Thanks, @dkoes.compstruct.org ! Agree in principle. Will surely consider this for the v3 release in the Q1 next year!

Happy holidays!! 🎆

These are more reflective of counts in PDB than anything else.

If one wants a subset of ligands that fit a particular interest - lots of annotations are provided, including tutorials on how-to!

Please consider reviewing ligand stats for the actual MLSB test set (346) set of de-leaked PLI systems

Thanks for the insightful review! Fully agree on the value of reasonable defaults. However, what is a universally "good" train/val/test set is a bigger discussion than permitted in 300 chars. We have some ideas in the works, & would be happy to discuss specific proposals via discord or other means.

As always with @mmbronstein.bsky.social & Bruno Correia.