Is it correct to have warnings of unknown value in feature.py? about torchdrug

[KiddoZhu]

Hi! This is a desired behavior. The feature function we use tokenizes rare atoms as unknown values, similar to the convention of representing rare words with [UNK] in NLP. This helps the model better generalizes across rare atoms. The warning is output to notify the user what kinds of rare atoms are encountered. We will revise the warning message to reduce ambiguity.

[schinto]

Many thanks.

Is there already the functionality to get a list of SMILES which lead to RDKit error?

[KiddoZhu]

This is actually a warning, rather than an error in RDKit. RDKit prints this as an error, but it can still return the nearest valid molecule of the SMILES. If RDKit fails to parse any SMILES, it will return an empty molecule, and in that case TorchDrug will log the SMILES on the screen. We may further wrap this kind of RDKit warnings for better user experience. You can safely ignore these cases if it happens less than 1%.