Hacker News Reader

Comment by dekhn | Hacker News Reader

They found a bunch of potential molecules that bind to 5HT2A, that's cool.

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.

GPCRs are one of the largest (if not largest) drug targets and work (including my own) showed that modelling protein dynamics using expensive simulations was likely to be necessary to make accurate predictions of GPCR binding molecules. If this paper is correct, AF2 is actually able to recapitulate much of the detail that we previously believed required expensive dynamics simulations, which is about the best possible news because it suggests that we can omit these costly calculations.

It's almost certain now that AF2 will win the Nobel Prize, as this represents some of the most exciting breakthroughs in the past few decades.

View on HN

Even if it has to go through extensive vetting on the theory side, the ability of these models to improve the starting points for the harder process is very strong. For example, in daylight modeling - and this is 2-3 years ago mind you - it has been found that "AI" is perfect for getting 99% closer to the actual luminance in something like 01% the computation. So that even if you feed those "precalculated" models into the traditional raytracers, the improvement to the first guess is well, well worth it.

I don't really think there's an analogy between radiosity simulations and drug discovery. Drug discovery will probably remain fairly ad-hoc and stochastic for some time.

That's fair. I was merely talking about being able to make better "first guesses." Nowhere near the predictability of a much simpler phenomenon such as radiance mapping, of course.

And as the article says, >"There were a lot of models that we didn’t even try because we thought they were so bad,...”

To me the analogy is that generic AI techniques with minimal problem domain specific adaptations is providing huge boosts in efficiency on a wide variety of applications. It’s like the steam engine causing mechanical work to become vastly more efficient. AI in many ways is doing the same thing to intellectual work.

The long-anticipated "mental gym" industry! Where people pay to exercise their atrophied mental muscles the same way they exercise their barely-needed physical muscles now.

With all the associated bullshit and snake oil and romanticizing of the "good old natty days" too, of course.

Big $$$ industry, at least until our brave future does away with voluntary reproduction altogether, being the reason people bother to exercise in the first place.

What are all video games if not mental gyms?

mental cake and junk food

I don't think this analogy works. You spend energy (just a different kind) playing games.

Not everybody is playing factorio or caves of qud. A lot of people are doing color match and fetch quests. It's about as mentally engaging as walking.

you say this as if walking didn’t have plenty of cognitive benefits

I meant the cognitive load of ambulation itself. Not the self reflection time or exercise.

I do hope people choose a nice walk over color match games.

As an expert, do you think this might lead to easier discovery of much more selective GPCR ligands?

Agree finding new 5HT2A ligands is cool, what would be truly interesting is having a suite of totally selective ligands for the entire 5-HT receptor family.

I feel like a lot psychopharmacological research into a lot of mental disorder is severely hampered by

1) the complex interactions of the various serotonin receptors(and also with other receptors like D2)

2) the lack of truly selective ligands, necessitating a hodge podge of less selective agonists combined with selective antagonists to isolate the effects of specific receptor interactions(making it hard to do this sort of thing in humans). Or using knockout mice, which is also kind of dubious because you've now made a fundamentally different brain and are trying to compare it to a normal one. And that's sort of hit and miss.

> 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....

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.

> It's almost certain now that AF2 will win the Nobel Prize ... most exciting breakthrough

Unsure of the merits of this finding myself, but I'd be cautious given just this past week DeepMind was accused of publishing gibberish by Chemists: https://twitter.com/Robert_Palgrave/status/17443839652705816...

That is unrelated work by a different team. The merits of AF2 have been demonstrated already and accepted by the community (which has reproduced those results both theoretically and experimentally) although the community is still arguing about just how good AF2 (and its descendants) are.

I agree the chemistry paper does not look g ood but it's hard to say yet. Like I've said elsewhere DM and Google Researchers are under huge pressure to make their results look as good as possible. Already people have serious concerns about another Google paper using ML for IC layout.

Awesome. What are a couple examples of what this may enable (drugs that would target the same)?

I would extrapolate from existing GPCR-targeting drugs, but the list is so long I can't really give a survey. As a protein family, GPCRs do a lot of different things but basically it boils down to signal transduction: there's something outside a cell and the cell wants to know that so it can respond. GPCRs mediate the binding of the thing outside to the surface of the cell, and transduct the information into the cell, delivering a signal to the nucleus, where the transcriptional machinery is manipulated to make new proteins (or stop making proteins). I left out some steps. This allows the cell to be responsive to the outside in a secure way, without the outside molecule coming into the cell.

The 3d structure of GPCRs is intimately associated with the cell membrane, with one end sticking out into the extracellular environment, and another sticking out into the interior of the cell. Simply getting that sort of protein, which is not soluble in water, was a massive challenge and the first structures didn't arrive until the early 2000s, when people learned how to crystallize GPCRs using detergents and other solvents that simulated cell membranes.

So many diseases are caused by misregulation of signal trandsduction... let's take some examples. Vasopressin is a drug that also happens to be a natural body product. In some diseases, people don't make enough vasopressin to regular the kidneys properly, leading to a form of diabetes, and simply giving people more of it helps reduce the symptoms. I guess in this case (not certain), people have mutations in the GPCR that receives the vasopressin signal that attenuates the signal "too much" and by just dosing the patient with a lot more, more signal gets sent inside the cell.

I think at least 10 nobel prizes in medicine have been awarded for research related to GPCRs, which has uncovered a wide range of medically relevant knowledge of physiology. The NERSC supercomputer "Cori" was named for Gertie Cori, who helped discover some of the core mechanisms in human metabolism during the golden age of metabolic biochemistry.

I think the simplest way to think about it is being able to target specific GPCRs is like being in front of a switchboard with buttons and knobs controlling every critical detail of how human bodies manage and maintain themselves.

Every time I hear about some complex function of the cell it really makes me amazed at us being collections of trillions of them...

A better question would be what don't they enable. Something like 30+% of FDA approved drugs operate along that pathway. Besides for the boring stuff like being able to trigger cancer cells to kill themselves, we could have a molecule you inject into your eye that opens up your vision to be able to see infrared and ultraviolet light(and beyond).

It's almost certain now that AF2 will win the Nobel Prize, as this represents some of the most exciting breakthroughs in the past few decades.

It seems weird to give the peace prize to a model. What about the people who made the model, or is that what you mean?

Reply on news.ycombinator.com