Forums  > Off-Topic  > AlphaFold2  
     
Page 1 of 1
Display using:  

Maggette


Total Posts: 1279
Joined: Jun 2007
 
Posted: 2020-12-03 11:13
I am not med/bio savvy.

Any insights on how groundbreaking that really is?
Thx

Ich kam hierher und sah dich und deine Leute lächeln, und sagte mir: Maggette, scheiss auf den small talk, lass lieber deine Fäuste sprechen...

chiral3
Founding Member

Total Posts: 5190
Joined: Mar 2004
 
Posted: 2020-12-03 14:20
I am not an expert. I've messed around with these things at various times in my life, tangentially touched, and almost took a job 20 years ago working on the problem with a co called schrodinger, but I don't have any real substantive understanding. My read on what's probably the same news you saw the other day is they hit a milestone with nobody really around to see what they did. Folding is a bit like fusion - when many-flop supercomputers became more accessible the problems were supposed to fall quickly. That was twenty years ago. I don't think the experts really know what afold2 accomplished. It could be completely theoretical or idealized at this point. It struck me like a D-wave press release.

Nonius is Satoshi Nakamoto. 物の哀れ

ronin


Total Posts: 630
Joined: May 2006
 
Posted: 2020-12-03 14:37
Like @ch3, I haven't really looked at this stuff in more than 15 years or so. Still, my 2p.

There are two steps to what they do. Step 1 is pretty traditional bioinformatics - you pick bits of the sequence that you already know about, and you put them together. That's really not new - there were websites doing it for free, in seconds, 20 years ago.

The new bit is the second part, where you do some constrained optimization to find the optimal geometrical packing of these bits you found. That's new, and that's the magic ingredient. Only they don't really do optimization, they trained a dnn to do it.

Based on what they write, the dnn sems to be a bit overtrained. It's really, really good at predicting x-ray structures of protein crystals. But it's sh.t at predicting nmr structures (i.e. proteins before you pack them into crystals) and interactions. The thing is, protein crystals aren't a thing in their own right. Humans make them because they are easier to study.

So it's a big step, but it answers the question a bit too literally. And the question it answers is the ever so slightly the wrong question.

In the immortal words of Homer Simpson, I give them partial credit.

"There is a SIX am?" -- Arthur

chiral3
Founding Member

Total Posts: 5190
Joined: Mar 2004
 
Posted: 2020-12-03 15:10
Isn't it really about the CASP (Critical Assessment of protein Structure Prediction) score? I don't know enough about it.

Nonius is Satoshi Nakamoto. 物の哀れ

ronin


Total Posts: 630
Joined: May 2006
 
Posted: 2020-12-03 15:22
Yes, CASP is a blind test of prediction power. They did really well in it this year.

So it is a big step.

But, like in all science, it's what they can't do that is more important.

"There is a SIX am?" -- Arthur

Strange


Total Posts: 1674
Joined: Jun 2004
 
Posted: 2020-12-03 16:40
I read the paper, its pretty interesting but it's not going to bring the new golden age of enzyme engineering or something like that. It's been a while since I looked at it, but that is very close to what I worked on for my PhD (my interest was in discovering the effect of minor mutations on protein folding).

The general idea is that it's relatively easy to predict which portions of the protein will form an alpha helix, which will form a beta-sheet and which are going to stay as loops. These predictions essentially produce structures that looks like a bunch of wine corks and bottle labels glued to a string. That's called a "secondary structure" and it's been around for decades (more or less since the Great Communist Scientist discovered them).

Figuring out how these sheets and cylinders fit together to form a tertiary structure is very hard as the search space is extremely large (in fact, not every protein will even fold back to it's original shape after it's been denatured). What these guys are doing they are looking at the known structures (which are discovered through protein crystallography, an extremely tedious process) and using whatever ML technique to minimize the search space. In some sense, this is similar to many other problems that were recently solved by ML - it's not really a scientific breakthrough per se, but rather a combination of a critical amount of data, increases in computational power and improvements in smart optimization/search.

This said, it's likely will change the approach to crystallography, now that you have a pretty good guess at the structure.

'Progress just means bad things happen faster.’

chiral3
Founding Member

Total Posts: 5190
Joined: Mar 2004
 
Posted: 2020-12-03 16:54
So Strange, this is not dissimilar to the relationship in, say, HEP, where theory reduces a huge space and tries to get experimentalists to use their results? That would make sense. It's a breakthrough in a theory problem, but to have it realized others need to utilize the outputs, right?

Nonius is Satoshi Nakamoto. 物の哀れ

nikol


Total Posts: 1279
Joined: Jun 2005
 
Posted: 2020-12-03 17:27
@Strange

"is very hard as the search space is extremely large"

Can this research be automated? Robotic process automatization in pharma, where robotic hand places all combinatorics of ingredients into test-tubes, is already accelerating research time and quality (no errors).
To very extreme: perhaps, if somehow to connect dnn to that robotic can it get really fast? Or everything will be messed up...

Same disclaimer: I'm from HEP, so not an expert.

... What is a man
If his chief good and market of his time
Be but to sleep and feed? (c)

Strange


Total Posts: 1674
Joined: Jun 2004
 
Posted: 2020-12-03 17:30
@chiral3, HEP is a good analogy.

Here is the general issue - an "intermediate output" of the X-ray diffraction is a convolution of the actual structure onto itself. Regular (like polymer) crystallography gets through the deconvolution step pretty easily because it's easy to guess what the structure will look like.

Protein crystallography is a nasty and tedious process because the protein structure is so complex that it's impossible to unravel the convolution analytically. At the moment, what people do is they add high diffraction atoms to the specific points in the protein structure so they have a sort of anchoring points to perform the deconvolution and arrive at the structure. It's a tricky process, for example some proteins refuse to fold the same way once you add modified amino acids. Now that they can guess what the structure will look like, it's possible that solving for the structures will take weeks as opposed to months. It's also possible that large structures like ion chanells will be finally solved with the help of this technique.

'Progress just means bad things happen faster.’

chiral3
Founding Member

Total Posts: 5190
Joined: Mar 2004
 
Posted: 2020-12-03 17:36
The number i had in my head was about 2x the order of mag the # of atoms in the universe, or 10^140-10^150 in terms of the brute force problem, hence the dimensional reduction work.

Nonius is Satoshi Nakamoto. 物の哀れ

Strange


Total Posts: 1674
Joined: Jun 2004
 
Posted: 2020-12-03 17:41
@nikol well, the general problem with biomedical research is that even with automation it is still hard to arrive at a good statistical sample :) that's the reason why Merks of the world are investing a lot of work into computer simulations of protein interactions etc.

The issue here is that actually figuring out what the structure looks like is the slow step. In an ideal world of drug discovery, you'd want to know what a specific protein looks like so you can find drugs that block it's function or something like that. E.g. reverse transcriptase inhibitors came about because people have solved the structure of the protein itself.

'Progress just means bad things happen faster.’
Previous Thread :: Next Thread 
Page 1 of 1