David Ryan Koes @dkoes.compstruct.org · Nov 25 This is why I've been looking at profilin - there are ALS linked mutations that open up a destabilizing pocket (yellow). Boltz-1 gets the pocket right (unsurprising - it is in the training set), but not a single ligand gets docked there. Maybe this means it is "undruggable", but I doubt it.
29 likes 7 replies
Replies
That’s interesting! I struggle with tons of false positives in protein-protein AlphaFold-type simulations—since it seems to clobber its way through the interaction on the assumption it does interact. Maybe ligands have a false negative problem? I don’t have an intuition for it.
Do you have an intuition if any of those poses & binding sites predicted by Boltz-1 are remotely realistic?
This is great. Do you know if AF3 and Chai-1 behave differently?
Interesting! Would also be interested in seeing what happens when the prediction is conditioned on some residue(s) from the yellow pocket! (we should be adding support for pocket conditioning very soon)
I don’t believe in the word “undruggable”…only “recalcitrant”.
Could you just design a de novo small molecule or peptide binder? I know some models that will design drugs that fit into weird pockets, but you don’t even need a pocket if you’re willing to use peptides. I can design something even for flat interfaces. DM me if you want to discuss.
Quite interesting. I've seen some bias towards exposed vs buried pockets on AF3 and Chai-1 (didn't try this on Boltz-1 yet). Maybe we have some an intrinsic bias pocket types in the training dataset? Not sure if anyone has already investigated this, actually.