At least this it is what I wish for boltz.

And I think this is really what we should be aiming for! Not be limited by the amount of data, but only by how protein-ligand interactions are represented.

Yes, but it should not be bound by that. For example, the AutoDock-Vina scoring function (6 parameters) was fitted using only 800 complexes. However, I know performance mainly depends on having the correct pocket conformation (+ other factors), but not because it didn’t see a particular complex.

About the val&test, no clue how many truly novel protein-ligand complexes were deposited since 2023. This would directly affect our capacity to benchmark these methods.

That’s a bit problematic and means that we reached the limit of this type of architecture, right? If a model has to see first examples of a system for making predictions, then it will always lag behind the field, such as the discovery of a new pocket, new binding mode or chemistry.

So you changed the cutoff date to 2023-06-01. Why? Are you scared of Runs N’Poses? 😏

I want to thank my co-authors @jeeberhardt.bsky.social, @torstenschwede.bsky.social, @ninjani.bsky.social and all of our collaborators! RunsN’ Poses builds on PLINDER and OpenStructure—this work wouldn’t be possible without them!
Also thanks to @rokbreznikar.bsky.social for this amazing logo! 9/9

We actually had a similar benchmark (with LDDT-PLI) in the same CASP15 issue a while ago (onlinelibrary.wiley.com/doi/10.1002/..., Fig3B) conclusions were (1) pocket detection needed for physics-based (2) DL models overfit (3) nothing performs on non "re-docking". Was my main inspiration for PLINDER

For future reference, this is the slide (without the bottom cropped) that should have been posted here. Boltz-1 and AF3 are actually performing similarly on the Chymase and Mpro datasets, and differ only on the Autotaxin dataset for still unknown reasons.

Hi @ddelalamo.bsky.social unfortunately, this paper from Jain et al. contains falsehoods, misleading comparisons, seemingly deliberate omissions, and is written in a tone not intended as a serious research paper. Please see our detailed response: www.linkedin.com/pulse/respon...

Ok, make sense. Thanks!

Nice work! Any reason why using AutoDock-Vina 1.1.2 instead of the last version (1.2.5), and prepare PDBQT files with Meeko (which allow you to convert it back to MOL)?

Nope, no allosteric binders. The only molecule binding outside the canonical pocket was in the Mpro dataset, and it was stabilized by crystallographic contacts.