nvmolkit.embedMolecules does not generate a conformer for every molecule in the set

[Yannlefrancois]

Hi,
i'm using nvmolkit to embeded conformation in a large amount of molecule, i'm trying my current code on a small subset of molecule and i see that it only embeded comformation in part of the molecules.

here is my code

import nvmolkit.embedMolecules as embed
from rdkit import Chem
from rdkit.Chem import Descriptors
from rdkit.Chem import AllChem

    suppl=Chem.MultithreadedSmilesMolSupplier(f"{arg.temp_file_dir}/{file}", "\t", 0, 1, True, True,numWriterThreads=48)
    
    
    # create the list in advance to avoid unecessary memory requirement upon list expension using append

    mol_list=[None for i in range(arg.batch_size)]
    mol_id=[None for i in range(arg.batch_size)]
    mol_counter=-1
    for mol_1 in suppl:
        mol_counter+=1
        if mol_1 is not None:
            try:
                mol_list[mol_counter]=Chem.rdmolops.AddHs(mol_1)
                mol_id[mol_counter]=mol_1.GetProp(arg.property_name)
            except :
                pass
    


        params = Chem.rdDistGeom.ETKDG()
        params.useRandomCoords = True
        hardware_opts = HardwareOptions(
        preprocessingThreads=48,
        batchSize=-1,
        batchesPerGpu=-1,
        gpuIds=[0, 1])
        embed.EmbedMolecules(mol_list, params, confsPerMolecule=1, hardwareOptions=hardware_opts)

and if i check with mol.GetNumConformers() a non negligable amount as 0 500 out of 100000

is there anything that could explain that ?

rdkit version = 2026.03.1

nvmolkit 0.5.0

best
yann

[scal444]

Hi Yann, thanks for the report.

Is it consistently the same molecules? Does RDKit also fail on these molecules?

In general both RDKit and nvMolKit can definitely fail on poorly formed molecules, molecules where AddHs failed, or occasionally on good molecules that are hard to generate. Part of what controls success rate for the latter case is the max attempts param and the number of conformers you are trying to generate.

If you can provide me reproducing molecules or point me to where you're sourcing them from, I can follow up to see if it's a real bug!

[Yannlefrancois]

test1.txt
test2.txt
test3.txt

test.cxsmiles.txt

Test1 test2 test3 are three run with all the molecule who did not have a conformer after the embedding

and the test.cxsmiles is my subsample of my dataset.

i tried incresing maxIterations to 50k and i still have compound who failed but i dont have error message.
So i supposed it's just regular fails.

best
yann

[scal444]

I'll go over the test.cxsmiles.txt dataset to see if there are any statistical differences between success rates on rdkit/nvmolkit and get back to you, thanks

[scal444]

Ok turned out we had a real bug. RDKit > 2024.9 has a different treatment of stereocenter checks. I thought we had this in lockstep, but there was an error in the implementation. The consequence is that we have a 10x higher failure rate of one of the ETKDG pipeline steps than RDKit. If you have enough retries, it was hidden, but probably led to worse performance, since any conformer would require more retries.

[scal444]

Should be fixed in v0.5.1 -