Structure design with BoltzGen

Tom Goddard
Biophysics 204B: Methods in Macromolecular Structure, 2 weeks on structure prediction and design
January 13, 2026

January 13 agenda

Journal club, Team 3: BoltzGen: Toward Universal Binder Design
Use BoltzGen to design opioid peptides, start with pdb 8f7q with beta-endorphin. 3 designs keep best.
Copy designs to laptops and look at in ChimeraX hbonds and clashes.
Describe team projects.

Opioid peptide input files opioid_peptide.zip and with example output opioid_peptide_results.zip (2 MB).

Design peptides that activate the human mu-opioid receptor

Look at pdb 8f7q in ChimeraX showing just the recepter (chain R) and beta-endorphin peptide (chain P).
BoltzGen github look at example input file peptide_against_specific_site_on_ragc.

Make an analogous input file 8f7q.yaml for opioid receptor

entities:
  - file:
      path: 8f7q.cif

      # Target is opioid receptor chain
      include: 
        - chain:
            id: A  # Boltzgen uses mmcif chain ids, not author chain ids.

      # The 10 designed residues should be near these receptor residues within 4A of beta-endorphin in 8f7q.
      binding_types:
        - chain:
            id: A
            binding: 126,129,133,145,146,150,153,213,214,218,220,238,320,324,328

  # A designed peptide with 10 residues
  - protein:
      id: B
      sequence: 10

Optionally add other design requirements following input yaml example. For example, design cyclic peptide, or require peptide is a loop like enkephalin and endorphins.
I specify where the peptide should bind because without that it binds all over the place. I took every residue in the 8f7q structure within 4 Angstroms of the first 10 residues of beta-endorphin. Show the ChimeraX command to find those residues.

Using BoltzGen on RunPod virtual machines

BoltzGen doesn't work on Mac in my tests (the code specifically refers to Nvidia CUDA) so we will run on RunPod.io virtual machines.
Startup virtual machine as done in last class when we tried predicting an NTCA dimer with Boltz.
Make sure to choose Network Drive "Boltz and BoltzGen" to get the preinstalled BoltzGen.

Copy BoltzGen input to VM and calculate designs

Make a directory opioid_peptide with the needed files (8f7q.yaml 8f7q.cif, command) on your laptop and copy it to RunPod VM. You can use my example files opioid_peptide.zip.
To copy we can use scp. We will use an ssh key id_eds25519.txt that I will tell you how to get in class.
The ssh key must have permissions that only allow you to read it, otherwise scp won't work. So remove group/other read permission.

RunPod Details web page lists IP and port.
```
	  chmod go-r ~/Downloads/id_ed25519.txt
```
Set variables for the IP address and port for your RunPod virtual machine to use in the copying commands
```
	  IP=203.57.40.77
	  PORT=10077
```

Copy opioid_peptide directory from laptop to VM.

	  scp -r -P $PORT -i ~/Downloads/id_ed25519.txt ~/Downloads/opioid_peptide root@$IP:

Login to the VM from a terminal on your Mac laptop (/Applications/Utilities/Terminal)
```
	  ssh root@$IP -p $PORT -i ~/Downloads/id_ed25519.txt
```
Run the boltzgen predictions (should take 5-15 minutes). For a quick test run 3 designs and keep just the best scoring (--budget 1) Real research runs are 10,000 - 50,000 designs and take several days on one fast GPU.
```
	  cd opioid_peptide
	  time /workspace/boltzgen/bin/boltzgen run 8f7q.yaml --output results3 --protocol peptide-anything --num_designs 3 --budget 1 >& output3 &
```
To see progress in the output3 log file use the tail command
```
	  tail -f output3
```
When the timing info is output the boltzgen run has completed and you type Ctrl-C to exit the "tail -f" command which was monitoring the output file.
When design finishes copy the opioid_peptide directory with results back to laptop using scp on the laptop for analysis. Run the scp command from your laptop (not from the VM).
```
	  exit   # This logs out of the VM and gets your terminal back to your laptop.
	  scp -r -P $PORT -i ~/Downloads/id_ed25519.txt root@$IP:opioid_peptide .
```

8 clashes pink, 1 hydrogen bond blue

Look at design hydrogen bonds and clashes in ChimeraX.

Open the best design (of 3 calculated) in ChimeraX:

   opioid_peptide/results3/final_ranked_designs/final_1_designs/rank1_8f7q_2.cif

Show residues near peptide chain B.
```
    show /B :< 4
```
Show clashes
```
    clashes /B
```
Show hydrogen bonds peptide (chain B) to receptor (chain A)
```
    hbonds /B restrict /A
```

Design scores

Look at design scores .csv file in a text editor

   opioid_peptide/results3/final_ranked_designs/final_designs_metrics_1.csv

Can be easier to read scores by transposing .csv file

    designed_sequence,MTPLEEFLAL
    design_to_target_iptm,0.20324
    min_design_to_target_pae,17.35381
    design_ptm,0.755
    design_hydrophobicity,36.414
    ...

ipTM score 0.2 very poor, and minimum PAE between peptide and receptor 17 Angstroms very poor.
Next step would be to run time 100 designs to estimate how long 10,000 designs would take.

Installing BoltzGen

If no GPUs are available on RunPod at the US-IL-1 datacenter where the Network drive "Boltz and BoltzGen" is located you can instead get a GPU at any data center and install BoltzGen yourself.

From a shell on your RunPod VM make a Python virtual environment in your home directory called boltzgen
```
	  cd
	  python -m venv boltzgen
```
Then install boltzgen
```
	  boltzgen/bin/pip install boltzgen
```