The media would have you believe that researchers merely asked AI to solve the folding problem, and it magically did. This video is an excellent illustration that the problem needed to be really well understood by the researchers, and a method had to be found in which to frame the problem in such a way that it was suitable for machine learning to solve. This is not unlike traditional software development. For now, there's still a lot of human intelligence required in getting AI to do useful things.
I know this is a bit different from my usual videos! I've been very interested in how AI in science recently, and AlphaFold is a great case study. I was shocked when I heard that protein folding, one of the most notorious and important questions in computational biology, had suddenly been solved by AI. In this video I explain what the AlphaFold algorithm actually does. Next week I want to see if I can replicate it.
Thanks for this video and looking forward to the follow -up. I can understand how AI can make predictions about folding and structure based on the actual sequence. But I guess another important aspect is of course the function that a protein can perform or how active it is. Is this something that AI can also predict?
Any video is already a nice surprise! The connection to quantum computing makes it totally on-brand because I'm seeing so much excitement about how transformers and attention models will replace other AI... maybe quantum computing will fulfill the hype too!
@HuygensOptics the structure is pretty important to predicting how the protein will react apparently, but I’m not sure exactly how! I would love to understand this, so if anyone has some good recommendations about proteins that’d be helpful
The triangle inequality is the kind of mathematical construct that, when learning about, seems kind of useless or obvious but then shows up everywhere.
Gosh, I remember at the beginning of my undergrad talking to my friend about protein folding and how hard is it. Five years later, we were talking about how AlphaFold just solved the problem. It's crazy! Would love to see more diverse stuff from you!
I certainly would not say that its 'solved'. Under rigid conditions of determining singular structures based on the protein data bank, alphafold performs amazingly. That's only a small subset of our proteome. alphafold cannot generate ensembles of structures, which is an exceptional challenge.
@@yousufo.ramahi126 AlphaProteome was founded for tackling that problem. It won't be easy. But every challenge we thought was insurmountable, turned out not to be.
In undergrad, I used to work in a structural biochemistry lab. I worked a lot with predicting protein structures using homology modeling to comparing two proteins with similar primary structures and other characteristics and using that to predict the tertiary structure of our protein of interest. I graduated before alphafold came out but I remember the excitement of those working in my former lab when it came out. Back then with our ab initio protein modeling techniques, I think the best we could do was proteins with like 100-200 residues with limited accuracy but I think alphafold can due over a thousand with decent accuracy. I no longer do any of that stuff but it is pretty amazing to see such a massive breakthrough in my former field of study.
You did a very good job explaining this. I watched Nazim Bouatta's whole presentation after seeing your video, and then came back to re-watch this. The way you fill in some of my gaps in understanding that I missed to pick up in his presentation, and the way he goes into detail on the things you summarized is great! Really helped me understand this incredible model even as a beginner in ml. I applaud you for your work.
And Deep Mind isn't even a biotech-focused research lab. They are just a bunch of computer scientists doing cool things with neural nets: playing video games, playing Go, protein folding, controlling plasma in a tokamak, etc...
I remember playing Fold-it years ago to help scientists figure out proteine folding as computers couldn't do it. Glad to see that computers *can* do it now 😊
It's interesting to see some AI topics on your channel. I find quantum physics interesting, but a lot of it goes over my head admittedly. AI/ML is more in my wheelhouse, so I'm excited for the follow up video on this. I think a lot of the more interesting AI algorithms are those like AlphaFold which combine some expert/domain knowledge to provide structure to more numerical machine learning algorithms. If you're starting to explore more AI literature, do you think you would try to create a video combining your own domain knowledge of physics plus AI to solve something new?
I have been teaching/researching mainly in Text analytics for the last 4 years, and man, I'm just mindblown on how they have leveraged attention mechanisms to solve this problem. Thank you so much for the video
Does it just predict how known proteins fold, or can it predict how a completely novel protein would fold? Like if I just invented an amino acid sequence that doesn't actually exist in nature, could it predict how that protein would fold if it were actually made?
That’s a great question! It claims to work for “biological relevant” proteins. There are definitely pathological ones it can’t do well. But I wonder how well it does for “random” proteins
It doesn't solve protein folding. It solves protein structure prediction to a very good accuracy based on amino acid sequence, it potentially covers the conformational landscape a protein can have. It not solving protein folding, it doesn't fold the protein nor give any insights (as of yet) on the stages. In addition, the structures predicted of heavily biased based on the current state of the PDB, dominated by X-ray crystallography structures. Blurry interaction (or fuzzy), ensemble interactions are coming more prevalent, something that is only now being touched upon. It is incredibly useful for pairing with experimental structure determination, and generating hypothesis, for example for molecular replacement (X-ray), model building (Cryo-EM) and informing biochemical perturbation of a protein-protein interaction.
I don't know much biology, but isn't it the case that not all amino acid sequences are capable of folding into a stable shape? I wonder what AlphaFold would say about those. Would it say, "I mean, like, we can fold parts of it, but we can't get the whole thing to fold into a stable shape"?
Great question! It seems like, given a particular environment (eg, ph levels and other molecules around) the protein will fold a certain way. But changing the environment can “denature” a protein (it will fold differently)
This is statistic analyze, and preparing data to do statistic analyze is main in quality of final result of statistic approximation. Just sort your data and build statistic of data connections - and predict outcome of new data base on statistic. Chess was solved same way in the late 1980s. Now "statistical approximation" called as AI. As mentioned in other comment - "media present AI as magic that invent rules of universe" - when it hand made data analyze made exact for this task.
Yes, but the chess algorithms in the late 1980s didn't win against the strongest human players, yet. Today computer chess is completely dead because it wins almost all the time.
MAM How u done quantum mechanics as i was browsing through ur old videos i saw u done quantum mech degree so where and how u done it mam can a 12th grader do it?
A smart 12th grader can do pretty much everything a physics undergrad can do, probably even a little more if they are exceptional. The only problem you will have is that without help the material at the university level is kind of hard to digest. Since you can't even write proper sentences you are not a smart 12th grader, though. :-)
@@lepidoptera9337 hmm and u said "The only problem you will have is that without help the material at the university level is kind of hard to digest" there is a mistake ig its should be like the only problem you will have is what without "the" help "of" .... :-)
@@sanjay-cm7dl I never claimed to be either a 12th grader or a native speaker. Let me give you some more attention, though, since your basement is very cold, kid. :-)
There is a philosophical dilemma here - would you give credit for "solving" a protein-folding to alphafold itself or the architects who baked-in the various sequences, loops, inequality-checks, pairwise modifiers, etc. into the basic design of alphafold? A "vanilla" model would have been inadequate to solve the problem, right? Alphafold "just" does the number-crunching.
Given that the algorithm seem to require input from genetically related species, what kind of data does AlphaFold need to predict protein folding? does it need information on the full 3D shape of a protein of related species?
Nah. More like "Homology Modeling" merged with Attension. If you wander, what Homology Modeling is - Its roughly what the first half of the video talk about. Alphafold just uses attention to extract better inference from the homology data.
AI is great for solving specific types of problems. Too bad it's being sold as technomagic for everything, and an incredible amount of resources have been dumped into the technomagic bubble that's starting to pop.
It's mainly machine learning. May be we should stop calling it AI. May be we should stop calling it machine learning either, since that's mostly statistical inference recursively applied gazzilions of times.
To be fair, ordinary digital computing was sold in the same way, with the same fears / mania expressed by the media, and the same business opportunities, some legit, many exploitative.
@@WanJae42 kind of. It was sold as instant access to infinite knowledge, which is kinda true, along with some technomagical problem solving. It wasn't being sold as this is the solution to all problems and will make you insanely rich.
@adampope5107 You may be younger than me. It was going to take over the world. Robots were going to run everything and take our jobs. Computers in movies could do everything except perform the actual car chase ... well until KITT came along. People thought that ordinary digital computers were already a kind of sentience, and didn't really know the difference. Not unlike how too many people today think that the current take on AI is sentient / thinking. The media played along. Hollywood played along. Tron depicted a massive intelligence going on inside an early 80s arcade cabinet. I give Terminator credit for at least explaining that the AI depicted required a leap in CPU technology unlike what existed at the time. (For the record, I love Tron.)
Once again brilliant explination. It's a shame your too educated to be a high school teacher. Because with the way you're able to break complex things down and make them interesting, you could get a lot of bored high schoolers interested in physics.
You can't get high schoolers interested in physics. I had a much better physics teacher than she could ever be. He didn't get anybody interested in it who wasn't already. Physics is like medicine, the law or the arts. Either you got it or you don't. There is nothing your teacher can do about that. They can help you if you got the spark, but that's about it.
The media would have you believe that researchers merely asked AI to solve the folding problem, and it magically did. This video is an excellent illustration that the problem needed to be really well understood by the researchers, and a method had to be found in which to frame the problem in such a way that it was suitable for machine learning to solve. This is not unlike traditional software development. For now, there's still a lot of human intelligence required in getting AI to do useful things.
I was very surprised at how involved this algorithm is! As you say, there was a lot of knowledge about the problem that was needed to solve it
For now
@@MelindaGreen Ain't that the case. For NOW. Right after she made this video, AlphaFold 3 came out. Four months later, AlphaProteome.
I know this is a bit different from my usual videos! I've been very interested in how AI in science recently, and AlphaFold is a great case study.
I was shocked when I heard that protein folding, one of the most notorious and important questions in computational biology, had suddenly been solved by AI. In this video I explain what the AlphaFold algorithm actually does. Next week I want to see if I can replicate it.
Thanks for this video and looking forward to the follow -up. I can understand how AI can make predictions about folding and structure based on the actual sequence. But I guess another important aspect is of course the function that a protein can perform or how active it is. Is this something that AI can also predict?
I would also be interested in the followup of replicating it. even a simplified version would be great
Any video is already a nice surprise! The connection to quantum computing makes it totally on-brand because I'm seeing so much excitement about how transformers and attention models will replace other AI... maybe quantum computing will fulfill the hype too!
Please more AI videos, AI is a pivot point for Humanity!
@HuygensOptics the structure is pretty important to predicting how the protein will react apparently, but I’m not sure exactly how! I would love to understand this, so if anyone has some good recommendations about proteins that’d be helpful
The triangle inequality is the kind of mathematical construct that, when learning about, seems kind of useless or obvious but then shows up everywhere.
Right? I remember thinking, “why are we bothering to learn this?” When it first came up.
Gosh, I remember at the beginning of my undergrad talking to my friend about protein folding and how hard is it. Five years later, we were talking about how AlphaFold just solved the problem. It's crazy!
Would love to see more diverse stuff from you!
Yeah, I remember how certain people were that it was an intractable problem classically. AlphaFold really came out of the blue!
I certainly would not say that its 'solved'. Under rigid conditions of determining singular structures based on the protein data bank, alphafold performs amazingly. That's only a small subset of our proteome. alphafold cannot generate ensembles of structures, which is an exceptional challenge.
@@yousufo.ramahi126 AlphaProteome was founded for tackling that problem. It won't be easy. But every challenge we thought was insurmountable, turned out not to be.
In undergrad, I used to work in a structural biochemistry lab. I worked a lot with predicting protein structures using homology modeling to comparing two proteins with similar primary structures and other characteristics and using that to predict the tertiary structure of our protein of interest. I graduated before alphafold came out but I remember the excitement of those working in my former lab when it came out. Back then with our ab initio protein modeling techniques, I think the best we could do was proteins with like 100-200 residues with limited accuracy but I think alphafold can due over a thousand with decent accuracy. I no longer do any of that stuff but it is pretty amazing to see such a massive breakthrough in my former field of study.
Oh that's so interesting! If you still keep in contact with your former lab, do you know how their research has changed now?
A friend of mine is working in that field as well. Maybe I can set you guys up for a talk? Would love to know the answer as well haha
You did a very good job explaining this. I watched Nazim Bouatta's whole presentation after seeing your video, and then came back to re-watch this. The way you fill in some of my gaps in understanding that I missed to pick up in his presentation, and the way he goes into detail on the things you summarized is great! Really helped me understand this incredible model even as a beginner in ml. I applaud you for your work.
As someone who is trying to get my head around all of this, this was incredibly well articulated and made it easy for me to understand - thank you!
thanks! there have been a few explanations of alphafold in various videos, but now I can say that finally there is a good explanation video on youtube
Aww, thank you so much!
Insanely well presented content.
And Deep Mind isn't even a biotech-focused research lab. They are just a bunch of computer scientists doing cool things with neural nets: playing video games, playing Go, protein folding, controlling plasma in a tokamak, etc...
Come for the animation. Stay for the content. ❤
I remember playing Fold-it years ago to help scientists figure out proteine folding as computers couldn't do it. Glad to see that computers *can* do it now 😊
I remember that!! Crazy isn’t it that computers are now good at this problem?
Top tier content right here
It's interesting to see some AI topics on your channel. I find quantum physics interesting, but a lot of it goes over my head admittedly. AI/ML is more in my wheelhouse, so I'm excited for the follow up video on this. I think a lot of the more interesting AI algorithms are those like AlphaFold which combine some expert/domain knowledge to provide structure to more numerical machine learning algorithms. If you're starting to explore more AI literature, do you think you would try to create a video combining your own domain knowledge of physics plus AI to solve something new?
Thank you! I’m very interested in the intersection of quantum physics/ chemistry and AI. I’d love to make more videos about it!
The triangle inequality went a long way towards understanding for me, thanks
And Go. We thought Go was impossible for a computer. We learned that AI can see patterns that we overlook
Yes, very true.
Very clear explanation. Thanks!
I have been teaching/researching mainly in Text analytics for the last 4 years, and man, I'm just mindblown on how they have leveraged attention mechanisms to solve this problem. Thank you so much for the video
We missed you!!
Sorry!!
thank you indepth analysis for the alpha phold
Does it just predict how known proteins fold, or can it predict how a completely novel protein would fold? Like if I just invented an amino acid sequence that doesn't actually exist in nature, could it predict how that protein would fold if it were actually made?
That’s a great question! It claims to work for “biological relevant” proteins. There are definitely pathological ones it can’t do well. But I wonder how well it does for “random” proteins
Hi Mithuna, I really enjoy your videos. But could you take some time out to complete the linear algebra series, they do not have to be perfect?
Very well explained! Thank you very much for this. I have an upcoming presentation about AF and this helps a lot to understand the main idea.
Look at AlphaGo go, how AlphaFold folds, and Alphabet... Bets?
It doesn't solve protein folding. It solves protein structure prediction to a very good accuracy based on amino acid sequence, it potentially covers the conformational landscape a protein can have. It not solving protein folding, it doesn't fold the protein nor give any insights (as of yet) on the stages. In addition, the structures predicted of heavily biased based on the current state of the PDB, dominated by X-ray crystallography structures. Blurry interaction (or fuzzy), ensemble interactions are coming more prevalent, something that is only now being touched upon. It is incredibly useful for pairing with experimental structure determination, and generating hypothesis, for example for molecular replacement (X-ray), model building (Cryo-EM) and informing biochemical perturbation of a protein-protein interaction.
Attention is truly all you need
I don't know much biology, but isn't it the case that not all amino acid sequences are capable of folding into a stable shape? I wonder what AlphaFold would say about those. Would it say, "I mean, like, we can fold parts of it, but we can't get the whole thing to fold into a stable shape"?
That’s so interesting! I might give it a go. My guess is that it just takes a stab at it and gets it wrong
Given a one long string of amino acids, is there only one unique way that can fold?
Great question! It seems like, given a particular environment (eg, ph levels and other molecules around) the protein will fold a certain way. But changing the environment can “denature” a protein (it will fold differently)
Awesome animations. Really nicely done :)
This is statistic analyze, and preparing data to do statistic analyze is main in quality of final result of statistic approximation.
Just sort your data and build statistic of data connections - and predict outcome of new data base on statistic.
Chess was solved same way in the late 1980s.
Now "statistical approximation" called as AI.
As mentioned in other comment - "media present AI as magic that invent rules of universe" - when it hand made data analyze made exact for this task.
Yes, but the chess algorithms in the late 1980s didn't win against the strongest human players, yet. Today computer chess is completely dead because it wins almost all the time.
Thanks! It was fun to watch.😀
Thanks, that was very good!
MAM How u done quantum mechanics as i was browsing through ur old videos i saw u done quantum mech degree so where and how u done it mam can a 12th grader do it?
A smart 12th grader can do pretty much everything a physics undergrad can do, probably even a little more if they are exceptional. The only problem you will have is that without help the material at the university level is kind of hard to digest. Since you can't even write proper sentences you are not a smart 12th grader, though. :-)
@@lepidoptera9337 hmm and u said "The only problem you will have is that without help the material at the university level is kind of hard to digest" there is a mistake ig its should be like the only problem you will have is what without "the" help "of" .... :-)
@@sanjay-cm7dl I never claimed to be either a 12th grader or a native speaker. Let me give you some more attention, though, since your basement is very cold, kid. :-)
There is a philosophical dilemma here - would you give credit for "solving" a protein-folding to alphafold itself or the architects who baked-in the various sequences, loops, inequality-checks, pairwise modifiers, etc. into the basic design of alphafold? A "vanilla" model would have been inadequate to solve the problem, right? Alphafold "just" does the number-crunching.
Given that the algorithm seem to require input from genetically related species,
what kind of data does AlphaFold need to predict protein folding?
does it need information on the full 3D shape of a protein of related species?
No, just other amino acid strings!
ChatGPT on steroids except it speaks in angles and distance instead of English 🥴
Nah. More like "Homology Modeling" merged with Attension. If you wander, what Homology Modeling is - Its roughly what the first half of the video talk about. Alphafold just uses attention to extract better inference from the homology data.
very interesting
Holy moly!!
AI is great for solving specific types of problems. Too bad it's being sold as technomagic for everything, and an incredible amount of resources have been dumped into the technomagic bubble that's starting to pop.
It's mainly machine learning. May be we should stop calling it AI. May be we should stop calling it machine learning either, since that's mostly statistical inference recursively applied gazzilions of times.
To be fair, ordinary digital computing was sold in the same way, with the same fears / mania expressed by the media, and the same business opportunities, some legit, many exploitative.
@@WanJae42 kind of. It was sold as instant access to infinite knowledge, which is kinda true, along with some technomagical problem solving. It wasn't being sold as this is the solution to all problems and will make you insanely rich.
@@aniksamiurrahman6365 AI, and deep or machine learning are definitely misnomers.
@adampope5107 You may be younger than me. It was going to take over the world. Robots were going to run everything and take our jobs. Computers in movies could do everything except perform the actual car chase ... well until KITT came along. People thought that ordinary digital computers were already a kind of sentience, and didn't really know the difference. Not unlike how too many people today think that the current take on AI is sentient / thinking. The media played along. Hollywood played along. Tron depicted a massive intelligence going on inside an early 80s arcade cabinet. I give Terminator credit for at least explaining that the AI depicted required a leap in CPU technology unlike what existed at the time. (For the record, I love Tron.)
Didn't understand a word.. other than tube and London
Neat!
So, basically, homology modeling on steriods, ummm, I mean attension.
Yay you upload! Never knew you were into AI. Good for me i guess...
Edit: Obligatory "first" message
God's geometry of creation is waiting for you, your Esoteric studies will be realized in Time. ❤
Please take this nonsense elsewhere. Smug theists always show their whole ass when they can't understand something. Go read a book. Try and learn.
Once again brilliant explination. It's a shame your too educated to be a high school teacher. Because with the way you're able to break complex things down and make them interesting, you could get a lot of bored high schoolers interested in physics.
You can't get high schoolers interested in physics. I had a much better physics teacher than she could ever be. He didn't get anybody interested in it who wasn't already. Physics is like medicine, the law or the arts. Either you got it or you don't. There is nothing your teacher can do about that. They can help you if you got the spark, but that's about it.