Question about pre-trained ligand embeddings by Unimol tools

[srdjan125]

Summary

Hello,

I am using the pre-trained ligand embeddings from Uni-Mol unimol_tools. The documentation states that the atom-level representations align with rdkit mol.GetAtoms().

Could you please clarify how SMILES strings are processed internally by Uni-Mol tools? Specifically, do you apply any atom reordering or special sanitization steps that could affect atom order? I want to ensure that the ordering of atom types and embeddings corresponds exactly to the atom order in my own mol2/sdf files.

Thank you very much for your help!

Best regards.

Uni-Mol Version

Uni-Mol

Details

Hello,

I am using the pre-trained ligand embeddings from Uni-Mol unimol_tools. The documentation states that the atom-level representations align with rdkit mol.GetAtoms().

Could you please clarify how SMILES strings are processed internally by Uni-Mol tools? Specifically, do you apply any atom reordering or special sanitization steps that could affect atom order? I want to ensure that the ordering of atom types and embeddings corresponds exactly to the atom order in my mol2/sdf files.

Thank you very much for your help!

Best regards,

[Naplessss]

Hi, if you're using MOL2 or SDF files, you can directly take the atom and coordinate list as input, which ensures the atom order aligns with your original file. However, if you're using SMILES as input, the atom order may differ.

AND for details of SMILES processed to atoms/coordinates, you can reference: https://github.com/deepmodeling/unimol_tools/blob/main/unimol_tools/data/conformer.py#L233
we add H before use MMFF generate coordinates.

mol = Chem.MolFromSmiles(smi)
mol = AllChem.AddHs(mol)
atoms = [atom.GetSymbol() for atom in mol.GetAtoms()]

[srdjan125]

Hello @Naplessss thanx for answer.

Is there also a way to generate Uni-Mol embeddings by providing as input atom types + coords from mol2 files?

In your tutorial it was said that I can create embeddings by providing SMILES as input (miles_list = [smiles]
unimol_repr = clf.get_repr(smiles_list, return_atomic_reprs=True)). Actually I just found out on other documentation page (https://unimol.readthedocs.io/en/yn-docs/train.html#unimol_tools.predictor.UniMolRepr):

data –
str, dict or list, default=None, input data for unimol.
str: smiles string or path to a smiles file.
dict: custom conformers, should take atoms and coordinates as input.

If I do that, is it enough to provide dictionary of non-H atoms and their coordinates, or I must provide list of all atoms (and then use remove_hs=True which will remove H-atoms but not pertube order of other atoms?

Thanx for help.

[Naplessss]

If I do that, is it enough to provide dictionary of non-H atoms and their coordinates, or I must provide list of all atoms (and then use remove_hs=True which will remove H-atoms but not pertube order of other atoms?
I think both approaches should give the same result. If you set remove_hs=True, the hydrogen atoms will be removed without perturbing the order of the remaining atoms, and then the model will be fitted using the processed coordinates, so whether you provide the dictionary of non-H atoms directly, or use remove_hs=True on the full atom list, the outcome should be equivalent.

[Naplessss]

Oh, sorry for the confusion — you do need to make sure remove_hs=True when loading the weights for the non-H version of UniMol.
That’s a good question! We’ll make sure to clarify this in the parameter settings.

[Naplessss]

And we already archieved this repo part of unimol-tools, you can refer https://github.com/deepmodeling/unimol_tools for any updates and new features.

[srdjan125]

@Naplessss ok thanx. Eventually I have provided full list of atoms from mol2 files and then used remove_hs=True. I was not sure whether model was trained on full-atom structure with H-atoms, so if I would provide it only non-H atoms model would be confused and think that the molecule has no hydrogen atoms at all.