Alberto Perez: AI and Structural Biology From Protein Structure to Molecular Design (TSVP Talk)

Hello. Hello everybody. Thank you very much for coming um to today TSVP talk. So today we have Alberto Perez uh whom I'll introduce briefly. Um so Alberto did his undergrad and PhD in Barcelona. Then he moved to the US to work on the protein folding problem with Ken Dil. And then since 2008 um he leads his own research group at the University of Florida currently a associate professor and associate director of the quantum theory project. Um and today he will be talking about the AI in the context of structural biology. So yeah thank you all for coming. >> Hi everyone and thank you for being here. Uh when I asked who will come to the talk I wasn't sure if anyone would show up. So, it's really energizing to see so many people here. Um, I'm here for three months. I've stayed already six weeks. So, I have six more six more weeks to go. So, if you like any of these, please come talk to me afterwards. Um, what I thought I would do today is give you an overview of my field, the kind of things we do. So, I'm a computational stctile biologist. I work in trying to understand biomolelecules, how they work, how they interact with each other. And in our field, as in many other fields, AI has made uh a huge revolution. So I'm trying I'm going to try to tell you how we're going from protein structure prediction to molecular design. And one of the things that I came here to Okinawa to do is to try to understand better how enzymes work, how they are able to perform their work. How can we design novel enzymes that work better and do things that we desire them to do? And so the big idea about proteins and biomolelecules is that um there are many of them of many different types. This is an artistic representation of what might be going on inside one of our cells. So you have multiple different kind of proteins. They are diffusing around. They're interacting amongst themselves with other molecules membranes. There are some small molecules, there are some large molecules, molecular motors and whatnot. And so the big idea is that inside of our cell there's a lot of activity going on. We have a lot of uh these biomolelecules, proteins and whatnot that in order to satisfy their functional role they need to interact with each other. And so these molecules they are small. So if this is a cell proteins are about 5 to 10 nanometers. We cannot see them with our naked eye. They interact with all kind of molecules. Some of them like aspirin are much much smaller. And in order to for us to understand how these molecules work, we need to know what they look like. So knowing what they look like means understanding where every atom is placed. And to do that, we have some experimental techniques such as X-ray crystalallography. So we try to crystallize proteins, bombard them with X-rays, and then where they defract that must be the places that they collide with the nucle. So we can backmap what the stack of a protein looks like. We can also use NMR. So trying to use nuclear magnetic resonance in order to understand what atoms might be close to each other and traced out what the effect of the protein is. And finally we can use carry. So since the resolution revolution there has been like a massive amount of factors. There's a lot of expertise here at OIS about carry and so those are the three main techniques by which we can know how proteins look like. The challenge is that these three techniques they are expensive they might not be amenable to every protein and so there there might be other ways in which we can predict or try to understand what the structures of these proteins look like. And so for someone that as a kid like to play with Legos, proteins are the ultimate Lego. So they are made of 20 different building blocks amino acids and the way that they're made is like a bits on a string. So you're putting different amino acids one after another. Each amino acids. So each different color represents a different kind of amino acid. You can divide them in four major classes. So blue are positively charged amino acids, reds are negatively charged amino acids, green are polar, and white are hydrophobic. They like to shy away from water. And so under the influence of the environment and the properties of these amino acids, some of these protein sequences can undergo a transformation. They can evolve. And so I'm going to represent this string of amino acids with this string. And I'm going to highlight here four residues. These are four hydrophobic residues. They like to share away from the water. I'm going to put in this string four different uh places in yellow. I'm going to explain that later and one of the pieces in blue. And so what I'm going to show you is a process that we call process of protein folding. So by looking at the interactions between the different amino acids and looking at the interactions with the environment, you have every now and then some stoastic kind of interactions that ultimately lead to the folded uh protein. So at some point those four hydrophobic residues that sh away from the water, they interact with each other. They clamp together at the core of the protein and they're protected by other residues that are polar and interacting on the outside. So this is a folded protein. You see that the four different yellow pieces they are interacting together. They form what we call a beta sheet. And this blue region here is making an alpha helix. So this is how this protein falls. You have many different kind of proteins. They come in all shapes and sizes and they do different things. So you have some proteins that are enzymes. They're able to break and make bonds. You have proteins that are able to recognize other molecules maybe in disease. You have proteins that for example allow your muscles to contract and extend. And you have membrane proteins that allow things to go in and out of the cell either molecules or signals coming out of the cell. So one of the things that we want is to understand what the structures are like so that we can understand what are the functions of these molecules. And so I've shown you the process of protein folding. There are two kind of competing interests. On one side you have the chain entropy. So every time I put like a something inside my pocket it comes out like a tangled mess. So that's the entropy. And on the other side you have like the interactions that these uh amino acids can make with each other that favor folding. And so you have an entropic term an enthalpic term and there's a balance between the two. For some protein sequences what you find is that at low temperature you can find this native state this folded state. And this is a panel like representation where at high temperature you find unfolded states at low temperature you find the folded state. Now how many protein confirmations how many confirmations can a single sequence adopt? Many many many many many in fact more number of possible uh configurations than the number of hydrogens in the known universe. That's how many possible confirmations every single sequence can do. So if a protein had to go through every single possibility before finding a native state, it would take longer than the edge of the universe for a protein to fold. Obviously, we cannot wait that long. And so all proteins fold very fast. The fastest proteins can fold in milliseconds and even in microsconds. And so that's how how fast they fall. And so for the longest time, this led to some questions in in our field. It's what's called the protein folding problem. It's a 60 year old kind of problem and it has three parts. The f the first one is how do proteins fold so fast? What are the structures that they adopt? And third, can we use computers to predict those structures? And so these last two questions have been mostly solved recently and I will tell you how. These third questions, how do proteins fall so fast? That's still something that we are discovering. So I've told you this is a 60 year old kind of question. So for the longest time if you look at papers historical papers you will see oh we solved the protein folding problem and then someone else publish another paper. Oh we solved the protein folding problem. It turns out like yeah maybe they did something with one or two proteins and then you tried to apply the methodology to a new protein and it didn't work. And so one of the biggest things, one of the biggest contributions to our field is the emergence of blind uh prediction uh competitions. So these are events in which someone is going to say well we're going to start a competition. Everyone in the world can participate. What we're going to do is we're going to contact some experimentalists that are working on crystallizing or solving with Kimm some of the proteins. We're going to ask them for the sequence and then we're going to provide this sequence to everyone who wants to participate. So this usually happens every other summer. So during summer we have two months. Every day we're given about two sequences. The competition is going on right now actually. And for every sequence you have about three weeks to predict up to five stractors. So you submit your pipe structures and at the end of the competition someone who has not participated will be assessing all of the predictions. This is what we call a double blind experiment blind to the predictors because we don't know what sequences we're going to be given. So methods uh cannot be um geared up for a particular sequence and then it's blind to the assessors because they're given all the predictions but they don't know who predicted what. And so at the end what you get is what is the state-of-the-art of the field? What are the best methods to work on the protein solving uh protein structure prediction problem? How do you think they did on the first chance? Good, bad, bad, right? Yeah. So the first experience was aartic experience. So if I were to go back in time to 1995 and I were to say well this protein has all the atoms on top of each other I would have one CASP that doesn't have any physical kind of sense but according to the metrics that they use that would have been the best protein that's how bad we were at predicting proteins. All right did we get better over the the years? Well, this is a plot that is going to try to explain that the answer is no. We didn't get much better. So each line here is a different competition. So every two years we have a new line. On the Y ais you have the ability to have accurate predictions. So good means the atoms are in the right place. Bad means it's random noise. And so on the x-axis I have whether the protein sequence that we were given is easy or hard. How do you tell if a protein is or a sequence is easy or hard? Well, I'll tell you an anecdote. So I came to Okinawa. I'm originally from Spain and I was walking around and I heard something that looks like sounded like a duck. I thought I went around. I looked for it. It looked like a duck but not like any duck that I knew about. And so I said, well, it sounds like a duck. It looks like a duck. It must be an Okinawan duck, right? So they do the same for protein sequences. If a protein sequence looks like another sequence for which I know the structure, I'm going to say this sequence most likely falls to that structure. So that's how we define an easy sequence. And for easy sequences, we get good structures. For proteins that don't look like other proteins, they're sequenced. we have very bad ability to predict their structures. So all right so over 20 years almost no progress. What did change is how much data we accumulated. So back in the 1990s we knew about 50,000 sequences. Then came all the genomic and metagenomic projects. We now know about over 200 million uh protein sequences. We used to know about 4,000 structures deposited in the protein data bank. we now know about 300,000s. So this changed from a problem of like oh can we use physics or can we use a statistical mechanics to predict the structures of proteins to an AI kind of challenge and so in my field computer programs and um science have gone hand in hand. So in the 1990s IBM developed deep blue. It beat the reigning champion Kasparov at the game of chess. So basically the idea is that by brute force a computer can calculate more moves than a human can. At any point in time you can look at the chess board and you can decide who's winning. And so with that uh deep blue beat Kasparov. So IBM said oh now we're going to solve protein folding. And they didn't. So a little bit later, a company named Deep Mind is a subsidiary subsidiary from Google. They decided to play the game of gold. Now, very very typical game in in Japan. And the idea is that here you're putting stones either white or black and you're trying to uh capture the pieces of the of the opposite uh side. The problem is that there are many more possibilities than in the game of chess. So now you cannot guess by brute force predict who's going to win. And there's no way to know who's winning by looking at the board. You really have to play out the game. And so if we were just waiting for a computer to be fast enough, we would still not be able to beat the reigning champions. And yet AlphaGo was able to defeat the world champions. And so and one of the particular interesting things so this was five matches I think it won 4-1 and in one of the matches the com commentators there are commentators for these games said oh the machine has gone crazy it has made a move that makes no sense 100 or something turns later it turned out that that move was critical to win the game and so it had hallucinated a move that no human would ever make and Yet it was critical to win the game and this is something that will come back later because when AlphaGo won they said well our next challenge are proteins and so I've shown you that there was no progress in 20 years alpho came into the so alpha go changed to alpha fold and in two years we went from here to all the way here. At this point they were just newcomers to the game. they were just doing everything that everyone else was doing. So they were just they had more resources, they had more people working on it. So in a few months they were able to beat the whole field. Two years later they said well we can now develop new transformers that pay attention to the sequence and structure similarities and now this is the state-of-the-art. And this this technology was actually the one that recently won the Nobel Prize in 2024 for STO prediction. And so um of course I was at this conference where they announced the winners. I came back and my lab was like oh why did I ever join your lab like this is going to be a disaster. I'm never going to get a job afterwards. This is this is horrible. And of course me being the optimist that I am I said don't worry guys is just the beginning. So we have Alpha Pole. It's a great new tool. Let's use it to be able to do new science that we couldn't do before. And so the big thing is like if I if I go back to the duck that we saw earlier, you see the duck with the with the wings tucked in and you never know that a D can expand its wing and then flap them and start flying, right? So knowing something about the structure or one state doesn't tell you anything about the function. If I show you this protein called ATP synthes, so what? Yeah, it looks colorful. It looks nice. What does it do? Well, this protein is actually a motor that it captures uh it's a hydrogen pump. So, it captures hydrogens and then it's able to make ATP, the molecule that uh allows you to carry energy many of the react to do many of the reactions in the body in the cell. And so, it is critical to know the structure. Yes. But that's the start to be able to understand how these molecules work. And one of the key aspects of how molecules these molecules work is the concept of molecular recognition. So if you think about your daily life, so this morning I was on my bike biking here then you know right now I'm holding this I might be holding a pen interacting with a keyboard. So in order to do different things we're interacting with different objects. In the same way, proteins in order to carry out their work, they need to interact with other molecules. Some proteins will interact with nucleic acids to maybe express information in our genes. Some will interact with other proteins or peptides or even small molecules like aspirin. And so understanding molecular recognition is key to many fields. In biio medicine to design novel molecules that can stop disease. In biotechnology to design uh custommade sensors or um or possible enzymes that can catalyze reactors that we care about. Biomeaterials to do for example selfhealing. uh you put two peptides together, they form a hydrogel that can uh capture cells to close a wound or uh to the sign novel functionality that we have never seen because evolution never thought of it. And so if I'm thinking about molecular recognition, it goes between two extremes. On one end, you have something that is called lock and key kind of interactions. So this is where you have a molecule. it might have like a cavity or something where a small molecule can bind. And on the other hand, you have what's called dynamic recognition. So proteins that might be intrinsically disordered, they can adopt many shapes and once they find the right binding partner, they adopt very well- definfined shapes. So if I had to put a cartoon representation, this is kind of what I thought about. And on the other side you have your hands which can adopt many different confirmations depending on what they're interacting with. Right? So one [snorts] of the things is knowing a structure is not enough. Now we need to understand how this works. So for a lot of the things that I do in my lab like we work with physics based models. So in physics like [snorts] you try to put a physics model you use the equations of motion Newton's equations of motion you propagate them through time and you see the wiggling and jiggling of atoms and if you look at this for long enough you might be able to understand the behavior of biomolelecules. The only problem is that these simulations are computationally expensive very very expensive. So right now about 20 to 30% of supercomputer time is devoted to these kind of simulations and it produces terabytes of data and yet understanding what these molecules do or reaching time scales that are relevant is hard and so how do we predict binding? So I'm going to tell you tell you about the two different regimes the lock and key regime. So how do we predict binding force a small molecule like aspirin and how do we predict if we have many different compounds which of them might be the one that binds best and so with physics it's very hard to predict things in absolute terms which one is the best but if I give you something like this horse so if I give you something like this horse it looks fast this horse also looks fast but if I were to say which one of them is faster, the best way is to do a horse race, right? And so we do exactly the same with our molecules. So when we have this lo and key interaction, we compete different compounds in order to understand which one of them might have the best binding affinity, bind their strongest. And so this is something that we use routinely in pharma. So in drug discovery, this is the early stages of drug discovery. It's widely used. You can process thousands of compounds, no problem. On the other hand, when you have flexible molecules, now that becomes a problem because not only you have to care about what molecules are interacting with which ones, you also have to care about what are the different shapes that they might adopt and what is the right one. And so it's a much more complex problem. So we've worked in this area. We've developed some methods based on physics. And basically the idea is that we have a protein target. we have two possible uh compounds. They're disordered when they're on their own and they might adopt like a helix like extractor when they bind. So we can make these comparisons. It takes us about 30 GPUs running for five days non-stop in order to make one of these predictions. Now imagine that I have a 100,000 of these compounds to test. There's no way I can do it with a computer. So as soon as Alpha came out, what did we do? We said well you know they design alpha fold to predict the structors of proteins but a peptide binding to a protein that's a similar kind of problem it might work and so people started using alpha for all kind of things that it wasn't designed to do. It turns out it turns out that it did pretty well. It did really well at predicting protein protein protein peptide and other kind of interactions. It was also really good not only at predicting structures but people started doing things like uh flow matching and whatnot to predict ensembles predict different confirmations away from the static structure. They also use it to design novel proteins and so from alpha fold that was designed to do only one thing a lot of different functionality emerged and that has opened the field to many many different kind of things. So still we were like okay but we don't care only about knowing what's where something binds. We care about knowing which of these compounds might be the best. And so we said if I give you all of these little compounds that have go from strong binders to weak binders will alpha know how they bind and which ones are the strongest? So it turns out that alphapold did very well at predicting that all of these fine but it could not tell us anything about which of them was the best. >> Can I ask a question? So these molecules live in some environment aquas or something else. Where is that information in here? >> So in the AI model that's implicit because it's been trained on um either NMR structures which already captured some environment crystal graphic environment environment. So they capture whatever environment they were in. Right? In a physics model, we would explicitly put the environment, either a membrane environment, either an aquis environment with ions, a crowded environment, so on and so forth. Thank you. >> So AI is very confident about these different peptides should bind, but we have no idea which one of them is best. And so, um, if I were to ask you, do you like Japanese snacks? If I ask you about this snack, do you like it? Pretty confident. Yeah. Do you like this one? Yeah. How about all of these? I'm pretty confident you like all of these. But being a 100% confident on each one of them doesn't tell me anything about which one you would choose, right? So, confidence is not the same as preference. So we said well if alpha fault is really confident about all these predictions could we ask it something about which one to choose and so we started thinking so my student leeway he started thinking about like well maybe in this uh function that alpha fold is encoding it has learned something that looks like a scoring function and could we instead of predicting one peptide at a time could we feed it two peptides at a time And now it has to make a decision which one are you more confident with this one being in the active site or the other one. And by doing that we could predict which peptides fine and which ones are the strongest. And the big game for us is that when we use physics we take 30 GPUs running for five days. When we do this AI approach it takes us one GPU running for five minutes. And so now we can scale to hundreds of thousands of compounds. So what that means is we can go get whole protein libraries, screen those protein libraries, uh cut them all into all their little pieces and identify which ones bind, which ones do not bind and which ones are are the best. So we're working with some immune response proteins. We're now understanding many proteins that are able to bind these immune response proteins from viral proteins from bacterial proteins and even host regulatory proteins. And so this is allowing us to to learn new many new things. Now one of the things that we are seeing is like well AI still has its own information. It's fast. It's really good but it might have some biases. So what our new paradigm is is that we're combining the speed of AI with the interpretability that physics based models are giving us and with that we're now able to make more predictive uh molecular modeling. So basically we have a front end in which we use AI at the sequence level then at the structural level and then we go to increasing degrees of complexity of physics-based models. When we are done with all our filters, we go down and we send our predictions to our experimental collaborators and with that we have been able to identify novel uh proteins and we have also been able to distinguish some proteins that were previously characterized and were not biologically relevant. All right. So so far I've told you about protein structure prediction the problem of molecular recognition and then the next stage that I want to tell you the last part of this talk is about going from a structure prediction to design right so the idea of design is the inverse problem as protein folding so in protein folding for protein structure prediction we have a sequence and we want to know what is the 3D structure that it adopts And that has been mostly solved by AI in um in protein design. What we want to know is if I have a 3D structure of a protein, what is the sequence that is best suited to adopt that structure? And so the structure could be a structure that it's already known or it could be a noble structure, a structure that is hallucinated for example by alpha 4. And so the main idea here is that during evolution things are things evolve until a certain point. So you have some parameters. For example, most of our proteins have to be as at least stable at 37 Celsius. when you have a high fever, you are putting some cold eyes on your forehead or whatn not to lower the temperature so your proteins don't unfold or what not. And so with designed proteins, you can now go to proteins that are stable maybe 80, 90 Celsius, what not. And so one of the biggest players in the field is David Baker. For the longest time, designing proteins was a task that very few groups in the world could do. David Baker was the maximum exponent of what we could do and before AI they were already doing very impressive things. So this is uh one of the designs that they did. It's basically a protein that is able to fold into a 3D structor. It's able to assemble itself with other proteins of its type and make this cage like kind of compound. So you have 40,000 atoms here that when you solve the experimental structure, those 40,000 atoms are on top of the design. That's amazing. So the challenge is that only one or two groups were really good at doing this. The success rates were not high and it it required a lot of manual intervention. So it was very artisal in some way. And so the idea is could AI help in this front as well. And so from 2015 where Baker lab was the maximum exponent to today what we're seeing is that we have now democratized um design. So basically any lab can now within one or two days say hey I want to design a protein that binds this particular target. And within two days you have 10,000 designs. Now are all those designs good? No. You still have to find out what are the best. The success rates are 10 to 20%. And programs are getting better and better. But like having like this AI machinery has helped a lot. So this is the Nobel Prize. 50% went to David Baker for design and 50% to Disabon Jumper for sector prediction. So how does that how does the design work? So it works pretty much the same way as in the image generation field. So we use what's called diffusion models. So the idea is that if I have a picture of a dog, it's made out of pixels, right? I can randomly change a few of the pixels in this picture and I will get something that still looks like a dog, but not quite. I can learn the process of going from this picture to this picture, but also from this picture back to this picture, right? And I can do that again and I can do that again and again and again until I get to a random set of pixels. Now if I have this set of pixels, how do I change them to get a new image? I have no idea. But if I have a random set of pixels, I know how to generate another random set of pixels. And if I've learned the process of going from a random set of pixels back to an image, I can say, well, this is my new set of random pixels. And if I go back, oh, now I get a cat. So, of course, to do that, you have 10 of many, many, many images. Yeah. So, I can do exactly the same with proteins. I have about 300,000 proteins. I know the position of all the atoms in the protein. I can add a little bit noise to the positions of those atoms. And I can learn that step. I can also learn the step of going back. And I can go all the way until I get a random distribution of atoms in space. And from that random distribution, I can generate a new random distribution. Go back and generate a new protein. So that's how design works. Nowadays once I have the fall, I calculate what is the sequence that is best suited for this uh protein design. And so again like we've been using the latest AI technology combined with our physics based pipeline. So basically we start we use diffusion models in order to predict as um new proteins that will bind a particular target of interest. Then we do all of our AI pipeline to filter out which we think are the best designs. And then finally we use our physics based models to say like do these two things that have been completely orthogonally uh designed do they agree or do they differ? When they agree we send those for experimental validation and this has been successful successful in many cases. I started the talk by giving you an example about blindness structure prediction. There have been new challenges for designing proteins. So this was one that we participated a little while ago. It's called the biome challenge. It's to binders. So the idea is they gave us this target. So it's this is a cancer cell. It has a tumor antigen and usually a T- cell will have to recognize that antigen, bind to it, destroy the the the cancer cell. Right? So they said, could you design a mini protein that would bind this antigen? And if you can, then what we're going to do is we're going to take that mini protein, we're going to put it in a T- cell. So we're going to have a functionalized T- cell. This is called a CART cell. And so once once we have that, what we're going to say ask is, can those designs be successful at killing these cancer cells? So basically that's what they tked us to do. About uh 12,000 sequences from around the world were submitted. Each team could submit 500 sequences. So once they gather all the sequences then they did all these kind of functional assays. I have no idea about these functional assays exactly what they are. The basic idea is they tested whether the carti cells were able to prol proliferate in the presence of cancer cells. If these carti cells were able to um send signaling molecules that would activate um the the killing of the of the cancer cells. And finally, what was the ability of these T- cells uh to kill cancer cells without killing regular cells? And it turns out that our submission was the top submission in the whole competition. So it was really nice. They 3D printed uh awards for the top teams and they sent it to us. So this was the mini protein that we designed and this is how it binds its uh uh cancer antigen. And so like this was a nice blind competition. One of the biggest surprise is that the best binder was not the one that won. The best binders actually create exhaustion and so it doesn't allow cancer cells to be killed. Um and so this was one of the lessons that we learned along the way. So by now I have told you three things. I told you about protein structure prediction about how that's not enough. We need to understand molecular recognition how these proteins interact with uh different molecules and the hope is that we can use that to understand how proteins function. So basically this challenge of protein structure prediction is mostly done. A lot of molecular recognition is now what what's going on in our field like there are many things that we cannot still understand using AI models for example things of like how do cryptic pockets emerge or how does a lost happen like those are too complex for most of the AI models that we have right now and the future where everyone is trying to go is how do we uh look into protein function how do we go from a sequence to function and so um with that I just wanted to finish with picture of my group. Really happy to be here. I'll take any questions. Thank you. >> Sure. >> Yeah, please if you have any questions um just please wait for the microphone for the people on Zoom. >> Yeah. Thank you very much. It was so nice to see such a good presentation. Uh so I have two questions. The first question is about the approach when you combine AI and physics based approach. So what is the lack of just using AI and the sequence? Um what's the lag there? What we take from the physics? Is it just the more parameters? Because like physics is kind of embedded in the sequence I would say. as well. >> All right. So the main thing for us, so whenever you use an AI model, one of the bigger biggest question that reviewers can ask you is, oh, what about memorization? Like did you get the right answer because it's making a prediction or is it because it has already seen this and so it's giving you the right answer for the wrong reason. If it's giving you the wrong answer for the wrong reason, then it's gonna make predictions that are false positives. when you use our physics based model because it has it hasn't been trained on data. It just uses like you know how do bonds stretch, how do angles change and so on and so forth. If the answer with our physics model agrees with the AI model then that increases our confidence. We could still be wrong but we're much more confident because they have completely separate origins. >> Okay. And I have the second question also short one about the diffusion and the protein design. So are there any like common cases when it doesn't work >> or just by chance? >> Yeah. So so all right. So if I were to say okay I'm going to use RF diffusion to predict binders to this molecule and then I'm going to design the best sequences that are able to adopt this sector. The success rate is expected to be 10 to 20%. So that means one of one every five will fail. Of course, you can only test experimentally. Like my collaborators will try at most five, right? So if I choose five that are incorrect, they're never going to collaborate with me again. So we need to make sure as best as we can that what we are giving them is enriched from having one every five correct to three or four every five correct. And so that's why we do this complicated machinery. There are so deep mind has done a new AI that is proprietary. So this is not available to the I think it's called alpha proteio or something like that. So this one is not available but this one is much much much better at making. So it goes from 10 to 20% to a much higher percentage of success. >> Yeah. My question was more about like we know that let's say basic alpha fault doesn't work well for uh instly disordered proteins let's say maybe isn't this like the case for diffusion models as well >> so for disorder proteins for predicting and designing disorder proteins I don't think it's going to look very good >> like I think like there are some AI tools that are starting to predict how to bind disorder regions. >> Yeah. >> But not necessarily to make disorder region. I mean like there are other tools that can give you disordered regions. But usually I think the state-of-the-art right now is I have something that is disorder. I will use NMR and I will do a population of states rewe them so that they satisfy the NMR data. That's I think the best that you can. >> Thank you. Thank you Albert. That was really nice to see and um so so this is probably a very simple question but still you had in the beginning the like a landscape of energies and you you talked about the the native state when the lowest energy right and then you also had one question that was still you working on the how how they can solve this hard problem so fast in reality I guess I mean do do they always find the lowest one or or is No or I mean >> so first not all protein sequences will have a folded state >> some of these proteins will have like a more complicated landscape where like you have like maybe a misfolded state that can aggregate and so it they don't find they find this aggregated state before they find the native state. So you have like um um what are they called chaperons that will help fold some proteins that are misbehaving. So some of them need help. M >> but eventually they will find the absolute smallest one or >> well and when they don't we get all kind of diseases right like Alzheimer's Parkinsons and what not >> thanks I'll be here six more weeks so I'll be two floors above if you have questions >> can I add one more and then >> yeah just >> yeah okay so the I remember when that they got the Nobel Prize but you know the because it's a statistical you know method in essentially that >> I would not call it a statistic >> okay but I remember the statement was that now we understand how proteins are folding I it's a good predictor but it's not the explanation right >> it's not the explanation it's a black box still >> so all right let me backtrack a little bit so when I said the protein folding problem is these three questions. >> So for me the protein folding problem is not solved. The protein structure prediction problem is mostly solved. >> Right? >> And so depending on who you ask in the field, they will not distinguish those two things. For me protein folding is understanding what are the myriad pathways that will take you from unfolded state to the folded state. Each one of those pathways will have different kinetics and different time scales, but on average altogether they make life possible. So it's robust enough that on average a protein will be folded every x number of units >> because of that understanding how that is actually the path to taking will make us maybe think about drugs that prevent miscaping. Right? Yeah. So, so I think it's good to have the difference between understanding and making predictions. >> Yeah. >> Thanks. >> So, like thank you very much for the lecture. That's like what you come from the simple explanation to the complex ideas and that's that's very nice. Uh I have a question about like you want to know how to come from the structure to sequence and uh maybe why do we need it right now? Maybe we needed first to understand how the structure calls the function and then understand how to come from the structure to signals, >> right? It would it would be very so like let's go back to the to the D, right? Like say alpha fold for the DAG. It would predict the D with its wings stacked. That doesn't tell me anything about what what is it that the D needs in order to fly. Right? I need to understand this other state which is the flap the wings open and down and up. So those are the three states I need to understand for that right. If I understand those three states and what are the kinetics between them then I can understand how it functions how it flies. >> Yes. So so understanding the u like the path from the structure to sequence is the key for understanding the structure to function. So one more time. So going from structure to sequence is key to >> understanding structure how the we can define structure function function of the partition. >> Not necessarily I mean like I I think like they are independent tools but all together we will get more knowledge to try to understand function. So what is the degenerate code of sequences that are able to fall to a particular structure and what are the states that they are compatible with and like how does that relate to function that will be interesting but you know I think it all plays together but not necessarily we need this to learn the other thing. >> Yeah thank you very much. Yeah. >> Any further questions? >> Um, well, if not, yeah, thank you very much again for thank you for coming.