Request for clarification on preprocessing: splitting protein–ligand complexes into protein.pdb and ligand.mol2

[ndlongvn]

Hi,
Thank you for sharing your great work and open-source code!

I have a question regarding the data preprocessing pipeline used to generate the model inputs — specifically, the separate protein.pdb and ligand.mol2 files.

In the released version of your model, the input format expects these two files independently, but most datasets (e.g., PDBBind, PubChem, or RCSB entries) provide only complex structures where the protein, ligand, and sometimes cofactors are combined in a single .pdb file.
It would be extremely helpful if you could share the script or function that performs this separation.

Thank you very much!

[BFeng14]

Hi, I started from biolip, which is ideal for preprocessing large datasets because it already provides separated protein and ligand structures from the PDB. I recommend using biolip if you want to preprocess a large database.

For processing files, you can use Python libraries like Biopython. A simpler alternative is to extract all lines starting with "HETATM" from a PDB file to create a ligand-only file, and then convert it to the .mol2 format using Open Babel.

[ndlongvn]

Hi, I found the example in your virtual screening model has PDB ID 1R4L. Unfortunately, the processed protein and ligand file I split from the complex, which I downloaded from PDB, are different compared to those you provide, resulting in a really different prediction result. So it's the reason why I asked you for your preprocessing code.

[BFeng14]

I think this might be the problem with your PDB file, which is mostly due to different preprocessing methods. The preprocessing method of PDB files should be identical to that of BioLip, and I cannot figure out your problem for now. I think that maybe directly downloading from Biolip (https://zhanggroup.org/BioLiP) using pdbid is a good choice for rapidly solving this problem, where almost all pdb files can be found. I will also try to reproduce your problems and provide more easy-to-use Colab tools shortly.

[ndlongvn]

Thank you for your solution. I have tried to access BioLiP on the link you provided, but unfortunately, it raised the error "Not Found. The requested URL was not found on this server."

[BFeng14]

I think there might be a problem with their server. I just get their preprocessed files from my device, hope that this can help you.

1r4lA.pdb
1r4l_XX5_A_1.pdb

[ndlongvn]

Sorry, could you double-check the files in example/ace2_l.mol2 and example/ace2_p.pdb?
The files you sent me are similar to the ones I downloaded from the PDB, but they differ from those in the example folder.
I noticed that the example files have hydrogens added, but I wasn’t able to reproduce them to match your reference files exactly.

[BFeng14]

Thank you for raising this question. The hydrogen shouldn't make any difference because LigUnity will ignore all hydrogen atoms. The only difference I found was that there is a Zn atom in ace2_p.pdb file, I will check this ASAP and will let you know afterwards.

HETATM 9519 ZN ZN A 803 39.248 8.035 22.462 1.00 37.70 Zn2+

[ndlongvn]

Hi, I realized that in your dataset code, you explicitly removed all hydrogens before training or testing. However, in the pocket-parsing step, the presence of hydrogen coordinates might affect the cutting results. Does your pocket-parsing step require hydrogen input, or is this a bug in your code?

[BFeng14]

Thank you for raising this problem. The pocket-parsing step should not require any hydrogen input, and the presence of hydrogen might affect the cutting results (problems in the given example and DEKOIS 2.0 benchmark). Thank you for raising this problem again. I will double-check this example and the DEKOIS 2.0 benchmark and see whether this problem can cause a significant difference in test results.

[ndlongvn]

I'm looking forward to your result. If you have any newer results, please let me know. Thank you very much for your time and consideration.