3 ms·

> But the really important conclusion of the paper, which I truly hope bears out, is that AF2 models are useful for finding binders to GPCRs. What seems more l

by timr 3y ago

> But the really important conclusion of the paper, which I truly hope bears out, is that AF2 models are useful for finding binders to GPCRs. What seems more likely to you: that the (obviously worse [1]) structures produced by alphafold are somehow magically capturing protein dynamics, or…the virtual screening method (DOCK) used by the paper is relatively insensitive to the quality of the underlying model? They say right there in the article that neither screen was obviously enriched versus the other (the hit rate was around 50% for both branches, which was essentially the same as for the very different retrospective screening method in the same paper [2]). I’m putting my money on the parsimonious answer. Running structures through a large enough virtual screen will almost always find something that binds, even if the starting model isn't great. [1] at least in the case of the x-ray structure. Using a cryoEM structure is sort of interesting, since they’re blurry anyway. [2] the actual paper: https://www.biorxiv.org/content/10.1101/2023.12.20.572662v1.full.pdf https://www.biorxiv.org/content/10.1101/2023.12.20.572662v1....

dekhn 3y ago
I don't have a good answer. I don't think AF2 captures protein dynamics in a direct way. And I think docking programs have inadequate loss functions and sampling to really find accurate answers. I suppose AF2 could be modified to produce structures that are well mapped to docking programs. I think you're right, that running lots of structures through virtual screens will find lots of binders. In the middle days of protein design (well before AF2), people went from not being able to design anything, to being able to easily design "rocks"- proteins that were exceptionally stable, but didn't do anything interesting. Their loss function that they maximized for was stability. But that comes at a cost: high stability seems to prevent lability (which is useful for enzymes). And I think that if you just pass enough stuff through a good-enough program, you'll find lots of binders, but those binders won't actually be useful binders. We already know from decades of work that we can find lots of nanomolar and picomolar binders that don't do anything worthwhile (or have side effects). My conclusion is: we may very well be on the tip of solving another bottleneck in the drug discovery pipeline, or, we might just be bumbling around in a big space and found a nice local minimum. I don't think AF2 or similar systems on their own are going to completely solve the problem, and until DeepMind or somebody else creates something that can truly address the hard problems in human biology, the most we can hope for is finding more effective new drugs more cheaply and somehow couple that with improvements in genetic personalization.

by timr 3y ago

dekhn 3y ago

I don't have a good answer. I don't think AF2 captures protein dynamics in a direct way. And I think docking programs have inadequate loss functions and sampling to really find accurate answers. I suppose AF2 could be modified to produce structures that are well mapped to docking programs. I think you're right, that running lots of structures through virtual screens will find lots of binders. In the middle days of protein design (well before AF2), people went from not being able to design anything, to being able to easily design "rocks"- proteins that were exceptionally stable, but didn't do anything interesting. Their loss function that they maximized for was stability. But that comes at a cost: high stability seems to prevent lability (which is useful for enzymes). And I think that if you just pass enough stuff through a good-enough program, you'll find lots of binders, but those binders won't actually be useful binders. We already know from decades of work that we can find lots of nanomolar and picomolar binders that don't do anything worthwhile (or have side effects). My conclusion is: we may very well be on the tip of solving another bottleneck in the drug discovery pipeline, or, we might just be bumbling around in a big space and found a nice local minimum. I don't think AF2 or similar systems on their own are going to completely solve the problem, and until DeepMind or somebody else creates something that can truly address the hard problems in human biology, the most we can hope for is finding more effective new drugs more cheaply and somehow couple that with improvements in genetic personalization.