Suggestions for the formation of disulfide bonds

[Charlesjc-lab]

The work is incredibly constructive, particularly for RFdiffusion, which traditionally struggles to produce stable short peptides in the range of 10 to 30 amino acids.
Despite this, after a week of rigorous testing, I encountered considerable issues with the formation and subsequent recovery (AF2/AF3) of disulfide bonds within the peptides synthesized. I'm also at a loss regarding the methodology employed for the incorporation of these bonds. Could you offer some guidance on this matter? Has the author explored the use of pystapler, a structure-based method for the introduction of disulfide bonds?

[YuzheWangPKU]

Thank you for your support of our work and for your constructive suggestions!

1. In our prior tests, we found that AF-multimer (in single sequence mode) struggles to accurately recover the native poses of peptide binders within the 8-30 amino acid range. This challenge likely stems from the exclusion of chains shorter than 16 residues in AF-multimer’s training data, leading to out-of-distribution protein-peptide interactions that are not predicted correctly. We are currently benchmarking AF3's performance in predicting the binding conformations of short peptides (both w/ and w/o disulfide bonds), which may prove advantageous before directly utilizing AF2/AF3 for conformation recovery.

2. The rationale behind our disulfide bond construction pipeline is that, in addition to applying geometric criteria—similar to those used by Pystapler (a tool we were not previously aware of, thank you for bringing it to our attention)—we incorporate an additional filter. This filter ensures that only residues not critical to binding interface interactions are considered for substitution with cysteines to form disulfide bonds; in other words, we avoid substituting residues essential for binding. To achieve this, we assume that residues where the model exhibits lower "confidence" in residue type, as indicated by higher residue type entropy, are less critical to binding. Empirically, we have found this strategy effective in generating peptide binders with disulfide bonds.

We acknowledge that for certain targets, the current version of DiffPepBuilder still requires extensive sampling and manual adjustment of binding information to generate valid and high-affinity peptide binders. We are actively working to improve our model to enhance its success rate and robustness. Please stay tuned for our upcoming updates.

[Guo-Stone]

According to your first answer, I know it's challenging to evaluate the designability of the peptide binders. I have tried to use AF3 for evaluation, but there is a usage limitation everyday, which means that we can't evaluate these generated samples thoroughly. How about your current work of designability evaluation? I know that UMol (Structure prediction of protein-ligand complexes from sequence information with Umol) may work but haven't tried yet.