Date: | 2023, December 19 |
Time: | 11:30 a. m. |
Place: | G29-018 |
Author: | Glauer, Martin |
Title: | Data-Efficient Graph Grammar Learning for Molecular Generation |
One piece of feedback we received for our ChEBI-based class prediction model highlighted missing
representation certain classes, particularly those with low membership count.
Recently, Guo et al. from IBM proposed a "data-friendly" solution for molecule synthesis,
which I want to discuss in this talk.
They claim that their approach can extract a productive graph grammar from a small set of molecules,
taking into account non-differentiable metrics like synthesizability.
These grammars are then utilized to generate new molecular structures that possess defining features
of the targeted class of molecules.
Remarkably, their method is effective even with classes comprising only 20 instances, and they assert
its competitiveness when using 171 samples against existing methods that use a dataset of 81,000 data points.