That's true 😁 We have kernel density estimate plots and statistical tests in Appendix (Fig. 18 and 19)

While DrugFlow is really good in distribution learning, we also show that it can be easily tuned to generate molecules with optimised target properties. For example, we can boost docking efficiency and synthetic accessibility, or reduce the level of swill in samples @wpwalters.bsky.social

Our adaptive size mechanism allows DrugFlow to remove excessive atoms on the fly, adjusting the molecular size to the provided pocket constraints. In our experiments, we demonstrate that DrugFlow can avoid steric clashes if the provided molecular size is too big.

DrugFlow's atom-level confidence score is able to detect unlikely samples and correlates with molecular size and docking efficiency. In particular, DrugFlow tends to assign low confidence in cases when it introduces steric clashes with the protein or produces strange geometries.

(2/2) SynFlowNet: Design of Diverse and Novel Molecules with Synthesis Constraints openreview.net/forum?id=uvH...

(1/2) Multi-domain Distribution Learning for De Novo Drug Design openreview.net/forum?id=g3V...