search

YuzheWangPKU / DiffPepBuilder

Official repository for Target-Specific De Novo Peptide Binder Design with DiffPepBuilder
MIT License
63 stars 7 forks source link

Minimization failed after 100 attempts #1

Open Charlesjc-lab opened 3 months ago

Charlesjc-lab commented 3 months ago

In the redocking step, I encounter an error message that says "Minimization failed after 100 attempts." Can the author provide a solution?

YuzheWangPKU commented 3 months ago

Please try the latest version which includes updates to the redocking pipeline. If the problem persists, we recommend manually inspecting and removing any structures with backbone steric clashes, as these may cause the energy minimization process to fail.

We have empirically observed that including too many residues as hotspots or motifs may bias the model towards generating peptides with significant steric clashes. If you find that many peptides are exhibiting steric clashes, please consider adjusting the binding information provided to the model accordingly.

Charlesjc-lab commented 3 months ago

Yes, I have also tried the method you mentioned, using the new scripts and manual check to solve similar problems. I also have suggestions and ideas for generating conformations, and it is possible to filter the output structures using the af2 method in dl_binder_design/colabdesign. However, unfortunately, AF2 is not a good filtering method for diffpepbuilder. Are there any effective filtering methods that the author has actually used?

wfh1998 commented 3 months ago

Yes, I have also tried the method you mentioned, using the new scripts and manual check to solve similar problems. I also have suggestions and ideas for generating conformations, and it is possible to filter the output structures using the af2 method in dl_binder_design/colabdesign. However, unfortunately, AF2 is not a good filtering method for diffpepbuilder. Are there any effective filtering methods that the author has actually used?

Thank you for your suggestion. We previously attempted to perform redocking tests on several peptide structures using AF2 (as we are also working on peptide docking). The results indicated that AF2 performs poorly in predicting certain loop conformations of peptides in redocking tests. Therefore, we opted to use a relax protocol based on Rosetta's energy score for post-processing. However, the method you mentioned, such as dlbinderdesign, is also a reasonable approach for subsequent backbone and sequence optimization pepline, and it is worth exploring.

amin-sagar commented 2 months ago

Hello. Sorry for commenting on a closed issue. I am also observing structures with significant backbone clashes. Do you have some general recommendations on the number of hotspot residues, surface area of the selected pocket etc., based on your experience? Best, Amin.

Charlesjc-lab commented 2 months ago

When employing an input akin to RFdiffsion (a hotspot comprising 3 to 6 hydrophobic residues), results can be obtained. However, it is challenging to circumvent instances of conflict. I wish it is worth to you.

Furthermore, when employing a threshold value of T=0.001, the number of staple peptides identified is significantly reduced.

YuzheWangPKU commented 2 months ago

Thanks for the valuable comment from @Charlesjc-lab! In our experiments, we used alanine scanning to identify hotspot residues for the three targets (3CLpro, ALK1, TNF-α). We found this approach effective for our model. For targets that lack a native binding ligand or natural substrate, we acknowledge that iterative optimization of the hotspot residues may still be necessary, and we found the hotspot selection procedures from RFdiffusion to be beneficial.

If significant clashes persist, an alternative to consider is manually adjusting the pocket_cutoff parameter in the script experiments/process_receptor.py using the flag --pocket_cutoff. Our previous experiments indicated that an excessively large pocket can lead to generated peptide binders becoming overly "deep" within the pocket, leading to the clashes.

amin-sagar commented 2 months ago

Thanks @YuzheWangPKU and @Charlesjc-lab Does it make sense to include a structure quality estimate like Qmean in the workflow and filter out the structures with bad geometry? Especially, since DiffPepBuilder is quick fast in generating peptides.

YuzheWangPKU commented 2 months ago

Thanks @YuzheWangPKU and @Charlesjc-lab Does it make sense to include a structure quality estimate like Qmean in the workflow and filter out the structures with bad geometry? Especially, since DiffPepBuilder is quick fast in generating peptides.

Thank you for your suggestion! In our design pipeline, we primarily use Rosetta ddG as a criterion to filter out generated peptides that exhibit backbone clashes or distorted structures, and we have found this approach to be effective. We plan to integrate it into our pipeline in a more end-to-end manner in the next version.

amin-sagar commented 2 months ago

Thanks.

My binding site is quite charged. I had specified 6 residues as hotspot but none were hydrophobic. I am wondering if this is leading to peptides with backbone clashes.

For the three targets that you studied, did you always choose hydrophobic residues as your hotspots? Or did you have combination of hydrophobic and charged residues?

Thanks again. Amin.