This is a big fricking deal

You guys know that I'm normally in the 'wet blanket' crowd with scientific or technology news. Well, for once I'm going to withhold my pessimism. Because CASP just announced some extraordinary data from AlphaFold, the DeepMind protein folding effort:

https://www.sciencemag.org/news/2020...ein-structures

Long story short? It's really, really hard to predict the 3D structure of a protein in silico; the gold standard has typically been x-ray crystallography which is very hard indeed. People have been working on this problem for decades with modest improvements but definite limitations. Well, the data coming out today sure looks like they've blown away all previous attempts and have near-parity with x-ray crystallography using deep learning techniques.

What does this mean? Assuming this can be replicated in more proteins and can be done quickly and at scale, the pace of target discovery and rational drug design will have just increased dramatically. And studying new proteins (like, say, the capsule proteins on a novel virus) will get much faster, allowing for rapid identification of potential therapeutics.

I am absolutely shocked this happened so quickly, this has been a very very hard problem for a very long time. Kudos to DeepMind on this, and I can't wait to see what people come up with using their technology!

The age of AI will soon be upon us.

1.0 of this thing, alphastar, kicked everyone's ass at starcraft 2 earlier in the year.

Why is this in general chat? :confused:

One for Discussion shirley.

Quote:

---

Originally Posted by Timbuk2

Why is this in general chat? :confused:

One for Discussion shirley.

---

Well, it's generally awesome :p

Quote:

---

Originally Posted by Timbuk2

Why is this in general chat? :confused:

One for Discussion shirley.

---

I didn't think there was much to discuss unless we had any experts in protein folding who could provide further color. :o

And stop calling me shirley!

https://arstechnica.com/science/2020...ious-software/

Even bigger than AI-identified antibiotic candidate earlier this year :up: apart from, well *gestures at everything* this is a cool time to be alive. Has some implications for my work :)

Quote:

---

Originally Posted by Aimless

Even bigger than AI-identified drug candidates

---

Has that ever been a particularly big deal, though? I was under the impression that, while there was a lot of hype in press, people who actually work in the field consider it not much more than a tiny incremental improvement to the existing techniques.

Quote:

---

Originally Posted by BalticSailor

Has that ever been a particularly big deal, though? I was under the impression that, while there was a lot of hype in press, people who actually work in the field consider it not much more than a tiny incremental improvement to the existing techniques.

---

I've been pretty leery of in silico prediction for API selection against a target; I worked on a project a number of years ago where our collaborators were a deep pocketed pharma company with massive libraries of potential APIs, but their in silico prediction of efficacy did not perform very well vs. high throughput testing against specific targets (which itself was only so-so). Their prediction of potential safety/tox issues was a bit more helpful in that it allowed us to focus our efforts a bit.

I don't know how much these techniques have developed since then (this is maybe 5 years ago?) but there's definitely going to need to be some improvement and reduction to practice.

I'm much more excited about this; if they can demonstrate it works on a sufficiently varied population of proteins, it could be really helpful. Waiting to get a crystal structure is the worst.

Quote:

---

Originally Posted by BalticSailor

Has that ever been a particularly big deal, though? I was under the impression that, while there was a lot of hype in press, people who actually work in the field consider it not much more than a tiny incremental improvement to the existing techniques.

---

The AlphaFold development is genuinely revolutionary :up: what happened earlier this year (see also: https://www.quantamagazine.org/machi...ance-20200309/) was related to but much more exciting than the overblown startup hype being discussed (and rightly criticized) in the post you linked to—agnostic AI discovering drug candidates with novel mechanisms of action without guidance from investigators' assumptions. In doing so, AI not only speeds up drug development, but also helps us expand and deepen our understanding of disease, by guiding us to unexpected solutions we might not have thought to look for. Wrt drug discovery and development, these two types of breakthroughs will be able to complement one another.

Quote:

---

Originally Posted by BalticSailor

Has that ever been a particularly big deal, though? I was under the impression that, while there was a lot of hype in press, people who actually work in the field consider it not much more than a tiny incremental improvement to the existing techniques.

---

I don't know much about computational biology, but in general the field of AI has been making some truly massive strides lately. It stalled for a while the first decade of this millenium, but it's been outpacing almost all projections the past several years. So it's not hard for me to believe that AI solutions that weren't too impressive a year or two ago would be outperforming humans today.

Quote:

---

Originally Posted by Aimless

The AlphaFold development is genuinely revolutionary :up: what happened earlier this year (see also: https://www.quantamagazine.org/machi...ance-20200309/) was related to but much more exciting than the overblown startup hype being discussed (and rightly criticized) in the post you linked to—agnostic AI discovering drug candidates with novel mechanisms of action without guidance from investigators' assumptions. In doing so, AI not only speeds up drug development, but also helps us expand and deepen our understanding of disease, by guiding us to unexpected solutions we might not have thought to look for. Wrt drug discovery and development, these two types of breakthroughs will be able to complement one another.

---

I had missed that somehow; I stand corrected then, it is indeed quite a big deal.
The AlphaFold development is incredible. Reducing the need to rely on getting an X-ray structure of proteins alone is a great achievement. I've only worked with small molecules, and even then getting decent crystals is really difficult. The best method in my experience has been "dissolve the compound in whatever, leave the solution in a not-so-tightly-stoppered flask, go on vacation, return having forgotten about the whole deal, re-discover the flask whenever the next lab cleaning comes around". Proteins, I've heard, are really way over there in the "more voodoo than actual reliable methods" realm compared to small molecules.

Quote:

---

Originally Posted by BalticSailor

I had missed that somehow; I stand corrected then, it is indeed quite a big deal.
The AlphaFold development is incredible. Reducing the need to rely on getting an X-ray structure of proteins alone is a great achievement. I've only worked with small molecules, and even then getting decent crystals is really difficult. The best method in my experience has been "dissolve the compound in whatever, leave the solution in a not-so-tightly-stoppered flask, go on vacation, return having forgotten about the whole deal, re-discover the flask whenever the next lab cleaning comes around". Proteins, I've heard, are really way over there in the "more voodoo than actual reliable methods" realm compared to small molecules.

---

Perhaps not witchcraft, but definitely a craft. Even aside from basic skill requirements, it's [even more] unpredictably difficult to obtain—and manage—adequate crystals, often requiring very extensive and unpredictable—and tedious—optimization. I have no idea how something like AlphaFold might be incorporated into the process of determining 3d structure for previously unsolved cases; I feel like it might require the application of additional machine learning tools to the diffraction data, because I don't think AI is about to eliminate those steps any time soon. It's possible taking vacations after/during each step might help. Fortunately, it's not my field, so I don't have to feel bad about Melon Usking about the possibilities :)

Quote:

---

Originally Posted by BalticSailor

Reducing the need to rely on getting an X-ray structure of proteins alone is a great achievement. I've only worked with small molecules, and even then getting decent crystals is really difficult. The best method in my experience has been "dissolve the compound in whatever, leave the solution in a not-so-tightly-stoppered flask, go on vacation, return having forgotten about the whole deal, re-discover the flask whenever the next lab cleaning comes around". Proteins, I've heard, are really way over there in the "more voodoo than actual reliable methods" realm compared to small molecules.

---

Amen. A friend of a friend literally got their PhD because some samples they forgot about in a drawer had crystallized and yielded a decent structure. I would not be surprised if black magic was involved in getting most protein crystal structures. Don't even start me on how they do it for transmembrane proteins or the like.