Double and triple bonds not correctly recognized (Problem with reading smiles or potentials?)

[OlfwayAdbayIgbay]

Using the following smiles: C#CCN=C1CCN(C(=O)OC(C)(C)C)C1, the model creates a molecule that has neither the necessary planarity/double bonds nor the triple bond that it should have.

This is what the molecule should look like:

While this is what it looks like:

[Yoik]

I’ve also encountered this issue. Based on my analysis, there are likely two root causes. First, the output CIF and PDB files only describe the bonding information between atoms in the small molecule but lack bond order information, which causes PyMOL to fail in correctly displaying single/double bonds. When using #268 to output the SDF format for the small molecule, the single/double bonds may appear visually correct, but the bond angles resulting from these bonds are actually nonexistent. This points to the second issue: during the denoising/diffusion process, there might be no constraints enforcing proper bond orders. This is my current understanding of the problem, but my coding skills are insufficient to address or fix these issues.

[indazole]

I'm seeing this as well. I agree with Yoik's second comment, it appears some bond order information, on certain compounds, is getting lost during the denoising/diffusion process. In my case, compounds with exocyclic double bonds with explicit E or Z geometry may be inverted following prediction, eg: E flips to Z. Prediction using the same compound with two (highly homologous) proteins can give two different, E and Z, complexes. I've tried many different canonical and non-canonical smiles representations of the same compound, as well as adding explicit hydrogens, all to no avail.