DeepMind and Protein Folding

[Nert]

Following is a link to an article describing Google's application of their DeepMind platform to protein folding. DeepMind was entered into a competition where participants predict protein structures from a list of their amino acids. Deep Mind came in first.

https://www.theguardian.com/science/201 ... f-proteins

[bruce]

But it cannot predict the unlikely prospect showing how a protein can miss-fold?

[Nert]

But it cannot predict the unlikely prospect showing how a protein can miss-fold?

I think that they are looking for the "normal" results, but that's mostly a guess. I tried to dig through the CASP13 (Critical Assessment of Techniques for Protein Structure Prediction) information linked by the Guardian article, but I couldn't make much headway. The information there is well beyond my superficial understanding of these things. Maybe one of the science guys can provide some high level context.

[bruce]

Yes, the CASP is aimed at predicting the normal structure. It's not surprising that AI would be good at that.

FAH goals are different. FAH's goals include a thorough understanding of how and why proteins fold in a specific way and finding the exceptions (known as mis-folding). FAH is also capable of predicting normal structures, but that has never been a primary goal and it's not particularly efficient at doing that.

Many of FAH's recent studies have been broadened to include "normal" interactions, but often exceptions are also discovered which turn out to be very valuable, scientifically.

[Ricorocks]

In the early days of computer security, security industry was fragmented, with many small companies. Soon to be gobbled up by larger firms!! Better/worse for that industry can be argued, each having pluses & minuses.

Lot of folding going on, with different approaches to the same common goal. Gobbling up by large companies, would be problematic, does FAH, converse/share/have overlap/could be helpful, with say Rosetta. Or do the two working on the same thing, act proprietary, don't look at my work?

FAH seems to have not participated in CASP, did FAH gleam anything useful from CASP, or just ignored it?

[bruce]

FAH and CASP have different goals --- both scientifically useful. They've participated but not aggressively. Finding all possible paths to the most probable end-point is not the most effective way to find that end-point. Or, from tho opposite perspective, finding the most probable end-point is not the most effective way to discover alternate paths which happen to lead to a mis-folded end-point.

Lot of folding going on, with different approaches to the same common goal. Gobbling up by large companies, would be problematic, does FAH, converse/share/have overlap/could be helpful, with say Rosetta. Or do the two working on the same thing, act proprietary, don't look at my work?

I'm not that familiar with Rosetta so I can't give you a good answer. I have heard Vijay mention University of Washington in a positive way so there's probably some cooperation going on at some level. Like all scientific research, there's always a bit of completion, too -- like we'd like to be able to be able to claim that we made a big discovery first.

(Like the recent PBS special on the discovery of the double-helix in DNA.)

[Ricorocks]

Thanks Bruce!

[rbpeake]

Update on AlphaFold: It is getting much better at what it was designed to do.
https://www.nytimes.com/2020/11/30/tech ... lding.html

[MeeLee]

Apparently the AI in combination with folding, gets better results.
All the AI is doing, is calculating what chance is higher to get 'a hit', and prioritizes those WUs earlier on in the cycle.
That way more work can be done folding the actually interesting data, rather than folding blindfolded (+99% miss cycles)

[The_Bad_Penguin]

Sorry if a repost.

Google’s Deepmind claims to have created an artificially intelligent program called “AlphaFold” that is able to solve those problems in a matter of days.

https://www.independent.co.uk/life-styl ... reddit.com

[will.allison]

The problem of these algorithms is are restricted until now cannot demonstrate that a problem P=NP (or the opposite), that means that, we did not find a generic equivalence in which a non-polinomic problem (exponential time consuming, for example, n^k), is solved in a polinomic time... meaning that those algorithms are far for being generic problem solvers

[bruce]

It's a matter of slightly different goals.

If you want to find the most probable folded state very quickly, there are better methods that FAH.
If you sant to study all the ways that a protein might mis-fold and cause a disease, then AI isn't going to get there faster.

CASP* (Critical Assessment of Techniques for Protein Structure Prediction) has reasonable goals and they do generate good science, but they're not the same as FAH's goals. FAH has contributed to CASP in previous years, but an exhaustive study of all of the shapes that a protein can take on on the way to THE single final solution is not the fastest way to get to that unique shape.

[JimF]

Speaking of Vijay, here is his take on AlphaFold:
https://a16z.com/2020/12/06/16mins-deep ... ng-ai-bio/

[gunnarre]

Dr. Bowman's piece on it: Protein Folding and Related Problems Remain Unsolved Despite Alphafold’s Advance

TL;DR: DeepMind’s AlphaFold marks a tremendous advance in our ability to predict the dominant structure of a protein. How proteins get there, and many other problems related to protein folding, remain unsolved. Synergies between AlphaFold and Folding@home could enable more progress.

DeepMind recently claimed they solved the protein folding problem with their AlphaFold algorithm. They made a huge advance, but solving protein folding is a little like saving the world. Many have tried, some have made big steps forward. Yet protein folding won’t stay solved any more than the world will stay saved.

(...)

[Nert]

Speaking of Vijay, here is his take on AlphaFold:
https://a16z.com/2020/12/06/16mins-deep ... ng-ai-bio/

Thanks for this link. As a layman I found it mostly understandable. I think that I now understand the difference between Alpha Fold and FAH: Alpha Fold predicts what a protein will look like after it's folded. FAH determines the individual steps a protein goes through to get to that final shape. I believe that discovering weak spots that a disease protein shows as it folds (but are not exposed once folded) is the really interesting thing that FAH can contribute that Alpha Fold cannot.