Enable 3D Ligand Files

[rjrich]

When I used boltz-1 to predict the structure of a protein-ligand complex, the protein structure is fine, but the ligand pose is very distorted and rather remote from the pose in a reference x-ray crystal structure. The problem seems to be with using smiles strings as ligand files. Are there plans to enable using true 3D cooordinate ligand files such as SDF or MOL2? I have tried various conversions to smiles in attempts to preserve 3D and stereochemical information, but these have not worked satisfactorily. Thanks.

[jwohlwend]

Would you mind sharing with me your input config so I can take a look? Also does your ligand happen to have a CCD code I can compare against?

[rjrich]

Hi Jeremy, Many thanks for your reply to my issue post. I was not sure which input file(s) you might need, but I have attached a zip archive containing the input fasta for the protein and the two smiles strings for the two ligands. These input files were derived from PDB 6T4G. I hope that you discover that I have done something wrong and that boltz-1 can reproduce the crystal structure after all! Best wishes

…

On Thu, Nov 21, 2024 at 12:15 PM Jeremy Wohlwend ***@***.***> wrote: Would you mind sharing with me your input config so I can take a look? Also does your ligand happen to have a CCD code I can compare against? — Reply to this email directly, view it on GitHub <#34 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AAP6T7L7R3NUI4MAOJBDWY32BYII7AVCNFSM6AAAAABSFS7CLCVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZDIOJRHAZDGMJUGE> . You are receiving this because you authored the thread.Message ID: ***@***.***>

[rjrich]

Hi Jeremy, My profound apologies. I believe I sent you incorrect input files. The correct files for the protein and 2 ligands could be derived from PDB 6T4G. However, please beware, as it appears that the PDB has an incorrect FASTA file for this protein. The correct FASTA from Uniprot is attached. I have also attached the ligand files in SDF format. Rudy

…

On Thu, Nov 21, 2024 at 12:15 PM Jeremy Wohlwend ***@***.***> wrote: Would you mind sharing with me your input config so I can take a look? Also does your ligand happen to have a CCD code I can compare against? — Reply to this email directly, view it on GitHub <#34 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AAP6T7L7R3NUI4MAOJBDWY32BYII7AVCNFSM6AAAAABSFS7CLCVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZDIOJRHAZDGMJUGE> . You are receiving this because you authored the thread.Message ID: ***@***.***>

[RuikangSun]

The problem seems to be with using smiles strings as ligand files.

Sorry but I'm not sure if using 3D ligand files can make any meaningful results. One of the purpose of AlphaFold 3 and Boltz is predicting protein-ligand interaction. 3D ligand input generated in silico is not equal to conformation in protein binding pocket.
SMILES string can have stereochemical information (L-Proline is C(O)(=O)[C@@H]1CCCN1 for example). I have also tried generating protein-ligand interaction with SMILES as input. Although the binding pose is terribly wrong, the ligand structure (bond lenth, bond angle, etc) is reasonable.

Cheers!

[rjrich]

Perhaps at least part of the problem is getting the SMILES string correct, especially for those of us who are not fluent in SMILES. For example, the PubChem SMILES string for ethinylestradiol (EE2) is as follows:
(C[C@]12CC[C@H]3[C@H]([C@@H]1CC[C@]2(C#C)O)CCC4=C3C=CC(=C4)O
If that string is opened in a 3D viewer, one can see that the structure of the 5-membered ring is incorrect.

Moreover, I found various incorrect versions of SMILES strings for this compound on other databases.
In addition, if the 3D SDF file is downloaded from PubChem and converted to SMILES using OpenBabel, the string is not correct.

However, the ChEMBl string appears to be correct:
[H][C@]12CC[C@@]3(C)[C@@]([H])(CC[C@@]3(O)C#C)[C@]1([H])CCc1cc(O)ccc21

When entering the SMILES strings in the yaml file for submitting a boltz job, I have enclosed them in single quotes ('....').

Thus far, although boltz is superb at predicting my protein receptor structures, ligands are still distorted and not in agreement with crystal structures of complexes. For example, the ethinyl group in EE2 was converted to an ethyl group and the entire ligand was flipped 180 degrees around its long axis. In contrast, conventional docking with Vina reproduced the crystal structure almost exactly. Nevertheless, boltz is a spectacular and most welcome achievement in rapid and accurate protein structure prediction starting with only an amino acid sequence file!

[rjrich]

Would you mind sharing with me your input config so I can take a look? Also does your ligand happen to have a CCD code I can compare against?

Per your request, please find attached my input yaml file in a zip archive. This contains the CCD codes for the two ligands. The reference PDB ID for the complex is 4X1G.
4x1g_ee2_nonat_ccd.yaml.zip

[benf549]

I've also wondered about this feature and 3D ligand inputs can additionally be useful by keeping the atom order constant. Remapping between the order in one stage of a pipeline and another can be quite tedious...