nan in loss with those molecules

[thegodone]

basically due to the features.GasteigerCharge() that can be null few molecules with like whit "P" atoms will produce nans during training. Can you replace the nan features for those molecules by 0.0 in such case ?

[akensert]

This is the current implementation:

@atom_features.register(name='gasteiger_charge')
class GasteigerCharge(Feature):
    def __call__(self, atom: Chem.Atom) -> float:
        mol = atom.GetOwningMol()
        rdPartialCharges.ComputeGasteigerCharges(mol)
        val = atom.GetDoubleProp('_GasteigerCharge')
        return val if val is not None else 0.0

Isn't it already taking care of it? Can you give me an example molecule that causes the issue and I'll look into it.

[thegodone]

this is a molecule that failed : ['OCPH(CO)CO'] I add GasteigerCharges and it goes to nan I remove it and it works. I think the vales are not None but Nan.

EDIT string was interpreted by github here the correct smiles:

'OC[PH](CO)(CO)CO'

[akensert]

smiles = 'OCPH(CO)CO'
mol = molgraph.chemistry.molecule_from_string(smiles)

ValueError: Inputted `molecule` (None) is invalid

Cannot convert SMILES to an rdkit mol object. Also tried with rdkit.Chem.MolFromSmiles(smiles, sanitize=False)

[thegodone]

I fix the smiles in github issue page see upper