[QUEST]

[lucatondini]

I have been working on the disaccharide-binding proteins, but I am currently stuck on the RF3 step. I have tried multiple times, but I get the same outcome every time: RoseTTAFold3 cannot recognize a disaccharide molecule as a single entity.
The core problem is that, in the PDB format, most disaccharides are treated as two separate monosaccharides. However, when I tried to provide the SMILES or CCD code or SDF file of the disaccharide to RF3, it generated protein structures with two monosaccharides spaced apart, and, of course, this configuration negatively affects the protein folding. During the model generation, RF3 also compares the models with the backbones generated by RFD3 (where there is the disaccharide)

[rclune]

If the input is a cif file, have you tried creating a struct_conn between the monosaccharides? (LLMs tend to be pretty good at adding them.)

Did you get an error when attempting to use the SDF file?

[lucatondini]

I have got an error with the SDF file. What would you suggest? I am thinking to modify the PDB structure of the backbones generated by RFD3 in order to write the struct_connect bewteen monosaccharides and then give RF3 the SMILES code or the SDF file of the disaccharide. Does this approach sound good? Whay would you implement?

[roccomoretti]

If you can post (copy/paste) the error message you got here in the chat, there's a chance someone can figure out why you might be getting it.

The other thing to keep in mind is that SDFs vary in their quality, and foundry/atomworks may be having issues with the particular format your SDF is in. You might have more success if you get an SDF of the ligand from a different source, or if you use an alternative input method (e.g. SMILES or CIF).

Also, I'm not 100%, but I think that the struct_conn will only work with CIF files, and only between residues within the same CIF structure. (It's a CIF file format feature, not an RF3/foundry-specific feature.) -- You'll need to specify the input as CIF instead of SDF/SMILES/etc.

Also, take a look at the "Folding with a Covalent Modification" section of https://github.com/RosettaCommons/foundry/blob/production/models/rf3/README.md -- If you're modeling the disaccharide as a two CCD codes (or two separate residues otherwise), you effectively have a "covalent modification" of the bond between them. There's a separate "bonds" section of the input JSON where you should be able to specify the covalent linkage between the two, even if your input specification has them as separate residues/chains.

A final note: If you're re-predicting RFD3/MPNN designed structures in the standard design pipeline, you probably don't want to provide the design structure as a template. There's too much potential for information leakage from the designed structure. The point of forward folding is to double check if your designed sequence is sufficient to fold to the structure you want. If you provide the designed structure as a template, you don't necessarily know if the resultant predictions successfully folded because the sequence fold robustly to that structure, or if the usage of the template allowed the structure prediction program to "cheat" and grab bits of the structure from the designed template. -- I'd recommend doing the reprediction of designs as sequence only. If you do need to provide a template for some reason, provide the parent experimental/crystal structure, rather than the design structure.