search

aspuru-guzik-group / selfies

Robust representation of semantically constrained graphs, in particular for molecules in chemistry
Apache License 2.0
660 stars 127 forks source link

Model SELFIES VAE #121

Open KatynaSada opened 2 weeks ago

KatynaSada commented 2 weeks ago

Hi, do you have the trained SELFIES VAE available for download somewhere?

MarioKrenn6240 commented 2 weeks ago

Hi @KatynaSada, we dont have the trained version, but the code is here, and training it on QM9 is very quick on a standard laptop just on CPU.

KatynaSada commented 2 weeks ago

Thanks!! I thought it would require many resources to train, but I was easily able to train it. Another question, have you obtained good results training the model with the ZINC-250k drugs dataset (I noticed you also had that dataset)???

MarioKrenn6240 commented 2 weeks ago

For larger molecules, i think VAEs are not the best options. I would recommend you to have a look at the STONED algorithm based on SELFIES, or an extended representation that handles groups efficiently (group selfies). In general we discuss various extensions and applications a lot in this community review.