RFpeptide protocol

[DazLe-Q]

Hello authors,

I am working on a peptide-protein binding design project and am using a computational pipeline that involves RFdiffusion (for backbone generation) -> ProteinMPNN (for sequence design) -> AlphaFold2 (for structure prediction and validation).

I have two questions regarding the potential use of LigandMPNN to replace ProteinMPNN in this pipeline:

1. Replacing ProteinMPNN with LigandMPNN: Since my goal is to design a peptide that binds to a target protein, I understand that LigandMPNN is specifically designed for tasks with ligand context. Is replacing ProteinMPNN with LigandMPNN in this pipeline a recommended approach for generating high-quality protein-ligand sequences?

2. Error with pack_side_chains feature: I encountered an error when running LigandMPNN with the pack_side_chains feature enabled. I noticed that this is an optional feature. Is it a mandatory feature for generating high-quality designs, or can I skip it and still get reliable sequences for subsequent experimental validation?

Here is the error and the command I used:

Command:

python ../../run.py \
    --model_type "ligand_mpnn" \
    --pdb_path "./inputs/1BC8.pdb" \
    --out_folder "./outputs/ligandmpnn_default"
    --batch_size 3 \
    --number_of_batches 2 \
    --pack_side_chains 1 \
    --number_of_packs_per_design 1

**Traceback:**

Packing side chains...
Traceback (most recent call last):
  File "G:\LigandMPNN\run.py", line 990, in <module>
    main(args)
  File "G:\LigandMPNN\run.py", line 526, in main
    sc_dict = pack_side_chains(
                  ^^^^^^^^^^^^^^^^^
  File "G:\LigandMPNN\sc_utils.py", line 114, in pack_side_chains
    xyz14_noised = feats.frames_and_literature_positions_to_atom14_pos(
                       ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "G:\LigandMPNN\openfold\utils\feats.py", line 253, in frames_and_literature_positions_to_atom14_pos
    group_mask = nn.functional.one_hot(
                      ^^^^^^^^^^^^^^^^^^^^^^
RuntimeError: one_hot is only applicable to index tensor.

Thank you for your time and assistance!

[rclune]

It seems like these questions are more related to Ligand/ProteinMPNN than RFdiffusion. You will likely get more support and comments by posting this issue in the LigandMPNN or ProteinMPNN repositories.

1. According to the LigandMPNN repository LigandMPNN contains the inference code for both LigandMPNN and ProteinMPNN. They are run slightly differently, so it might be a bit of work to adjust a pipeline built around ProteinMPNN to use LigandMPNN instead, but the functionalities should all be there.

2. I'm not completely certain why that error is appearing, but what I can tell you is that disabling the side chain packer (setting pack_side_chains to 0) will make it so the output from Protein/LigandMPNN will not have packed side chains - so it is optional to use this feature, but it will definitely impact your outputs.

[DazLe-Q]

Thank you so much for your advices!