AlphaFold 2 is amazing but protein folding is still not solved | Dmitry Korkin and Lex Fridman

Lex Fridman Podcast full episode: • Dmitry Korkin: Evoluti...
Please support this podcast by checking out our sponsors:
- Brave: brave.com/lex
- NetSuite: netsuite.com/lex to get free product tour
- Magic Spoon: magicspoon.com/lex and use code LEX to get $5 off
- Eight Sleep: www.eightsleep.com/lex and use code LEX to get special savings
GUEST BIO:
Dmitry Korkin is a professor of bioinformatics and computational biology at WPI.
PODCAST INFO:
Podcast website: lexfridman.com/podcast
Apple Podcasts: apple.co/2lwqZIr
Spotify: spoti.fi/2nEwCF8
RSS: lexfridman.com/feed/podcast/
Full episodes playlist: • Lex Fridman Podcast
Clips playlist: • Lex Fridman Podcast Clips
CONNECT:
- Subscribe to this KZbin channel
- Twitter: / lexfridman
- LinkedIn: / lexfridman
- Facebook: / lexfridmanpage
- Instagram: / lexfridman
- Medium: / lexfridman
- Support on Patreon: / lexfridman

[@justindavid6181]

Wow! That's compLEX!

[@sallerc]

And now in 2024 we have AlphaFold 3, would be interesting to hear this guy talk about the new version-

[@SongbirdXX]

It’s all fun and games until you fold with the wrong person.

[@mcasanovaiii]

...I like the movie t sum of a fears ... re watching rn actually...

[@davidregi7571]

Lex I like your protein

[@almor2445]

2 years ago in AI is like decades in any other field. Surely this has been improved since GPT-4 (and if not, then soon) with Google Gemini?

[@germaned.lealnino631]

Here is a serious question for all of you reading the comments. Can crisp therapeutics make money out of the findings from alpha fold 2?

[@thryce82]

i think the business plan is to get pharma togive them sequences that they are interested in they build the protein model and now that pharma company has a huge leg up on its competitors.

[@johnbouttell5827]

Is that a real hedgehog -- or a Sears hedgehog?

[@00coppelia00]

Classic scientist problem, explaining it in complex terms but not explaining why we are in this quest of solving the problem for the lay audience.

[@arthurdiaz9184]

Are you thinking? Think about it & you can answer your own question.

[@00coppelia00]

I have no idea what you mean, since I didn't ask any questions in my message. It was an observation.

[@Hshjshshjsj72727]

Sounds like gsp

[@eatmyspam]

What does he mean by domains around 14 minutes? Can anyone elucidate?

[@dawnkeyy]

from wiki A protein domain is a region of the protein's polypeptide chain that is self-stabilizing and that folds independently from the rest. Each domain forms a compact folded three-dimensional structure. Many proteins consist of several domains. One domain may appear in a variety of different proteins.

[@davidmadariaga9675]

They are just regions of proteins

[@ilikenicethings]

I missed the reason why physics based algorithms don’t work

[@sandy19842]

Performance. Underlying problem-space is highly complex and we haven't come with ways to optimize it down to fit processing capabilities, if indeed there are enough optimizations waiting to be found out there to make a purely physics based approach workable to begin with which isn't guaranteed either.

[@MrSidney9]

A chain of amino acids has 10^50 to 10^150 possible ways of stably folding into a protein (the longer the chains of amino acids, the more possibilities). So you can’t predict solely based on electrochemical stability, which is what physics would be doing.

[@thryce82]

complexity i believe. to do a standard physics based simulation of a protein using an iterative model just takes wayyyyyyyy to long . you can do it for small peptides but it literally takes a supercomputer and fwhen you start adding up the fact most proteins are made of multiple peptides it becomes untractable.

[@earthbndmsfit]

How that shit working for yeas

[@jonabirdd]

Ok, I'm not sold, I think most people care about structure prediction, not protein folding, the latter seems more a means to the former. I'm wondering if this guy is confusing the trees for the forest...

[@aleph0540]

Correct.

[@raghuvenkatesan6792]

Structure prediction is protein folding

[@ashanmaynard4085]

@@raghuvenkatesan6792 Not really.Solving the Protein folding problem is knowing the folding pathway+final structure, not just the final structure. Alphafold doesn't have any idea about the folding pathway which is most important.

[@raghuvenkatesan6792]

@@ashanmaynard4085 lmao you need to know the folding pathway to know the final structure

[@ashanmaynard4085]

@@raghuvenkatesan6792 No. These are huge neural networks which we don't know how hidden layers work. Seems like Y doesn't know how neural networks work.

[@mcasanovaiii]

Protein shakes have been baroundforever...

[@robertmihalinac]

What if AlphaFold 3 could predict with 100% certainty the structures of every protein in human body, would then the protein folding be solved?

[@anuraganand2736]

I dont think It can ever be 100% ,it is a NP-Complete problem so yes they can increase the accuracy more n more n more but with a bottoms up approach I don't think its possible to ever reach that limit coz the model will always be contraint by the limitations of the dataset it sees.

[@robertmihalinac]

@@anuraganand2736 Thank you. If we produce syntetic protein, does it have to be folded with 100% accuracy (like natural protein) to have a medical value? What if we get into body protein that is folded with 95% accuracy, will it heal or will it stagger other proteins in the body and produce more medical problems?

[@anuraganand2736]

@@robertmihalinac @Robert Mihalinac that I don't think I know enough to answer it properly but as also mentioned in the video , its much complex than that and also with the fact that any bottom up approach will be constrained by how much observations /data it encounters. yes actually many diseases such as type 2 diabetes occur from wrong unfolded structure but I think and I could be wrong and do tell me if you think I am,it should help us in synthesizing more effective medicines/ vaccines as more and more accuracy will cutoff the time as well as cost that it usually takes with all the stages with a vaccine manufacturing where we can first simulate and then only start with the other stages and as with any np complete problem somewhere in the process they will encounter some non deterministic step where some scientist will have to take the decision to go through one particular path .

[@robertmihalinac]

@@anuraganand2736 Thank you

[@thryce82]

​@@robertmihalinac yes and no . proteins "breathe" they have jiggls and dynamic movements. most of them do not matter. what will matter is how the protein binds to say the pocket of a receptor. this is not an all or nothing thing but is generally very sensitive. like if it doesnt bind within a few angstroms you wont get conformation change. Now it doesnt have to hold this state with 100% accuracy either the kinetics of receptor interactions are imp too. drugs generally target enzymes that do a function you want to tune not stop or start completely . as for causing damage unlikely. your body has plenty of unfolded and misfolded proteins lying around. some, prions , do cause damage but it is extremely rare and those conditions usually strike when you are elderly. If the protein doesnt work your not gonna take it for 40 years so thats not a huge worry

[@Kevoshea]

Lol

[@jorgediaz1605]

2nd

[@jorgediaz1605]

Third *

[@iankelly6632]

#Evolution did it 😂