The Most Important Unsolved Issue of Modern Science

If someone asks you what AI is really good for, say AlphaFold. We just used it to solve what had been called “one of the most important yet unsolved issues of modern science.” The issue they were talking about is protein folding - predicting the 3D structure of a protein with just its amino acid building blocks.

This short is part of Huge If True, our optimistic show about science and tech. We made a longer episode all about AI in music! Subscribe to support us and to see more.

#shorts #tech #stem #ocean #deepocean #deepsea #animation #science

[helloworld]

Hi Cleo, I work on the protein folding field, and there is a big controversy about how alphafold was oversold. It is true that it can predict some structures much better than any other algorithm, yet it can't predict the structures of new folds or proteins with multiple folds. It is exceptionally good interpolating but really bad extrapolating, which is usually expected of AI. I know it is a small difference, but it is usually key for convincing a funding entity that there is still a problem needed to be solved.

[kurtnelle]

Yes, but now they need to validate each one by x-ray bombardment. This is because sometimes the AI hallucinates. You always have to check its work.

[kamenwaticlients]

This is some Star Trek level stuff being developed. Fascinating.

[joydipmajumdar]

Protein folding is a really complex process. The AI being talked about is good at predicting but it's not THAT efficient. No one in the academic field would actually trust it 100%

[heathrel]

Please, be realistic(

I am working with AF2 and it is not so optimistic as you state. It’s predictions aren’t true (you can check, how he spaghettifies a lot of proteins)

Yes, It’s huge, but we are far away from predicting protein structure by sequence

[RaeneYT]

There's something so endearing about hearing someone so passionate about something. Even better when it's about the things that truly change the world for the better.

[nuah]

Ah! That’s so cool you’re covering this topic, AI’s help in proteomics is so cool because it also means software alone can help boost identification of proteins which is far cheaper than upgrading hardware things like purchasing the latest mass spec model (or so I’ve been told). It has so many applications beyond human health, like in animals (eg my masters project on fish growth and how different proteins affect this or are reflected in this). Such a cool topic to bring to light, keep up the vids!

[AshnaBihani]

Your videos are so amazing, you explain science so clearly and effectively, I never get confused❤❤

[veona]

As a type 1 diabetic, I am super exited for this.

[YoungGandalf]

Folding proteins is easy. I'll be truly impressed when AI is able to fold my laundry for me.

[marcuscetina]

I listened to an NPR story about this very thing, and your short left me with an understanding and visualization that they were completely unable to! Love the content, keep it up!

[tinyprecious]

Thank you for sharing this and also for emphasizing optimistic stories 🙏🏼

[ilovekurkure]

I wish people who were experts in the field talked about this stuff, even at the beginner level, because videos like this one really try to oversell one benefit while ignoring the asterisks included with the text, which can be only fully understood and explained if you go deep into the subject and not just by reading few newspaper articles

[TheRobson]

I don’t understand any of this, but it sounds brilliant

[Royahmar]

I've always used Phyre2 for 3d protein models. What AI would impress me with is if it can start comparing the 3d folded protein to each for binding site comparisons. Can come in handy in A LOT of research.

[Segen_Bell]

Now this is what we should be doing with ai, helping the world; not making fake nudes, porn, and stealing from artists.

[mackenzieharty]

You and your content make me less scared of the world ❤

[philliplee]

Your content is absolutely freaking amazing. Done and done.

[CLove]

AI was announced as predicting cancer better than doctors...then it turns out it used an unrelated correlated variable and was just coincidentally right (lucky).

Lawyers were excited when generative AI gave them relevant cases...until they learned the AI made the cases up.

I'll be excited about this when we confirm that these structures are accurate, and when the AI is finding new proteins and novel ideas and approaches.

[rizma]

This is so cool, thank you for sharing this optimistic science story!