search

mattragoza / LiGAN

Deep generative models of 3D grids for structure-based drug discovery
GNU General Public License v2.0
221 stars 43 forks source link

feedback about generated molecules #45

Closed ChengchengGutenTag closed 2 years ago

ChengchengGutenTag commented 2 years ago

Hi, thank you for publishing this fantastic work. May I give some feedback about the generated molecules? I tested liGAN (posterior and prior model) on our protein target with reference molecule, and the test sets in the demo. When I view the SMILES of the generated molecules, it seems that a high ratio of the molecules have weird structures. Is it normal? For example, it occurs quite often that the structures have multi-fused cyclic compounds, which are not so realistic. Another example is that a super big ring (which may contain 7 or 8 atoms) shows up.

Bests Chencheng

mattragoza commented 2 years ago

The approach we recommend is to generate many samples (100-1000) and then filter the molecules based on some criteria such as drug-likeness. You can also get more realistic molecules by reducing the variability factor, though they tend to be more similar to the reference molecule.

ChengchengGutenTag commented 2 years ago

Thank you very much for your kind reply.

Bests