Nanobody and an unusual glucose molecule

[xushaojie-cola]

Dear authors,
I want to design only the CDR3 region of nanobodies, but I encountered several problems during the actual process.
Here is my YAML file:
entities:

• file:
  path: 2D29-G.pdb
  include:
  - chain:
  id: A

  design:
  - chain:
  id: A
  res_index: 98..111

  structure_groups:
  - group:
  visibility: 0
  id: A
  res_index: 98..111

  binding_types:
  - chain:
  id: G
  res_index: 201

Run command: boltzgen run 29.yaml --output workbench/29 --protocol nanobody-anything --num_designs 1000 --budget 50

Question 1
I want to design only the CDR3 region of a nanobody. In my YAML config, I didn’t include the antigen chain (chain G) in include:, but I did specify it in binding_types: as the target. The run completed successfully. However, I’m not sure whether the generated designs actually target chain G. During folding, I saw this log message:
[Info] Number of files to process (including already processed ones): 1000
Found 1 targets and 1000 remaining designs...
Does “1 target” refer to chain G? How can I confirm that chain G was used as the binding target?

Question 2
To solve the first problem, I added chain G into include:. Everything else was the same. But then the run failed during the inverse folding step. The error said that a glucose molecule in the database is different from the glucose molecule in my chain G. In fact, my chain G has an extra azide group compared to a normal glucose. How can I properly define chain G so that BoltzGen accepts it ?

I would be very grateful for any guidance you can provide.
Best regards

[chr-wn]

Hi, thanks for the detailed write-up! Here're my thoughts without testing it exactly in code. If you send me the specific task, I can try to replicate the error or get a colab notebook with a solution up and running.

Disclaimer: I'm not an author of boltzgen.

q1: was chain G used as the binding target?

No. In your first config, chain G was not included in the include: section, so it was not part of the structure that boltzgen operates on. Even though binding_types references chain G, the binding annotations are applied to the full PDB structure first but then masked by the include filter (see schema.py line 2343). Since chain G's residues are not included, all binding information for chain G is discarded.

The "1 target" in the logs Found 1 targets and 1000 remaining designs... refers to the number of different design specification files used (i.e., your one .yaml file), not to chain G. It's basically bookkeeping for grouping designs by their source specification. So the 1000 designs generated were just CDR3 loop redesigns without any target-binding guidance.

To fix: add chain G in include: so it becomes part of the structural context during design, and use binding: to mark which residues are binding sites.

q2: glucose/azide mismatch during inverse folding

When you include chain G, boltzgen has to parse all its residues, including non-polymer chains (like glycans). It errors because your PDB labels the modified sugar using a standard CCD code (like GLC or NAG), but the actual atoms in the file include extra azide atoms. Boltzgen looks up the CCD reference molecule and sees it mismatch the atom identity/count.

To workaround you can either:

• Exclude the problematic residue(s): If the azide-modified sugar is not at the binding interface, simply exclude it:

exclude:
  - chain:
      id: G
      res_index: <index_of_modified_sugar>

• Edit the PDB file to remove the modified sugar (or replace it with the standard glucose if appropriate).
• Rename the modified residue to a unique 3-letter code and provide its SMILES or a pickled RDKit mol in the molecules directory.

Also note: In your binding_types specification you had res_index: 201. I think the correct syntax to mark binding residues is binding: 201 (not res_index:).

Here's a skeleton you can try:

entities:
  file:
    path: 2D29-G.pdb
    include:
      - chain:
          id: A
      - chain:
          id: G
    exclude:
      - chain:
          id: G
          res_index: <modified_sugar_residues>
    design:
      - chain:
          id: A
          res_index: 98..111
    structure_groups:
      - group:
          visibility: 0
          id: A
          res_index: 98..111
    binding_types:
      - chain:
          id: G
          binding: 201

Hope this helps --- feel free to follow up if you run into any other issues!

[xushaojie-cola]

@chr-wn Thank you very much for your detailed reply. It helped me understand why I was losing chain G before and what "1 target" actually meant, and also corrected my syntax error. After posting my question, I did some troubleshooting on my own. First, I tested using ccd: GLC and found that the results generally included both chain A and chain G—which indirectly proved that my previous chain G was not being used as a binding target (please forgive this seemingly foolish move, but it was an effective method I could think of at the time). At the same time, through similar tests, I realized that my azide-containing glucose was indeed not easily understood by boltzgen. So in the end, I switched to using the SMILES format. Here is the YAML file that finally ran successfully for me:

yaml
entities:

• file:
  path: 2D29-G.pdb
  include:
  - chain:
  id: A

  design:
  - chain:
  id: A
  res_index: 98..111

  structure_groups:
  - group:
  visibility: 0
  id: A
  res_index: 98..111

• ligand:
  id: G
  smiles: "C1C@HOCCCCCN=[N+]=[N-]"
  binding: A
  I later chose to define the ligand via ligand and set its binding: A.
  This is my first time posting a question on GitHub. I truly feel honored that my question received your attention and such a detailed reply. Finally, thank you again very much for your answer!

[chr-wn]

Great, I'm glad you got it working (your empirical test with ccd: GLC was pretty clever). Happy I helped clarify some syntax for the future, and good luck with your CDR3 designs!