kexinhuang12345 / DeepPurpose

A Deep Learning Toolkit for DTI, Drug Property, PPI, DDI, Protein Function Prediction (Bioinformatics)
https://doi.org/10.1093/bioinformatics/btaa1005
BSD 3-Clause "New" or "Revised" License
974 stars 272 forks source link

[Question] Enantiomeric/Stereoisomeric compounds #98

Closed albertma-evotec closed 3 years ago

albertma-evotec commented 3 years ago

Are the encoding protocols sensitive to enantiomeric and stereoisomeric compounds? It matters because two isomers have different binding affinities. If they are going to have exactly the same encoding, it may cause problem in training.

kexinhuang12345 commented 3 years ago

Hi, it depends on the encoding schemes. For SMILES strings encoding, it should be sensitive as long as the SMILES string is sensitive and the training/testing set both have the same encoding scheme. But for molecular graph (MPNN), it is not. It is actually one of an active research directions to put chirality into the graph neural network: https://arxiv.org/abs/2012.00094