Inquiry: Does Shepherd Support Scaffold Decoration?

[Manjubn777]

Hello,

I have been exploring its capabilities and wanted to ask whether scaffold decoration is currently supported or planned.

By scaffold decoration in my use case, I mean providing a fixed carbon-based core scaffold and expecting the model to:

1. Keep the scaffold structure intact
2. Replace atoms within the scaffold (substituting an carbon with a nitrogen)
3. Decorate the scaffold by attaching new atoms or groups at chemically reasonable positions

Does Shepherd currently support this kind of scaffold-anchored generation? If not, are there any suggested workarounds or plans for incorporating it?

Thank you!

[kentoabeywardane]

Hello,

Apologies for the late response. We currently do not support scaffold decoration.

To do so in the manner that you described, you may consider adjusting the inference procedure to inpaint the scaffold itself. Say your complete molecule is $x_1$ with $N_{x_1}$ atoms and a scaffold $x_{1,\text{scaffold}}$ where $N_{1, \text{scaffold}} < N_{x_1}$.

Potential paths forward:

Maintain the exact scaffold

• Inpaint both the atom types and positions for $x_{1,\text{scaffold}}$, in addition to the interaction profiles.

Keep the scaffold structure intact, but allow atom substitutions

• You could inpaint only the atom positions, and let the model infer what the atom types are. However, note that this strategy is implicitly biased in a couple of ways: 1) molecular geometry may leak information about the atom types (i.e., differences in bond lengths) and 2) conditioning on the interaction profile of the scaffold may also leak information about the atom identities of the scaffold
• Both could potentially be mitigated by stopping inpainting early for $x_{1,3}$.
• You must also be sure not to exactly inpaint the pharmacophore features of the scaffold that you condition on so that it can sample different atoms.

In both cases, as long as $N_{x_1} > N_{1, \text{scaffold}}$, the model will automatically attach new atoms or groups at chemically reasonable positions on the scaffold to match the conditioning interaction profile.

Let me know if you have any further questions or suggestions.

[aqemia-bill-tatsis]

That will be a very interesting application of Shepherd!
Will it be possible to share an example with code? Starting from a reference ligand that is not included in your data.
Thanks :)

[adrienchaton]

following!

@kentoabeywardane I am trying the previous SQUID model to generate shape-similar molecules to my own query molecule.
It would be great to have examples on how to do similar tasks with SHEPERD, if possible, for example:

• conditioning on a custom molecule which is not part of the preprocessed data
• inpainting a user-provided fragment

[kentoabeywardane]

I've added a couple examples notebooks found in the examples folder. Please let me know if you have any more questions.

[adrienchaton]

@kentoabeywardane this is great, the two notebooks you added look very useful!
I will try them as soon as possible and let you know if something is still unclear or not working.
Thanks a lot for the examples!

[adrienchaton]

@kentoabeywardane thanks again, the notebooks are very useful

conditional generation runs through, following your instructions, I input my query molecule in SMILES format and let embed_conformer_from_smiles prepare it

about atom inpainting, you start from a .sdf with ligand extracted from the PDB
mol = Chem.SupplierFromFilename('./data/WX7.sdf', removeHs=False)[0]

when I provide my molecule as .sdf, should I make sure it has hydrogens using e.g. Chem.AddHs?

[kentoabeywardane]

@adrienchaton Yes, ShEPhERD was trained on molecules with explicit hydrogens so the conditioning molecule should also include hydrogens.

If your .sdf file does not include hydrogens, I suggest using the Chem.AddHs(addCoords=True).