redesigning experimentally validated binder

[dhiraj82]

Hi
I am so far unsuccessful in designing a binder against tetrameric complex. all I am getting are bad less than 0.5 iPTM, more than 10, 15 iPAE binders. Earlier, I designed and validated a binder against one of the dimeric partners of the complex. So I am trying to use protein-redesign option to redesign the binder against the heterotetramere which has some steric clash. However, while redesigning, all I see is my target complex falling apart. During intermediate design step, the target monomers are not assembling correctly to form correct tetramer. How can I redesign only binder portion to prevent steric clash. I am redesigning chain E.
below is my YAML file.
Thanks
Dhiraj

entities:
  # Specify a designed protein chain
  # random number between 80 and 140 of designed residues (inclusive)
  # Specification of the target which is extracted from a .cif file
  - file:
      path: target_binder.pdb
      # Which chain and residues in the .cif file to use as target (uses only chain A here)
      include: 
        - chain:
            id: A
            res_index: 1..220
        - chain:
            id: B
            res_index: 1..220
        - chain:
            id: C
        - chain:
            id: D
        - chain:
            id: E
      design:
        - chain:
            id: E
      structure_groups:
        - group:
            visibility: 1
            id: A
            id: B
            id: C
            id: D

[buerbaumer]

Hi @dhiraj82,

One issue in the YAML is that the structure_groups block repeats the id key inside the same mapping:

id: A
id: B
id: C
id: D

Most YAML parsers keep only the last duplicate key, so this likely applies the structure group only to chain D. That would leave A-C without the intended structural conditioning, which could explain why the tetramer falls apart during the design step.

I would write the target complex as separate group entries, all with the same visibility value so their relative arrangement is specified:

structure_groups:
  - group:
      id: A
      visibility: 1
  - group:
      id: B
      visibility: 1
  - group:
      id: C
      visibility: 1
  - group:
      id: D
      visibility: 1

and keep only chain E in design. If only a local steric clash needs fixing, it may also be better to restrict design to the clashing residues on chain E using res_index rather than redesigning the whole chain.

If the goal is sequence-only redesign on the fixed experimental backbone, I would also consider --only_inverse_fold instead of full backbone diffusion redesign.

Regards,
Harald (Contributor)

[dhiraj82]

thank you @buerbaumer. the complex is still falling apart after your suggested edit.
However, I observed that BoltzGen does not model the complex accurately. The native complex is a β-strand-swapped homodimer (A:B) in which a β-strand from one monomer contributes to the β-sheet of the partner monomer. In the refolded structure, this strand-exchange interaction is not there. instead the beta strand is contributing to self beta sheet instead of the one belonging to interaction partner. C and D interacts with A:B in such a way that they interacts with both chain A and B of dimer. if A and B are not assembled properly, C and D will not interacts with A:B making A:B:C:D. This explain everything.
just to add - A:B is a dynamic complex but there is no monomer structure of A:B in PDB so I believe boltz should be modeling it as dimer.
Also, this behaviour is specific to the redesigning protocol and protein-anything protocol is behaving as expected. also, I noticed this very early on at intermediate design state. is it a bug or there is some mistake in my yaml file? my yaml file is -

entities:
  - file:
      path: target_binder.pdb    
      include: 
        - chain:
            id: A
            res_index: 1..220
        - chain:
            id: B
            res_index: 1..220
        - chain:
            id: C
        - chain:
            id: D
        - chain:
            id: E
      
      structure_groups:
        - group:
            id: A
            visibility: 1
      structure_groups:
        - group:
            id: B
            visibility: 1
      structure_groups:
        - group:
            id: C
            visibility: 1
      structure_groups:
        - group:
            id: D
            visibility: 1
      design: 
        - chain:
            id: E

Thank you
Dhiraj

[buerbaumer]

Hi @dhiraj82,

Thanks for the detailed follow-up. The strand-swapped A:B interface you describe is a plausible reason this target is difficult, but there is still a YAML issue in the updated file.

In the updated YAML, structure_groups: is declared four times under the same file: mapping. YAML parsers normally keep only the last duplicate key, so BoltzGen likely sees only the final block:

structure_groups:
  - group:
      id: D
      visibility: 1

That means A, B, and C are not getting the intended structure-group conditioning.

Please try exactly one structure_groups: key with multiple list entries:

entities:
  - file:
      path: target_binder.pdb
      include:
        - chain:
            id: A
            res_index: 1..220
        - chain:
            id: B
            res_index: 1..220
        - chain:
            id: C
        - chain:
            id: D
        - chain:
            id: E

      structure_groups:
        - group:
            id: A
            visibility: 1
        - group:
            id: B
            visibility: 1
        - group:
            id: C
            visibility: 1
        - group:
            id: D
            visibility: 1

      design:
        - chain:
            id: E

Using the same nonzero visibility value for A/B/C/D tells BoltzGen to condition their relative geometry together. Different visibility values would condition chains separately, not their mutual arrangement. Leave chain E out of structure_groups if you want its backbone/redesigned region to remain flexible.

If the complex still falls apart after this exact YAML, then the issue may be a limitation of the redesign diffusion step for this strand-swapped assembly rather than just YAML. For an experimentally validated binder where you mainly want to change sequence on a fixed backbone, I would test:

boltzgen run your.yaml --protocol protein-redesign --only_inverse_fold

or restrict design: to only the clashing residues on chain E.

Regards,
Harald