use of hotspots located on the target proteins. Generate at least 10 binders towards your target protein using the hotspots below.
- Is there a particular binder topology that RFdiffusion prefers to generate?

Target	UniProt ID	PDB ID	Hotspot Residues (chain and residue index from PDB)
PD-L1	Q9NZQ7	5O45	A56, A115, A123
IL-7Rɑ	P16871	3DI3	B58, B80, B139
TrkA receptor	P04629	1WWW	X294, X296, X333
IFNAR2	P48551	2LAG	B52, B82, B98
Bet v 1-A	P15494	1BV1	A24, A28, A43

Milestone 4: Potential Optimization

Potentials can be powerful ways to bias the generation process. Try at least 3 different combinations of potentials, generating 5 backbones with each. Be sure to use hotspots for these generations too!
- How did the potentials change your outputs?
- Did you find a particularly useful configuration of potentials?

Milestone 5: Analysis & Selection

Compare your unconditional vs. hotspot-guided vs. potential-optimized designs
Identify your top 3-5 most promising binder designs
Document what makes these designs stand out (binding pose, topology, contact with hotspots, etc.)

Troubleshooting Guide

Common Issues & Solutions

Problem	Likely Cause	Solution
“CUDA out of memory”	GPU memory exhausted	Reduce `inference.num_designs` to 1-2, or design smaller proteins
Script hangs at “Initializing model”	Missing model weights or incorrect paths	Verify RFdiffusion installation and model checkpoint paths
“FileNotFoundError” for input PDB	Incorrect file path or missing file	Check that PDB file exists in current directory with `ls *.pdb`
Designs look extended/unfolded	Insufficient denoising or inappropriate settings	Check `inference.num_steps` (should be ~50), verify contig syntax
Binders don’t contact target	Incorrect contig specification or chainbreak	Verify `/0` chainbreak in contigs, check residue numbering in target PDB
Binders don’t contact hotspots	Hotspot residues too far apart or incorrect chain ID	Verify chain IDs and residue numbers match your PDB file exactly
“ImportError” or module not found	Conda environment not activated	Activate RFdiffusion environment: `conda activate RFdiffusion` (or appropriate env name)
Script runs but produces no outputs	Output directory doesn’t exist or permissions issue	Check that output directory exists, verify write permissions
Symmetric oligomers don’t look symmetric	Incorrect symmetry specification or potentials	Verify symmetry string (e.g., “C4”, “D2”, “T”), check potential settings
“BlockAdjacency” error in fold-conditioned	Known bug in model_runners.py	Edit `rfdiffusion/inference/model_runners.py` line ~751 as noted in activity
Very slow generation times	Normal for large/complex designs	Tetrahedral oligomers can take 30-60 min. Consider using `screen` or `tmux`
All designs look very similar	Insufficient diversity sampling	Increase contig length ranges (e.g., `[60-100]` instead of `[80-80]`), adjust noise scales

Tips for Success

Always check your PDB file first: Use PyMOL or ChimeraX to verify chain IDs and residue numbering before running
Start small: Test with num_designs=1 first to verify your setup works
Save your commands: Keep a log of successful parameter combinations
Use descriptive output names: Include key parameters in output_prefix (e.g., hotspot_A59_A83_A91)
Check the logs: RFdiffusion creates log files - read them if something goes wrong