For benchmarking, we placed a lot of emphasis on distribution learning capabilities because this reflects the training objective of generative models. But we also show how downstream preference optimization can be used to further improve molecular properties.

(4/4)

These models are equipped with a few new features including:

1. protein side chain modeling
2. adaptive ligand sizes
3. confidence score
4. preference alignment

(3/4)

We (together with @igashov.bsky.social, @adobbelstein.bsky.social, Thomas, @mmbronstein.bsky.social, and Bruno) introduce two new models for target-conditioned drug design in 3D (DrugFlow and FlexFlow), which sample new molecules using a mixed continuous/discrete generative framework.

(2/4)

Compared to the preprint (biorxiv.org/content/10.1...), we added an optimised design pipeline using AlphaFold and LigandMPNN, and super cool tumour cell killing results from Maddalena.