Has AI 'solved' protein folding? | Hacker News Reader

Has AI 'solved' protein folding?

explainthispaper.com

29 pointsstuartbman23 comments

Funny D-Wave claimed to have solved this problem years ago http://blogs.nature.com/news/2012/08/d-wave-quantum-computer...

Yeah but not really "It worked, but not particularly well. According to the researchers, 10,000 measurements using an 81-qubit version of the experiment gave the correct answer just 13 times. This was owing, in part, to the limitations of the machine itself, and in part to thermal noise that disrupted the computation. It’s also worth pointing that conventional computers could already solve these particular protein folding problems."

To be fair, "solves protein folding problem" is a lot narrower than "solves protein folding".

E.g. "Multivac solves physics problem" vs. "Multivac solves physics"

This paper provides an approach for a simplified representation of a problem, so called lattice protein folding, for a very small protein.

In particular, they find a 2D lattice fold of a 6-amino acid sequence (PSVKMA). 2D lattice fold means that every amino acid must occupy discrete Euclidean coordinates (i.e. (0,2) or (2,1)), which allows you to convert the search into optimization problem.

Couple of years later (using more tricks and newer D-Wave machine), we pushed the limit to 10 amino acids on the planar grid, and generalized the approach to 3D grids as well (where we were able to obtain a lattice fold of 8 amino acid sequence) [1]. This is actually still well below what your laptop can search with some effort.

On the other hand, AlphaFold2 predicts coordinates of the C-alpha atoms in the continuous space, so the predicted structure would actually be an potential direct substitute for experimental structure (similar to, i.e. a structure obtained using homology modelling).

[1] https://arxiv.org/abs/1811.00713

I guess the bigger question is: what can we actually do with this? They keep talking about new drugs for things like cancer, but even if you find a "match" with x-ray crystallography today, that doesn't get you anywhere near a successful drug - there are hundreds more variables.

From what I understand a quick and accurate algorithm for protein folding would allow researchers to more easily create proteins with the structure they want. Create a sequence, fold it, see if it has the desired structure, tweak sequence until done.

This is quite useful since proteins perform basically all work in life

My impression as a non-expert from reading the last thread is:

DeepMind didn't "solve" protein folding in the game theory sense that perfect play is now possible and there are no better solutions to be had. That's probably what most of us expect from "X has been solved."

But it "solved" it in the sense that it answered a core unanswered question of an entire field. Algorithmic protein folding was a research field trying to answer whether it was even possible to climb from 40 to 90 on the CASP competition, or whether physical experiments were inherently superior. Obviously entering that field implied that you thought it was possible, but it wasn't known. DeepMind has now answered that question: yes, computational folding can work as well as experiments. That's a solution, if you take the open problem to be "can this be done at all?" rather than "what is the perfect way to do it?"

Great result, in sore need of external validation (just like every other piece of research). In a few years we will see through the hype.

The contest is external validation.

For an in depth review of what this work was actually about (and how it differed from AlphaFold 1, hint it was probably transformers), see this great video from Yannic Kilcher [0]

[0]: https://www.youtube.com/watch?v=B9PL__gVxLI

Maybe I'm misunderstanding the purpose of the site, but this is silly. I was mad at the TLDR not including the answer to the premise, until I realized that the post itself doesn't even make an attempt. The information contained is basically the same as scrolling through the HN comments.

True TLDR; Yes as far as we can tell. But we'll have to see.

> Any headline that ends in a question mark can be answered by the word no.

Should it be updated?

"Any headline that ends in a question mark should be moved the end of the article."

The article also fails to mention that the structure of many proteins doesn't correspond to their "natural folding" but is modified by other proteins (e.g. chaperones), or is determined by the complex to which the protein belongs, i.e. cannot be predicted with data based on isolated proteins. Furthermore most proteins were measured in crystallized (i.e. not their natural) form, so the resulting static structure is likely not representative (even less if parts of the protein are "moving" to perform their function). So it would be quite naive to assume that from now on you can just throw a sequence into the black box and the right structure comes out.

Thanks for mentioning this, it's always difficult determining which levels of information to include, and which to leave out. I agree this is worth mentioning and I've now added this in to the article.

I think one should not withhold from the Python generation the fact that molecular biology is much more complex than one might realize in just looking at a few data. It's fine that AI can make scientifically relevant contributions, but to call it the greatest of all discoveries and Nobel Prize candidate, as done in many posts, is rather pretentious.

I wish these articles that don't really try to contribute wouldn't be upvoted :(

Maybe the flagging system is too slow, or we need downvoting, there should be a better way to make sure that HN front page articles are informational.

I genuinely thought this was helpful! I wrote the article to make the methods more transparent and understandable, sorry if it's not worked for you.

I genuinely thought it was helpful, too! I've only dabbled in AI, though, and my bioengineering coursework stopped well before protein folding, so the simple terminology and basic explanation to an outsider was probably more appropriate.

True, you didn't explain how the new algorithm achieves its shockingly impressive result, but that's probably not something that anyone could provide in a 500-word article that's easy to approach for non-experts. Though these guys were able to teach a computer to fold proteins more accurately than anyone expected it could be done, maybe they could extend their model to have it also write the paper for them...

The article is more about how AlphaFold works, not about how useful it is in practice, which the article should focus on with this title. You should have selected a title that's more in line with the article.

Or what's more interesting is to have an article about the current top use cases and companies that need protein folding, and how much money they can save on finding better drugs...now that would be really interesting to me... that would be the appropriate content for the title that you gave.

Thanks for the feedback. As a researcher I'm more interested in the methods of how they've achieved this performance and this was what I wanted to share. The top use cases are briefly listed here but frankly unlimited within and beyond pharma- as another commenter has said, if this is applied a Nobel is likely. Until then we'll need to wait for the full paper.

Sure, but for the method I really liked the video from Yannic Kilcher 3 days ago, so I compared this articlecto that (I'm subscribed to his channel, I love his detailed explanations):

https://youtu.be/B9PL__gVxLI

I found it very useful, as it went quite into detail and explained why this step was significant.

I tend to distrust "X solves Y" headlines from regular news lack nuance and understanding so often that it's better to think of them as wrong.

Reply on news.ycombinator.com