AlphaFold 2는 아미노산 서열로부터 단 몇 초 만에 3D 구조를 예측하여 단백질 폴딩에 혁명을 일으켜 구조 생물학을 변화시켰습니다.

### 단백질 접기의 획기적인 발전: 알파폴드 혁명
* 알파폴드 2는 한때 불가능하다고 여겨졌던 **단백질 폴딩을 정확하게 예측**하여 구조 생물학 및 과학 전반에 기념비적인 발전을 이루었습니다.
* 🔬 생물학의 핵심인 단백질은 아미노산 서열에 의해 특정되며, **단백질의 3D 구조는 생물학적 과정과 약물 상호작용에서 그 기능을 직접적으로 결정**합니다.
* 📊 단백질 접힘 문제는 **1972년 노벨상 수상자인 크리스티안 안핀센이 생물학에서 페르마의 마지막 정리에 비유**하며 그 복잡성을 처음으로 설명한 이래 **50년 넘게 과학자들의 도전 과제**였습니다.
* ⏳ 기존에는 단백질의 구조를 결정하려면 단백질 하나를 결정화하고 시각화하는 데 박사 과정 학생 전체가 몇 년이 걸리는 고난도 실험 방법이 필요했지만, **AlphaFold 2는 단 몇 초 만에 3D 구조를 예측**할 수 있습니다.
* 🌍 크리스마스 기간 동안 AlphaFold 2는 **인간 프로테옴의 모든 2만 개 단백질의 구조**를 성공적으로 예측하여 구조 생물학자들이 구글 검색처럼 단백질 구조를 '조회'할 수 있게 함으로써 연구 방식에 혁신을 가져왔습니다.
* ⚙ 알파폴드의 개발에는 **30개가 넘는 알고리즘을 통합**하고, 물리학 및 진화 생물학의 제약 조건을 활용하고, **자체 증류를 통해 학습 집합을 혁신적으로 확장**하여 예측 기능을 강화하는 과정이 포함되었습니다.
* 🔄 거리 매트릭스를 생성했던 이전 버전인 AlphaFold 1과 달리 **AlphaFold 2는 아미노산 서열을 3D 구조로 직접 변환하는 엔드투엔드 방식**으로 작동하여 인간 설계자보다 더 효과적으로 제약 조건을 학습하는 시스템의 능력을 보여줍니다.
* 🌌 알파고에서 알파폴드로의 여정은 **모방 학습에서 자기 지도 학습으로 AI가 진화하는 과정**을 보여주며, 과학적 발전을 추구하기 위해 복잡한 시스템을 구축하고 이해하는 것이 중요하다는 점을 강조합니다.

So, here's another problem that was thought to be impossible to solve, which is protein folding, and AlphaFold, or specifically AlphaFold 2, did just that. It solved protein folding. I think it's one of the biggest breakthroughs, certainly in the history of structural biology, but in general in science. Maybe from a high level, what is it and how does it work? And then we can ask some fascinating questions after. Sure. So, maybe to explain it to people not familiar with protein folding is, you know, I first of all explain proteins, which is, you know, proteins are essential to all life. Every function in your body depends on proteins. Sometimes they're called the workhorses of biology. And if you look into them, and I've, you know, obviously as part of AlphaFold, I've been researching proteins and structural biology for the last few years, you know, they're amazing little bio-nano machines, proteins. They're incredible. If you actually look at little videos of how they work, animations of how they work, and proteins are specified by

their genetic sequence called their amino acid sequence. So, you can think of it with their genetic makeup, and then in the body, in nature, they, when they fold up into a 3D structure. So, you can think of it as a string of beads, and then they fold up into a ball. Now, the key thing is you want to know what that 3D structure is, um, because the structure, the 3D structure of a protein, uh, is what helps to determine what does it do, the function it does in your body, uh, and also if you're interested in drug drugs or disease. You need to understand that 3D structure because if you want to target something with a drug compound about to block something the protein's doing, uh, you need to understand where it's going to bind on the surface of the protein. So, obviously, in order to do that, you need to understand the 3D structure. So, the structure is mapped to the function. The structure is mapped to the function, and the structure is obviously somehow

specified by the amino acid sequence, and that's the protein folding problem. Can you just, from the amino acid sequence, the one-dimensional string of letters, can you immediately computationally predict the 3D structure, right? And this has been a Grand Challenge in biology for over 50 years. So, I think it was first articulated by Christian Anfinsen, a Nobel Prize winner in 1972, as part of his Nobel Prize-winning lecture. And he just speculated this should be possible to go from the amino acid sequence to the 3D structure, but he didn't say how. So, I, you know, it's being described to me as equivalent to Fermat's Last Theorem but for biology, right? You should, as somebody that very well might win the Nobel Prize in the future, but outside of that, you should do more of that kind of thing in the margin, just put random things that will take like 200 years to solve, set people off for 200 years. It should be

possible and just don't give any DET exactly. I think everyone exactly should be. I'll have to remember that for future. So yeah, so he said off, you know, with this one throwaway remark just like Fermat, you know, he set off this whole 50e field really of computational biology. And they had, you know, they got stuck, they hadn't really got very far with doing this and um until now, until Alpha came along. This is done experimentally, right, very painstakingly. So the rule of thumb is, and you have to like crystallize the protein, which is really difficult. Some proteins can't be crystallized like membrane proteins. And then you have to use very expensive electrom microscopes or x-ray crystallography machines, really painstaking work to get the 3D structure and visualize the 3D structure. So the rule of thumb in experimental biology is that it takes one PhD student their entire PhD to do one protein. And with Alpha 2, we were able to predict

the 3D structure in a matter of seconds. Um, and so we were, you know, over Christmas, we did the whole human proteome or every protein in the human body or 20,000 proteins. So the human proteome is like the equivalent of the human genome but on protein space and, uh, and sort of revolutionize really what a structural biologist can do because now, um, they don't have to worry about these painstaking experimentals, you know, should they put all of that effort in or not. They can almost just look up the structure of their proteins like a Google Search. And so there's a data set on which it's trained and how to map this amino acid. First of all, it's incredible that a protein, this little chemical computer, is able to do that computation itself in some kind of distributed way and do it very quickly. That's a weird thing. And they evolve that way because, you know, in the beginning, I mean, that's a great invention, just the protein itself of Earth, and then they, there's, I think

probably a history of like, uh, they evolved, uh, to have many of these proteins, and those proteins figure out how to be computers themselves in such a way that you can create structures that can interact in complex with each other in order to form high-level functions. I mean, it's a weird system that they figured it out well for sure. I mean, we, you know, maybe we should talk about the origins of life too, but proteins themselves, I think, are magical and incredible, uh, uh, uh, as I said, little bionano machines, and um, and, and actually Leventhal, who was another scientist, uh, contemporary of Anenson, uh, he, he, he coined this Leventhal, what became known as Lal's Paradox, which is exactly what you're saying. He calculated roughly a pro, an average protein, which is maybe 2,000 amino acids bases long, is um, is, is can fold in maybe 10 to the power 300 different confirmations. So, there's 10 the^ 300 different ways that

protein could fold up, and yet somehow in nature, physics solves this in a matter of milliseconds. So, proteins fold up in your body in, you know, sometimes in fractions of a second. So, physics is somehow solving that search problem. Maybe you can give some interesting sort of beautiful things to you about these early days of AlphaFold, of solving this problem, because unlike games, this is real physical systems that are less amenable to self-play type of mechanisms. Sure, the size of the dataset is smaller than you might otherwise like, so you have to be very clever about certain things. Is there something you could speak to, um, what was very hard to solve and what are some beautiful aspects about the solution? Yeah, I would say AlphaFold is the most complex and also probably most meaningful system we've built so far. So, it's been an amazing time, actually, in the last, you know, two, three years to see that come through because, um, as we talked about earlier, you know, games is what we started on, uh, at building things.

like AlphaGo and AlphaZero, but really the ultimate goal was to, um, not just to crack games. It was just to build, use them to bootstrap General Learning Systems we could then apply to real-world challenges. Specifically, my passion is scientific challenges like protein folding, and then AlphaFold, of course, is our first big proof point of that. And so, um, you know, in terms of the data, uh, and the amount of innovations that had to go into it, we, you know, it was like more than 30 different component algorithms needed to be put together to crack the protein folding. Um, I think some of the big innovations were that, um, kind of building in some hardcoded constraints around physics and evolutionary biology, um, to constrain sort of things like the bond angles, uh, uh, uh, in the protein and things like that. Um, a lot, but not to impact the learning system, so still allowing, uh, the system to be able to learn the physics, uh, itself, um, from the examples that we had. And the examples, as you say, there are

only about 150,000 proteins. Even after 40 years of experimental biology, only around 150,000 proteins have been. The structures have been found out about, so that was our training set, which is much less than normally we would like to use, but using various tricks, things like self-distillation, so actually using AlphaFold predictions, some of the best predictions that it thought was highly confident in, we put them back into the training set right to make the training set bigger. That was critical to AlphaFold working, so there was actually a huge number of different innovations like that that were required to ultimately crack the problem. AlphaFold 1, what it produced was a disr, so a kind of a matrix of the pairwise distances between all of the molecules in the protein, and then there had to be a separate optimization process to create the 3D structure. What we did for AlphaFold 2 is make it truly end-to-end, so we went straight from the amino

acid sequence of uh bases to uh the 3D structure directly without going through this intermediate step. And in machine learning, what we've always found is that the more end-to-end you can make it, the better the system. And it's probably because, um, we, you know, in the end, the system is better at learning what the constraints are than we are as the human designers of specifying it. So anytime you can let it flow end-to-end and actually just generate what it is you're really looking for, in this case, the 3D structure, uh, you're better off than having this intermediate step, which you then have to handcraft the next step for. So, um, so it's better to let the gradients and the learning flow all the way through the system, um, from the endpoint, the end output you want, to the inputs. So that's a good way to start on a new problem, handcraft a bunch of stuff, add a bunch of manual constraints with a small intent learning piece or a small learning piece, and grow that learning.

piece until it consumes the whole thing. That's right. And so you can also see, you know, this is a bit of a method we've developed over doing many sort of successful Alpha. We call them alpha x projects, right? Is, and the easiest way to see that is the evolution of alpha go to Alpha zero. So Alpha go was a learning system, but it was specifically trained to only play go, right? So, uh, and what we want to do in with the first version of alpha go is just get to world champion performance, no matter how we did it, right? And then, and then of course Alpha go zero, we, we, we remove the need to use human games as a starting point, right? So it, it could just play against itself from random starting point from the beginning. So that removed the, the need for human knowledge, uh, about go. And then finally Alpha zero then generalized it so that any things we had in there, the system, including things like symmetry of the go board, uh, were removed so the alpha zero could play from scratch any two-player game. And then muzero, which is

the final, our latest version of that set of things was then extending it so that you didn't even have to give it the rules of the game. It would learn that for itself, so it could also deal with computer games as well as board games. So, that line of Alpha Go, Alpha Go Zero, Alpha Zero, Mu Zero, that's the full trajectory of what you can take from imitation learning to full self-supervised learning. Yeah, exactly, and learning the high structure of the environment you're put in from scratch, right? And bootstrapping it through self-play, yourself. But the thing is, it would have been impossible, I think, or very hard for us to build Alpha Zero or Mu Zero first out of the box, even psychologically, because you have to believe in yourself for a very long time. You're constantly dealing with doubt because a lot of people say that it's impossible to exact. So, it's hard enough just to do Go, as you were saying. Everyone thought that was impossible or at least a decade away.

from when we did it back in 2015, 204, you know, 2016, and um, and so yes, it would have been psychologically probably very difficult, as well as the fact that, of course, we learn a lot by building alphao first, right? So it's, I think this is why I call AI an engineering science. It's one of the most fascinating science disciplines, but it's also an engineering science in the sense that, unlike Natural Sciences, um, the phenomenon you're studying doesn't exist out in nature. You have to build it first, so you have to build the artifact first, and then you can study how and pull it apart and how it works.

노벨 화학상 딥마인드 데미스 하사비스 인터뷰 (알파폴드)

BZCF | 비즈까페

알파폴드의 획기적인 단백질 접기 기술

알파폴드 2는 50년 이상 생물학의 핵심 난제였던 단백질 폴딩이라는 오랜 문제를 성공적으로 해결했습니다.

모든 생명체에 필수적인 단백질은 아미노산 서열에 따라 3D 구조로 접혀 그 기능을 결정합니다.

이 3D 구조를 이해하는 것은 약물 표적화 및 질병 치료에 매우 중요하므로, 알파폴드의 성과는 구조 생물학 및 과학 전반에서 중요한 돌파구가 될 것입니다.

알파가 단백질 구조 예측에 미치는 영향

알파의 등장은 기존에는 박사 과정 학생들이 몇 년씩 걸리던 단백질의 3D 구조를 단 몇 초 만에 예측함으로써 계산 생물학 분야에 혁명을 일으켰습니다.

2만 개의 단백질로 구성된 인간 프로테옴이 단기간에 매핑되어 구조 생물학자들의 접근 방식이 바뀌었습니다.

복잡한 계산을 수행하고 서로 상호 작용하는 단백질의 독특한 능력은 단백질의 진화적 성공과 고차원적 기능 개발의 핵심이었습니다.

알파폴드의 복잡성과 아름다움

AlphaFold는 혁신적인 알고리즘과 하드코딩된 제약 조건을 사용하여 복잡한 문제를 순식간에 해결함으로써 단백질 폴딩에 혁명을 일으켰습니다.

훈련 세트에 단백질이 15만 개에 불과한 AlphaFold는 자체 증류 및 기타 트릭을 사용하여 학습을 향상시켰습니다.

AlphaFold 1의 쌍별 거리 매트릭스에서 AlphaFold 2의 엔드 투 엔드 접근 방식으로의 전환은 3D 단백질 구조 예측에 상당한 진전을 이루었습니다.

엔드투엔드 학습은 시스템이 중간 단계 없이 3D 구조와 같은 원하는 결과물을 직접 생성할 수 있으므로 더 효과적입니다.

이 접근 방식은 시스템이 인간 설계자보다 제약 조건을 더 잘 학습하는 데 도움이 됩니다.

이 방법은 알파고에서 알파제로로 발전한 알파 프로젝트에 성공적으로 적용되어 인간의 지식이 필요 없는 2인용 게임으로 학습을 일반화했습니다.

최신 AI 버전은 모방 학습에서 자기 지도 학습으로 전환하면서 게임 규칙을 스스로 학습하도록 진화했습니다.

알파고, 알파고 제로, 알파 제로, 뮤 제로 등이 이러한 궤적을 따르며 셀프 플레이와 부트스트랩의 힘을 보여줬습니다.

이러한 모델을 개발하는 여정은 심리적으로 매우 어려웠으며, 잠재력에 대한 강한 믿음과 초기 버전으로부터의 학습이 필요했습니다.