Top rank Relaxed Structure obtained from notebook run is not giving the cyclic peptide strucutre

[N-Akash10]

config.json

Context/Issue

Keeping the input parameters default, the top relaxed structure obtained from colab notebook, for PDB-1SFI, gives the peptide with NH2 at N-terminus and COOH at C-terminus. Hence, visualizing in stick mode in pymol, the peptide is not forming the cyclic conformation. The distance between "N" of N-terminus and "C" of C-terminus is 2.8Å, however for cyclic peptide formation, this bond length should be around 1.33Å. I might be missing something. Please help.

[N-Akash10]

Hi, I'm still looking for a solution. Please help.

[blacktanktop]

This is as previously POSTED in X. If you run it as default, it will close the ring like this.
https://twitter.com/black_tank_top/status/1745357798966448168

Can you show me the output PDB or .ipynb?

[N-Akash10]

Thanks a lot for your reply.

Yes in default mode it is closing the ring. But when I'm giving "num_relax" value as 1 then the resultant relaxed file, the ring is not closing.
Here is the .pdb file attached. [Please find the .zip file and the run it after unzipping it]

1SFI_relaxed_b9c30_0_relaxed_rank_001_alphafold2_multimer_v3_model_5_seed_000.zip

Waiting for your reply as soon as possible.

[blacktanktop]

Once relaxed, the cyclic constraint is released.
This is because AlphaFold is originally a predictive model for linear chain proteins, and if atoms are close enough to bind to each other, relaxation avoids repulsion, so the protein is no longer cyclic.

Using Rosetta's minimize properly avoids minor crash with the target protein (as described in the paper).

I don't know what your objective is, but I think the result will only be the same within the margin of error level whether you relax or not (although minor crashes may be avoided).

[N-Akash10]

Hi,

Thank you so much for the reply and your comments.

Here I'm showing one example where the relaxed structure is more native like than the unrelated one. In the relaxed [rank_1], even though the ring is not closing, the structure is more native like where I observed two expected disulfide bonds. On the other hand, in the unrelaxed[rank_1], only one disulfide bond is forming that too is distorted.

Here I'm trying to use your notebook to predict the protein-cyclic peptide interaction. I'm getting well-docked results in the rank_1 models but the disulfide bonds are the ones which is giving the trouble.

I am attaching the comparative results obtained for relaxed and unrelaxed. Also, I'm putting the predicted IDDT plot for reference.

Also, I've observed that only rank_3 of the unrelaxed structure is more similar to the relaxed rank_1 [attached].

So in this case is it a good idea to take rank_3 unrelaxed instead of rank_1?

I know that it's a long query but still, I am hopeful that you reply soon.

Thanks.

[blacktanktop]

The rank is only determined by AlphaFold (ColabFold) criteria. (e.g., pLDDT is high, pAE is low, etc.) So, top1 is about whether it is plausible by that criterion as an overall structure, and cyclic offset is hacking AlphaFold, so it cannot guarantee all disulfide bonds.
It is only a matter of using the RANK that the person using it thinks is correct for the right reasons.

The issue of whether it is good or bad as a science will not be dealt with in this issue. It is only an implementation issue.

Another limitation of the current implementation is that cyclic peptides can bind at impossible valence. Please submit a pull request for suggestions to improve this issue.

There was also a previous report called HighFold, which is specialized for disulfide bonding, so you might want to try that one. You will have to implement it yourself, but I have tried it.