构建通用 AI 药物设计引擎：从 AlphaFold 3 到超人类创造力

我们从公司成立第一天起就立志追求这个宏伟目标。这不是针对某个特定适应症或靶点开发疗法，而是思考如何打造一个非常通用的 AI 药物设计引擎——不仅能应用于单一靶点甚至单一模态，还能反复应用于任何不同的疾病领域。这就是我们目前正在推进的方向。

We have set up the company from day one to really go after this big ambition. This isn't about developing therapeutics for a particular indication or a particular target. It's really thinking about how do we create a very general drug design engine with AI, something that we can apply to not just a single target or even a single modality, but we can apply this again and again across any different disease area. And that's what we're stepping towards at the moment.

今天，我们非常高兴地邀请到 Isomorphic Labs 的首席 AI 官 Max Jaderberg。Isomorphic Labs 脱胎于 DeepMind，旨在利用 AI 彻底改变药物发现。去年夏天，他们发布了 AlphaFold 3，这一惊人突破使我们不仅能建模蛋白质结构，还能建模所有分子及其相互作用。这促使 Demis Hassabis 去年获得了诺贝尔化学奖。Max 描述了他们对药物设计圣杯模型以及科学智能体的愿景。他将自己的经历与构建 AlphaStar 和夺旗游戏以及更广泛的智能体与游戏研究方向进行了类比。具体来说，面对 10^60 种可能的药物分子结构，我们需要构建生成模型和能够学习探索并搜索整个潜在设计空间的智能体。Max 还描述了他对该领域 GPT-3 时刻的愿景，将其比作 AlphaGo 著名的第 37 手——当我们开始看到 AI 药物设计中展现出超人类创造力、甚至令人类自身惊叹的事物时。这是我最喜欢的节目之一。享受节目。Max，非常感谢你今天在伦敦加入我们。

Today we're excited to welcome Max Jaderberg to the show, chief AI officer of Isomorphic Labs, which launched out of DeepMind with a goal of revolutionizing drug discovery using AI. Last summer, they released AlphaFold 3, a stunning breakthrough that allows us to model not just the structure of proteins, but of all molecules and their interactions with each other. That led to Demis Hassabis winning the Nobel Prize in chemistry last year. Max describes their vision for what a holy grail model for drug design and what agents for science look like. He draws parallels to his experiences building AlphaStar and Capture the Flag and the research directions of building agents and games more broadly. Specifically, with 10^60 possible drug molecule structures, we need to build both generative models and agents that can learn how to explore and search through the whole potential design space. Max also describes his vision for what a GPT-3 moment for the field might look like, describing it more akin to AlphaGo's famous move 37 when we start to see things that exhibit superhuman levels of creativity in AI drug design and that stun even humans ourselves. This is one of my favorite episodes yet. Enjoy the show. Max, thank you so much for joining us today here in London.

不，很高兴能和你在这里。

No, it's a pleasure to be with you here.

是的，太棒了。时机也恰到好处，AlphaFold 3 的发布和 Demis 获得诺贝尔化学奖，这确实证明了您和您的团队在过去几年所做的一切。

Yeah, it's fantastic. Awesome timing too with the launch of AlphaFold 3 and with Demis winning the Nobel Prize in chemistry, which is a true testament to everything that you and your team have done over the last couple years.

是的，2024 年对我们来说绝对是忙碌的一年。许多重大突破。诺贝尔奖真是令人难以置信。我认为这是对这项开创性工作的惊人认可。

Yeah, 2024 was definitely a busy year for us. Lots of big breakthroughs. Nobel Prize was just incredible to see. I think amazing recognition for this seminal piece of work.

是的。我想先聊聊您的个人经历。您从一开始就在深度学习领域有着非凡的职业生涯，在 DeepMind 期间撰写了多篇开创性论文，包括夺旗和 AlphaStar，这些都是深度学习领域的突破。您能谈谈当时在深度强化学习研究领域的一些关键问题吗？

Yes. Well, I'd love to start with talking a little bit about your own personal story. You've had an incredible career in the world of deep learning from the very start, authoring many seminal papers while at DeepMind including for Capture the Flag and AlphaStar, breakthroughs in the world of deep learning. Can you walk us through some of the key questions that you had in your field of research around deep reinforcement learning at the time?

是的。在 DeepMind，我研究了很多东西，早期包括计算机视觉和深度生成模型。但真正让我着迷的是强化学习。当时 DeepMind 是世界上研究强化学习的最佳地点。我们心中的问题是：我们如何才能达到让 AI 能够去执行任何你希望它完成的任务？当时的主导范式是监督学习。监督学习与强化学习非常不同。它们都是学习技术，但监督学习需要知道问题的答案，并用它来训练模型。在监督学习中，你给出一个示例，然后向模型提供该问题的答案。如果你已经知道训练这个 AI（这个神经网络）要解决的所有问题，那很好，但大多数时候你并不知道。世界上有很多问题我们不知道答案，不知道解决方案。想想我希望 AI 如何应用于世界：是的，在我们人类已经擅长的领域应用 AI 会很棒，但真正的前沿是，我们能否开始将 AI 应用于人类不知道如何做或人类表现存在局限的领域？这正是强化学习作为关键工具之一并具有真正前景的地方，因为在强化学习中，你不需要知道问题的答案。你只需要能够判断模型给出的答案是好还是不好，甚至可能判断有多好或多不好。这开辟了一个全新的问题领域来训练这些模型。因此，从 DeepMind 早期在 Atari 等游戏上的重大突破开始，强化学习面临的问题是：我们如何将其从 Pong 和太空入侵者的世界扩展到真正看起来像现实世界问题的事物？于是，随着我们扩展这些方法，出现了一条精彩的研究轨迹。

Yes. So at DeepMind, I worked on a whole host of stuff, early days of computer vision and deep generative models. But it was really reinforcement learning that ended up hooking me there. DeepMind was the place in the world to be working on reinforcement learning at that time. And really the question in our minds was how can we actually get to a point where we could get an AI that could go off and do any task you wanted it to do. And the dominant paradigm at that point in time was supervised learning. Supervised learning is very different from reinforcement learning. They're both learning techniques. But supervised learning, you need to know what the answer to your question is and that's how you train the model. So in supervised learning, you give an example and then you supply the model with the answer to that question. Now that can be great if you already know everything about the problem that you're training this AI to do, this neural network to do, but most times you don't. I mean there are just so many problems in the world where we don't know what the answer is. We don't know what the solution is. And if you think about how I want AI to be applied to the world, yes, it's going to be great to be able to apply things where we're already good as humans here, but really the big frontier is can we start applying AI to places where humans don't know how to do this stuff or there's a limit to human performance there. And that's where reinforcement learning is one of the key tools and has real promise here because in reinforcement learning you don't need to know what the answer to the question is. You just need to be able to say whether the answer that the model gave you was good or not good. Maybe even how good or not good. And so this opens up a completely new field of problems to train these models against. And so reinforcement learning, starting from one of the big breakthroughs of DeepMind in the early days, working on games like Atari. The question was okay so how can we scale this up from the world of Pong and Space Invaders to things that really start to look like real problems in the world. And so there was an amazing track of research as we scaled up these methods.

您知道红杉资本当年是 Atari 的第一个投资者吗？

Did you know that Sequoia was the first investor in Atari back in the day?

哦，真的吗？我不知道。太不可思议了。

Oh, really? I didn't know that. That's incredible.

是的。不，那些 Atari 游戏实际上很有趣，可以回顾并玩一下，比如我们有一个智能体，我顺便玩一局 Pong。红杉办公室有一面很棒的墙，上面列出了所有传奇的 IPO 和并购名称。有一个我称之为“披萨公司”，我喜欢问人们是否知道那是什么，实际上它来自 Chuck E. Cheese's，当时是红杉的原始投资。太棒了。

Yeah. No, those Atari games were great fun actually to go back and play in the context of, hey, we've got an agent and I'm just going to have a game of Pong on the side as well. So there's a wonderful wall at Sequoia in our office where we have all these names of legendary IPOs and M&As that have happened. There's one I think it's called the pizza company and I love asking folks if they know what that is and it's actually from Chuck E. Cheese's which was an original Sequoia investment at the time. Amazing.

太棒了。

Amazing.

那么，夺旗和 AlphaStar 在当时是令人难以置信的突破。您能分享一下这些突破具体是什么，以及为什么选择这些特定游戏吗？

So Capture the Flag and AlphaStar were incredible breakthroughs at the time. Can you share a little bit about what exactly those breakthroughs were and maybe why you chose those specific games?

是的。如果你思考 AI 使用电子游戏的历史，我们为什么要用电子游戏？电子游戏是那种可塑的、完美封装的世界，作为研究人员和科学家，我们可以操控它们，在其中测试不同的算法，设置不同的情境。因此，它们是我们开发新算法的完美试验场。然后你可以想象，作为一名强化学习研究员，一个思考如何让 AI 尽可能通用的人，你总是在想：好吧，我们攻克了 Atari，如何获得更复杂的游戏？我个人痴迷的是，我希望这些智能体能够完成任何任务。

Yeah. So if you think about the history of AI using video games, why do we use video games at all? Video games are these sort of malleable, perfectly encapsulated worlds that as researchers and scientists, we can manipulate them. We can test out different algorithms in them. We can set up different situations. So they are the perfect test ground for us to develop new algorithms. And then you can imagine as an RL researcher, as someone who's thinking about how can we get AI to be as general as possible, you're always thinking, okay, we've cracked Atari, how do we get a more complex game? And the thing that I was personally obsessed with is I want these agents to be able to do any task.

这与当时人们在 Atari 上训练的做法略有不同。通常强化学习的思路是：这里有一个游戏，你训练它并变得擅长，然后将同样的算法从头开始训练在不同的游戏上。我更喜欢另一种场景：我们训练一个智能体，然后可以把它迁移到任何新任务上，并且这个智能体无需额外训练就能在该任务中表现良好。要做到这一点，你真正需要的是跨任务空间的泛化能力。这意味着你需要大量的训练任务。因此，在面向智能体的强化学习中，训练数据变成了任务，而不是图像或文本片段，而是任务。你可以想象，你可以组建一个完整的游戏工作室，手动设计数百个不同的任务，在这些虚拟世界中制作许多小游戏。我们确实这样做了，而且做了很多。然后你可以想，我们实际上可以超越手动设计：我们可以程序化地生成这些任务和游戏，生成世界、地图和不同的目标，我们也这样做了。但你会不断遇到复杂度的天花板，人类能手动设计或创造的复杂度是有限的。而多人游戏正好解决了这个问题。因为一旦从单人游戏转向多人游戏，就不再只是智能体在玩，游戏中还有另一个玩家。那个玩家或其他玩家可以表现出许多不同的特征和行为。因此，你面对的每一个不同的玩家、每一种不同的策略，都会从根本上改变游戏以及智能体试图完成的任务。我回想一下，为什么人们仍然痴迷于下棋？为什么职业棋手还在不断下棋？棋局看似相同，但实际上不同，因为你每天面对的是完全不同的对手，而且新棋手不断涌现。所以游戏在不断变化。因此，多人游戏和多智能体游戏真正包含了仅因其他玩家存在而可能遇到的巨大任务多样性。所以《夺旗》实际上是我们首次尝试利用多人游戏来真正拓展强化学习算法的能力，迫使我们深入思考如何泛化到新任务，以及如何处理多智能体动态。《夺旗》是一个了不起的突破。它真正展示了我们可以在这些多人第一人称游戏中达到人类水平的表现。当然，随后《星际争霸》增加了巨大的复杂性，成为我们接下来必须攻克的前沿。

And this is a slightly different paradigm from what people were doing at the time with training on Atari, where normally in reinforcement learning you think: here's a game, now you train on it and get good at it, and then you apply that same algorithm from scratch training on different games. I'd love a different scenario where instead we train an agent and then we can lift it and put it on any new task. And that agent will be able to perform well in that task without any more training. So to do that, what you're really asking for is generalization over task space. And that means you need lots and lots of training tasks. So the training data in this RL for agents becomes tasks, not images, not pieces of text, but tasks. So you can imagine you could go and sit and take a whole game studio and try to hand-author hundreds of different tasks, lots of little mini-games in these virtual worlds. And we did that, we were doing lots of that. And then you can think, we can actually go further than hand-authoring: we can procedurally generate these tasks and games, generating worlds and maps and different objectives, and we did that. But you keep running into this complexity ceiling, that there's only so much complexity that you can hand off or you can design humanly. But that's where multiplayer games come in. Because as soon as you go from single player to multiplayer, it's not just the agent playing; you've got another player in this game. And that other player or other players can take on many different characteristics and many different behaviors. So every different player, every different strategy that you're up against changes fundamentally the game and what the agent is trying to do. I go back and think, why are people still obsessed with playing chess? Why does a professional chess player still keep playing chess? It's the same game, but it's actually not, because you're playing completely different opponents day after day, and new people come into the world. So the game is continually changing. So multiplayer games and multi-agent games really encapsulate that huge diversity of tasks that you might encounter just from other players being there. So capture the flag was actually one of our first forays into how we can use multiplayer games to really stretch what our reinforcement learning algorithms can do, to really force us to think strongly about how we can generalize to new tasks, how we deal with these multi-agent dynamics. Capture the flag was a fantastic breakthrough. It really showed that we could get to human-level performance for these multiplayer first-person games. And then of course StarCraft added a huge amount of complexity and was sort of the next frontier that we had to go after for this.

你在这个领域起步非常早，这些概念中有很多在今天语言领域仍然非常相关。看到这些工作继续发挥作用，你有什么感受？

You were so early in this, so many of these concepts are very relevant today in the world of language. How does it feel to see some of this work continue to be played out?

是的，这太棒了。实际上非常精彩。你知道，我们在七年前，也就是 2015、16、17、18 年讨论的许多东西，如今看到所有这些核心基础概念在大语言模型领域真正有用且适用，并带来了当时我们只能梦想的性能，这真的非常令人满足。

Yeah, it's brilliant. It's just fantastic actually. You know, there were so many things that we were talking about seven years ago, you know, 2015, 16, 17, 18. And to see all of these core fundamental concepts being really useful and really applicable today in the world of large language models, and resulting in performance that we could only really dream about at the time, that's incredibly satisfying actually.

那么，用你自己的话说，你从搭建玩具转向了寻找真正的应用。你是什么时候知道自己找到了正确的配方？

So then, in your own words, you said that you moved from building toys to then finding real applications. When did you know that you found the right recipe?

你知道，我就是热爱深度学习。我已经痴迷深度学习十到十五年了。我热爱它的原因是，它有这些底层的核心概念，这些基本的构建块，它们在不同应用领域之间具有令人难以置信的可迁移性。所以，我们在 2012 年计算机视觉中使用的构建块，与后来语言生成模型早期、强化学习等中使用的构建块是一样的。我一次又一次看到的是，我们可以利用这些核心概念，同样的核心概念，召集那些了不起的人，他们就像把这些概念和不同构建块组合在一起的大厨，组建一支优秀的团队，去攻克真正具有挑战性的问题——那些你在会议上与领域顶尖研究者交流时，他们会说“不，不，不，这还需要十年”，而你心里知道，我们其实基本上已经解决了。我一次又一次看到这种情况发生。你把优秀的人才、优秀的算法、强大的算力投入到真正困难的问题上，我们现在可以找到破解许多问题的配方。所以，到了某个节点，我一直非常关注这些方法的应用。我希望看到这项技术对世界产生真正的、变革性的积极影响。因此，我们需要开始真正追求这一点，而我认为时机已经成熟好几年了。

So, you know, I just love deep learning. I've been obsessed with deep learning for, you know, 10, 15 years now. And the thing that I love about it is that you have these underlying core concepts, these fundamental building blocks that are somehow incredibly transferable between different application spaces. So you know, it's the same building blocks that we were using in computer vision in 2012 as we were using in generative models early on in language, then reinforcement learning, etc. So what I was seeing again and again was this ability to take these core concepts, the same core concepts, take incredible people who understand how to, they're almost like master chefs of putting these concepts together and these different building blocks together, take a team of incredible people and go after really challenging problems, problems that you go to conferences at the time and you talk to leading researchers in the field, they say no no no, this is 10 years away, and in the back of your mind you know okay we actually basically cracked it. And I saw that happen again and again. You take amazing people, amazing algorithms, amazing compute on really challenging problems, and we can find recipes now to crack so many problems. And so it just got to the point where, and I've always been quite obsessed with the application of these methods. I want to see this technology have real transformative positive impacts in the world. And so we need to start actually going after that, and the time has been right for, I think, a few years now.

那么，你现在已经与这个时代最伟大的科学家、技术专家和创始人之一 Demis 合作了十年之久。他是在你还在牛津时给你打电话的。然后你的公司 Vision Factory 和 DeepMind 都在 2014 年左右被谷歌收购，差不多同一时间，你们俩开始合作。到现在已经超过十年了。与 Demis 共事是什么感觉？

Well, so you've now had a decade-long relationship working together with one of the greatest scientists, technologists, and founders of our lifetime, Demis. He called you while you were still at Oxford. And then your company, Vision Factory, and DeepMind were both acquired by Google back in 2014, around the same time, and that's when the two of you started to work together. Now for over 10 years. What was it like or what has it been like to work with Demis?

是的，Demis 是一个了不起的人。他很有个性，也是一个真正的远见者。同时他也非常有人情味，平易近人，我认为这很能激励人。所以，只需五分钟的交谈，他就能让你感受到他思考的雄心壮志的深度，以及迈向这些雄心的潜力的紧迫性。因此，我认为他有一种伟大的能力，能够为一群非常聪明的人注入大量能量，让人们看到眼前之外的东西。我记得有一次，我站在早期 DeepMind 办公室的大厅里，那是一次庆祝 DeepMind 第一篇《自然》论文的祝酒会。Demis 说，这实际上只是几十篇《自然》论文中的第一篇。

Yeah, I mean Demis is an incredible person. A real character and a real visionary. And also amazingly human and relatable, and I think that really inspires people. So you know, it only takes a five-minute conversation for him to sort of really bleed out the depth of ambition that he thinks about and just the immediacy of the potential to step towards these ambitions. So I think he has this great ability to inject a lot of energy into a group of very smart people, get people to see beyond what's right in front of them. I remember moments sitting, well standing in the lobby of one of the early DeepMind offices. I think this was a toast, a celebration we were having for the first Nature paper from DeepMind. And Demis was saying, you know, this is actually just going to be the first of dozens of Nature papers.

当时，这是第一篇发表在《自然》杂志上的机器学习论文，就是那篇 Atari DQN 论文。而展望未来会有几十篇《自然》论文，听起来有点不切实际。实际上，他更进一步说：‘我们会因此赢得诺贝尔奖。’那是十年前的事了。太不可思议了。他那种远见——我称之为‘推演思维’。也许来自他下棋的经验，但他总是在推演未来：现在需要哪些步骤才能实现那个宏大目标？所以这十年非常精彩。我和他共事大约十年了，现在在 Isomorphic Labs 仍然紧密合作。雄心一如既往。

And at the time, this was the first machine learning paper in Nature. This was the Atari DQN paper. And the prospect of dozens of Nature papers seemed a bit farfetched. And actually, he went further and said, 'And we're going to be winning Nobel prizes as a result of this.' And that was 10 years ago. That's incredible. The forethought that he has—he's got what I call one of these rollout minds. Maybe it comes from all his experience playing chess, but he's always rolling out into the future. What are the steps now that are going to lead to this big ambition? So it's been fantastic. I've been working with him for about 10 years now. We still work really closely together on Isomorphic Labs. And the ambition is as big as ever.

听到你和他在一开始就有这样的雄心，真是太有趣了。而且它真的实现了，令人难以置信。我想聊聊 Isomorphic。你现在正投身于我们这一代最雄心勃勃的使命之一：用 AI 重新构想药物发现和药物开发。如果一切顺利，你实现了 Isomorphic 的愿景，世界会是什么样子？

It's so interesting to hear that you had this ambition and that he had this ambition from the very start. And it's incredible that it's played out that way. Well, I'd love to talk a little bit about Isomorphic. You're now embarking on one of the most ambitious missions of our generation: to reimagine drug discovery and drug development with AI. If everything goes right and you realize your vision for Isomorphic, what does the world look like?

是的，我们在 Isomorphic 想得非常大。我们想在这里解决所有疾病，真正达到那个规模。关键在于，我们正在构建的技术——以及整个 AI 领域——将彻底改变我们理解生物学的方式，以及我们操控和设计化学来调节生物学的能力。所以我们设想一个未来，我们解决所有疾病，AI 不仅帮助我们发现、创造和设计新疗法，还能让我们更深入地理解生物世界——细胞如何工作，疾病的根本原因是什么——从而开辟我们可以调节的新途径。我们从公司成立第一天起就致力于实现这个宏大目标。这不是为了某个特定适应症或靶点开发疗法，而是思考如何用 AI 创建一个非常通用的药物设计引擎，不仅能应用于单个靶点甚至单一模态，还能反复应用于任何不同的疾病领域。这就是我们目前正在推进的方向。

Yeah, we think really big at Isomorphic. We want to be solving all diseases here, genuinely at that scale. The point is that this technology we're building—and AI as a whole field—is going to be completely transformative in how we understand biology, in our ability to manipulate and craft chemistry to modulate that biology. So we really think about a future where we are solving all diseases, where AI is not just helping us discover, create, and design new therapeutics, but also understand so much more about our biological world—about how our cells are working, what the root causes of disease are—and therefore opening up new pathways that we can think about modulating. We set up the company from day one to really go after this big ambition. This isn't about developing therapeutics for a particular indication or a particular target. It's really thinking about how do we create a very general drug design engine with AI, something that we can apply not just to a single target or even a single modality, but we can apply this again and again across any different disease area. And that's what we're stepping towards at the moment.

从一开始就带着这种通用性的雄心，如何改变了你实际构建的方式？

How does setting out with this ambition of being general change how you built in practice from day one?

好问题。当我思考 AI 在药物设计中的现状时，化学和生物学中已经有很多机器学习模型的应用，但我认为这些应用的第一代大多是局部模型。你可能有一些关于特定靶点或某类分子行为的数据，然后你会拟合一个小型多层感知机来生成预测，指导下一轮设计。这与我们试图做的完全相反。所以从第一天起，我们就着手创建能够跨化学和靶点空间泛化的模型。一个关键例子就是 AlphaFold 和 AlphaFold 3。这是一个可以应用于大量不同靶点的模型。你可以将其应用于蛋白质组中的任何蛋白质，甚至整个蛋白质宇宙。你可以将其应用于你能想到的任何小分子设计，而无需微调或拟合任何局部数据。所以你可以想象，如果化学家不需要针对每个应用调整模型，这完全改变了他们使用这些模型的方式。我们内部的每一个研究项目——顺便说一句，当我思考我们需要什么来构建这个突破性的药物设计引擎时，我们需要大约六个 AlphaFold 级别的突破。AlphaFold 只是故事的一部分。所以从第一天起，我们就设立了这些内部研究项目，去攻克这六个问题。我们在 AlphaFold 和结构预测方面取得了重大突破，也在其他关键领域取得了进展。所有这些模型都是通用的；它们可以应用于任何靶点，至少可以应用于许多不同的模态。

It's a good question. When I think about some of the status quo of AI in drug design, there's been a lot of use of machine learning models in chemistry and biology, but I would call a lot of the first generation of these applications to be more local models. You might have some data about a particular target or about how a particular class of molecules is behaving, and you'll fit a small multi-layer MLP against this data to help you generate some predictions that lead to your next round of design. This is the complete opposite approach of what we were trying to do. So from day one, we were setting out to create models that generalize across chemistry and across target space. A key example of this is something like AlphaFold and AlphaFold 3. This is a model that you can apply to a whole different host of targets. You can apply it to any protein in the proteome, in the universe of proteins. You can apply it to any small molecule that you can think of designing without needing to fine-tune it, without needing to fit any local data. So you can imagine that completely changes the way that chemists can use these models if you don't need to be adapting this model to every single application. Every single one of our internal research projects—and by the way, when I think about what we're going to need to get this breakthrough drug design engine that we've been building, we need like half a dozen AlphaFolds. AlphaFold is just part of the story. So from day one, we've been setting up these internal research programs going after these half a dozen problems. We've had significant breakthroughs obviously in AlphaFold and structure prediction, but also in other key areas. And in all of these, these models are general; they can be applied to any target. They can be applied to any modality or lots of different modalities at least.

这是我第一次听你说‘六个 AlphaFold’。能再多解释一下吗？

That's the first time I've heard you say 'half a dozen AlphaFolds.' Can you share a little bit more about what that means?

AlphaFold 显然是在理解生物分子结构方面的重大突破——蛋白质的结构是什么，现在有了 AlphaFold 3，还包括蛋白质与小分子以及 DNA、RNA 等的结构。这是一个根本性的阶跃变化。它让我们能够以实验级别的精度理解生物化学的一个核心概念。这为化学家开启了一整套思考和设计工作。但我想说的是，我们可能还需要大约六个这样的突破——达到生物学和化学不同核心概念的实验级精度——才能将它们整合成真正变革药物设计的东西。药物设计非常困难。它不是一个单一问题。不仅仅是理解蛋白质的结构，甚至不仅仅是设计一个能以你想要的方式调节该蛋白质的分子。你希望这个分子能作为药片服用，通过身体，以正确的方式被吸收，到达正确的细胞类型，进入细胞，并且不被肝脏以某种方式分解。所以药物设计师需要处理非常多的复杂性。而每一个这样的问题都像 AlphaFold 级别的突破，我们正在创造它们。

AlphaFold was obviously a massive breakthrough in understanding biomolecular structure—what is the structure of proteins, and now with AlphaFold 3, structure of proteins with small molecules and things like DNA and RNA. That's a fundamental step change. It allows us to get experimental level accuracy of a really core concept of biochemistry. That unlocks a whole bunch of thinking and design work for chemists. But my comment here is actually we're probably going to need something like half a dozen more of these sorts of breakthroughs—getting to experimental level accuracy of different core concepts of biology and chemistry—to be able to put this together into something that's really transformative for drug design. Drug design is really hard. It's not just a single problem. It's not just about understanding the structure of a protein. It's not even just about designing a molecule that will modulate that protein in the way you want. You want this molecule to be able to be taken as a pill, go through the body, be absorbed in the right way, reach the right cell type, go into the cell, and not be broken down by the liver in a certain way. So there's just so much complexity to hold on to as a drug designer. And each one of those is like an AlphaFold-level style breakthrough that we've been creating.

我还听你提到‘药物设计的圣杯模型’和‘科学智能体’。能再解释一下吗？

I've also heard you use the words 'a holy grail model for drug design' and 'agents for science.' Can you explain a little bit more about what you mean?

我们一直在攻克的一些研究领域——预测这些分子的结构和性质，以及所有这些生物分子如何相互作用并随时间演变——这些确实是药物设计的圣杯级预测问题。

Some of these research areas that we've been going after—predicting structure and properties of these molecules and how all of these biomolecules interact and play out over time—these really are sort of holy grail predictive problems for drug design.

我们在那里取得了一些令人难以置信的突破，这些突破确实让我们的化学家感到震惊，并彻底改变了我们在 Isomorphic 内部进行药物设计的方式。但我认为一个非常有趣的想法是，你可以创建世界上最好的预测模型，达到实验水平甚至超过实验水平，来预测分子的某个特定属性，例如预测真实实验的结果。我们可以拥有一整套这样的模型，但这仍然不能解决药物设计问题。思考这个问题的方式是，有一个数字 10^60，这大概是所有可能存在的类药分子的数量。这考虑了很多因素。所以我们甚至可以将这个数字减少 20 个数量级，降到 10^40。那仍然是一个很大的数字。即使你拥有世界上最好的预测模型，假设你可以筛选十亿个不同的分子，那是 10^9。所以现在我们仍然有 10^31 个分子没有被探索。因此，即使拥有最好的预测模型，你仍然没有触及到你应该探索的分子空间的表面。这就是为什么我们需要超越仅仅预测实验的模型，还需要像生成模型这样的模型，像智能体那样能够真正导航整个 10^40 到 10^60 的空间。

And we've made some incredible breakthroughs there which have really stunned our chemists and step-changed how we do drug design internally at Isomorphic. But what I think is a really interesting thing to think about is that you could create the best possible predictive model of the world, at an experimental level or even better than experimental level, to predict a particular property about a molecule, for example, to predict the outcome of a real experiment. So we could have a whole suite of those, but that still wouldn't solve drug design. And the way to think about this is that there's this number 10^60, which is perhaps all of the possible drug-like molecules that could exist. That takes into account a lot of things. So we could even reduce that by 20 orders of magnitude to 10^40. That's still a lot of things. And even if you had the best predictive models in the world, let's say you could screen a billion different molecules, that's 10^9. So now we're still like 10^31 molecules left on the table. So even with the best predictive models, you're still not even scratching the surface of molecular space that you should be exploring. And this is why we need to go beyond just predictive models of experiment but also models like generative models, like agents that can actually navigate that whole 10^40 to 10^60 space.

这太有趣了。显然，使用我们的预测模型来理解如何导航，这样我们就不必穷举搜索，因为我们永远无法穷举搜索整个分子宇宙，如果这说得通的话。就像 AlphaGo 无法穷举搜索所有可能的围棋走法一样，对吧？不像国际象棋，你可以穷举搜索所有可能的走法。

That's so interesting. Using our predictive models obviously to understand how to navigate that, so we don't have to exhaustively search because we can never exhaustively search the whole universe of molecules, if that makes sense. Just in the same way that AlphaGo couldn't exhaustively search all of the possible Go moves, right? Unlike chess where you could exhaustively search all possible chess moves.

是的。但分子设计更像围棋而不是国际象棋。所以这就是生成模型发挥作用的地方，智能体利用生成模型，利用搜索技术以及这些惊人的预测能力，真正打开整个分子空间。对我来说，即使没有人工智能，我们也能在这个 10^60 到 10^40 的空间中找到药物，这仍然令人惊讶。这只说明可能有很多冗余，很多潜在的设计。如果你考虑一个特定的疾病适应症，一个特定的靶点，应该有很多设计适合它，并且符合这种治疗药物的产品特性。我认为真正的潜力在于这些生成模型，这些智能体，能够搜索这个空间，真正揭示整个潜在的设计空间。

Yeah. But molecule design is much more like Go than it is like chess. So that's where generative models come into play, agents that utilize generative models, that utilize search techniques as well as these amazing predictive capabilities to really open up the entirety of molecular space. Now to me, it's actually still amazing that even without AI we managed to find drugs in this 10^60 to 10^40 space. It just says there's probably a lot of redundancy, a lot of potential designs. If you think about a particular disease indication, a particular target, there should be many designs that would be good for that and would be the right sort of product profile for this therapeutic. And I think the real potential here is for these generative models, these agents as well, to be able to search through this space and really uncover that whole potential design space.

这太有趣了。我认为用非常简单的通俗语言来说，你既在建模、学习和模拟游戏，又在试图构建最好的玩家来解决不同类型的游戏。

That's so interesting. I think in very simplistic layman terms, you're both modeling, learning and modeling the game and trying to build the best player to solve different types of games.

所以，我的意思是，我对游戏有着难以置信的偏爱。我从小就开始玩电子游戏，在那个世界里长大。但这正是我对它的看法。我们必须创建我们的世界模型，我们的生化世界模型，我们的生物世界模型。然后我们不止步于此。我们实际上还需要创建智能体和生成模型，它们能够找出如何探索，如何穿越那个空间，并基本上揭示化学空间中这些令人惊叹的沧海遗珠，这些可能成为数百万人的改变生命的疗法。

So, I mean, I'm incredibly biased by games. I've been playing video games since I was a kid, grew up in that world. But that's exactly how I think about it. We've got to be creating our world models, our models of the biochemical world, our biological world. And then we don't stop there. We actually then need to be creating agents and generative models that can work out how to explore, how to traverse that and to basically uncover these amazing needles in the haystack in chemical space which could be life-changing therapeutics for so many millions of people.

我喜欢这个。这就是我们今天的点睛之笔。那么，AlphaFold 3 确实是开创性的。你让我们从只能建模蛋白质的结构，到现在能够建模所有分子及其相互之间的结构。你能分享一下我们应该如何看待它在准确性、速度、效率方面的影响，以及可能探索以前无法解决的问题空间的能力吗？

I love that. That is our punchline today. So, AlphaFold 3 is truly groundbreaking. You've taken us from being able to model just the structure of a protein to now being able to model the structure of all molecules and their interactions with each other. Can you share a little bit about how we should think about that in terms of the impact in accuracy, speed, efficiency, and also potentially in being able to explore problem spaces that we couldn't solve before this?

是的。所以 AlphaFold 2 是最大的突破，能够理解蛋白质的结构。然后有一个叫做 AlphaFold Multimer 的东西，它不仅能让你理解蛋白质本身的结构，每个单独的蛋白质，还能理解蛋白质在所谓的复合物中结合在一起时的结构。所以这些蛋白质如何结合在一起，这开启并帮助我们回答生物学中的许多问题。但距离设计治疗药物还有很大一步。而一大类治疗药物就是所谓的小分子。这些分子不是蛋白质。它们会是像咖啡因或扑热息痛这样的东西，通常可以做成药片服用。这些治疗药物，这些小分子的作用方式是，它们通过身体，进入细胞，然后实际上附着到这些蛋白质上。这些蛋白质是生命的基本组成部分。它们通过与其他蛋白质相互作用形成这些分子机器。所以你可以想象，如果你有另一个分子，你的药物，它进来并附着在这里的一个蛋白质上，那么它可能会破坏该蛋白质与另一个蛋白质相互作用的能力，而这是它日常正常机器的一部分。所以你在用这个小分子调节该蛋白质的功能。这就是药物设计以及治疗药物如何工作的本质。所以你可以想象，作为一名化学家或药物设计师，你试图设计一个小分子，它将适合这里的这个蛋白质，并破坏它的正常功能，或者在某些情况下增强它的正常功能。因此，了解这个小分子如何与蛋白质相互作用将非常有帮助。它可能形成什么样的结构？有什么相互作用？这些是字面意义上的物理相互作用。所以这真正激发了 AlphaFold 3 的创建，现在我们有了一个模型，它不仅预测蛋白质的结构，还预测这些蛋白质如何与小分子相互作用，以及其他基本的分子机器构建块，如 DNA 和 RNA。这基本上开启了结构理解的能力，这是药物设计的核心部分。小分子，它开启了新的靶点类别，你知道，有些东西像转录因子，它们是坐在 DNA 上并读取 DNA 的蛋白质。

Yeah. So AlphaFold 2 was the biggest breakthrough, right, to be able to understand the structure of proteins. And then there was something called AlphaFold Multimer which then allows you to understand not just the structure of proteins by themselves, each individual protein, but the structure of proteins as they come together in what we call complexes. So how these proteins fit together opens up and helps us answer a lot of questions in biology. But there's still a big hop to designing therapeutics. And one of the big classes of therapeutics is what's called small molecules. So these are molecules that are not proteins. These would be things like caffeine or paracetamol, things that more often you can take as a pill. And the way that these therapeutics work, these small molecules, is that they go through the body, they go into the cell and they actually come and attach themselves to these proteins. These proteins are the fundamental building blocks of life. They form these molecular machines by interacting with other proteins. And so you can imagine that if you have another molecule, your drug, that comes in and attaches itself to a protein over here, then it might disrupt the ability for that protein to interact with another protein, part of its normal machine in day-to-day life. And so you're modulating the function of that protein with this small molecule. And that's the essence of drug design and how therapeutics work. And so you can imagine as a chemist or a drug designer, you're trying to design a small molecule that's going to fit to this protein over here and disrupt how it normally functions or in some cases enhance how it normally functions. And so it'd be really helpful to understand how this small molecule interacts with the protein. What's the structure that it might make? What are the interactions? These literally physical interactions that are being made. And so that really inspired the creation of AlphaFold 3 where now we have a model that not only predicts the structure of proteins, but how these proteins interact with small molecules, also other fundamental molecular machine building blocks, things like DNA and RNA. And this basically opens up the ability to structurally understand, which is a core part of drug design. Small molecules, it opens up new classes of targets, you know, there are things like transcription factors which are proteins that sit on DNA and read DNA.

你可以想象一下，现在要设计一个小分子来改变或破坏某个功能，你真的很想能亲眼在 3D 中看到这一切。如果我对我的小分子做了改动，那会如何改变它与这个蛋白质在生物分子系统中的相互作用？AlphaFold 3 现在非常非常准确，它让我们能够纯粹在计算机上回答很多这类问题。而以前你必须去实验室，真正地结晶这些东西，这可能需要六个月，甚至几年，有时甚至根本不可能。现在在 Isomorphic，我们的药物设计师就坐在笔记本电脑前，通过浏览器界面就能理解、修改他们的设计，并看到改动的影响。太不可思议了。

And you can imagine now trying to design a small molecule to change or disrupt the function of something like that. To do that, you'd really want to be able to see literally in 3D how this all looks. And if I make changes to my little molecule, how will that change the way it interacts with this protein in this biomolecular system? So AlphaFold 3 is now very, very accurate. It allows us to answer a lot of these questions purely in silico, purely on a computer, where before you would have to go to the lab, literally crystallize this stuff. This can take six months, it can take years, sometimes it's even impossible. Now at Isomorphic, our drug designers are literally sitting with their laptop, browser-based interface, able to understand, make changes to their designs, and see the impact of that. Incredible.

AlphaFold 3 关注几种相互作用：蛋白质与核酸、蛋白质与配体、以及抗体与抗原。你能给我们举一些很好的例子，说明 AlphaFold 3 对这些不同类型蛋白质和分子相互作用的影响吗？

So there are a couple of interactions that AlphaFold 3 is focused on: proteins and nucleic acids, proteins and ligands, and antibody to antigen. Can you give us some good examples of the impact that AlphaFold 3 now has on the interaction of these different types of proteins and molecules?

是的。蛋白质与配体，也就是蛋白质与小分子，这两个术语是同义词。这让我们能够理解小分子药物是如何相互作用的。然后我们可以考虑蛋白质-蛋白质相互作用。有一整类疗法叫做生物制剂，比如抗体。这让我们能够理解它们如何与我们的靶点相互作用，开辟了新的模式。这也涵盖了抗体-抗原界面。所以如果你在设计抗体，你想了解你的抗体设计将如何与蛋白质表面相互作用。所以这是同一个模型，可以用于所有这些不同的应用。

Yeah. So protein and ligands, that's the same as protein and small molecules. Those two terms, ligands and small molecules, are synonymous. That allows us to understand how small molecule drugs interact. Then we can think about protein-protein interactions. There's a whole class of therapeutics called biologics, these are things like antibodies. That allows us to understand how they might interact with our targets. Opens up new modalities. And that also encapsulates the antibody-antigen interface. So if you're designing an antibody, you want to understand how your antibody design is going to interact with the protein surface. So it's the same model that we can use across all of these different applications.

训练像 AlphaFold 3 这样的模型有什么细微之处？使用基于扩散的架构有什么好处？

What are the nuances of training a model like AlphaFold 3 and what are the benefits of using a diffusion-based architecture?

这是个好问题。要让 AlphaFold 3 工作，我们克服了很多挑战。最有趣的事情之一就是如何把只处理蛋白质的 AlphaFold，扩展到输入这些新的模态，比如 RNA、DNA、小分子这些新数据类型。所以我们不仅要解决如何对蛋白质进行分词（我们大致知道怎么做），还要解决如何对 DNA 和小分子进行分词。对于 DNA 和 RNA，这比较明显，我们可以按碱基分词。但对于小分子，我们尝试了很多不同的方法，最终发现这种原子级分辨率的分词效果非常好。然后问题来了：如何实际预测这种不同分子类型混合物的结构？你不能用 AlphaFold 2 的框架。这就是扩散模型真正大放异彩的地方。我们可以对每一个单独的原子及其坐标进行建模，让扩散模型生成这些 3D 坐标。而我们刚才提到的分词，就是为扩散过程的推理提供条件。这非常有趣，是一个巨大的突破。所以，在我们的排行榜上，这是一个巨大的阶跃变化，特别是在小分子-蛋白质相互作用准确性方面。这是一个巨大的阶跃变化，真正为项目的其余部分扫清了障碍。

Yeah, it's a great question. There are a lot of challenges we had to overcome to get AlphaFold 3 to work. One of the most interesting things was actually just how do we take something like AlphaFold, which was only working with proteins, and then input these new modalities, these new data types of RNA, DNA, small molecules. So we had to work out how to tokenize not just proteins, which we kind of knew how to do, but how to tokenize DNA, how to tokenize small molecules. For things like DNA and RNA, that's a little bit more obvious, we could tokenize in the bases. But then for small molecules, we tried a whole bunch of different stuff, and it really ended up that this atomic resolution tokenization worked super well. And then you have the question of okay, how do you actually predict the structure of this mixture of different molecule types? You couldn't use the same framework as AlphaFold 2. And this is where diffusion modeling really shone. Here we could actually model every single individual atom and the coordinates of every atom individually, and have a diffusion model be producing those 3D coordinates. And the tokenization that we talked about is conditioning the inference of that diffusion process. So interesting, and this was a huge breakthrough. So, you know, we're talking about on our leaderboard just a massive step change, particularly in small molecule-protein interaction accuracy. It was a massive step change and something that really unblocked the rest of the project.

数据、算力和算法，我们知道这三者在所有其他相邻领域都很重要。但我很惊讶地读到一篇对 Demis 的采访，他说我们在生物学中并不受数据限制。你能分享你的观点吗？

So, data, compute, and algorithms, we know those three are important in all other adjacent fields. But I was surprised to read an interview with Demis where he shared that we're not data constrained in biology. Can you share your point of view on that?

我认为，无论你身处哪个机器学习领域，你都会感受到一些数据限制。Demis 的观点是，这并非真正的瓶颈，因为我们利用现有的数据和我们能生成的数据就能取得进展，而且确实能取得实质性进展。并不是说我们得坐等 50 年，等全世界生成数据才能产生影响。不，我们完全不是这样。有些建模领域的数据已经存在多年，我们可以利用它们取得远超以往的巨大进步。但这并不意味着生物学数据没有机会？绝对不是。数据将是我们继续开发这些模型和系统的基础部分，关键在于我们去生成什么样的数据。在我看来，这里有一个巨大的机会。用于机器学习的生物学数据实际上还没有被创造出来。是的，有很多历史数据，但这些历史数据并不是为了机器学习的目的而创建的。所以当你思考如何创建数据来训练模型时，你的思考方式与过去人们生成数据的方式截然不同。这其中有很大的探索机会。

You know, I think it doesn't matter what field of machine learning you're in, you're going to feel some data constraints. And I think the point here from Demis is that it's not a real bottleneck, as in we can make progress with the data that is out there, the data we can generate, and real progress can be made. It's not like, oh, we've got to sit and wait 50 years for the world to generate data before we can actually make impact here. No, we're not seeing that at all. There are modeling spaces where the data has been sitting around for years that we can see that we can make really substantial progress beyond anything that people have experienced before. Now, does that mean there's no opportunity for data in biology? Absolutely not. Like, it's going to be a fundamental part of how we continue to develop these models and these systems will be what data we go out and generate. And there I think there's just a massive opportunity in my mind. The data for machine learning in biology hasn't actually been created yet. Yes, there's a lot of historical data, but that historical data hasn't been created for the purposes of machine learning. And so when you're going out and thinking how do I create data to actually train my model, you're thinking in a very different way to how people have gone out and generated data in the past. And that there's a big opportunity there to explore.

你认为我们现在缺少什么样的数据？你认为我们需要合成数据吗？

What kind of data do you think we're missing here right now? And do you think we need anything in synthetic data?

是的。我是合成数据的超级粉丝。实际上，从我职业生涯一开始就是这样，当时我生成合成文本数据，只是为了克服一个事实：我作为博士生只能访问几千张图像，而谷歌有数百万张图像，所以我生成了大量的合成数据，这解决了问题。我们在化学领域也看到了同样的情况。我们有很好的理论。我们实际上对物理学了解很多。我们有量子化学和量子力学的理论，并可以据此创建模拟器。我们可以近似模拟，创建更具可扩展性的分子动力学模拟。这为大量合成数据提供了基础。然后我们还有模型本身，特别是生成模型，它们可以生成数据，我们可以使用评分系统来帮助真正增强这些数据的信息含量。但我认为一个大的开放领域将是所谓的体内数据。也就是通常在真实动物（如小鼠或大鼠）身上测量的数据。

Yes. So I'm a massive fan of synthetic data. Actually, I have been since the very beginning of my career where I was generating synthetic text data just to overcome the fact that I was a PhD student with access to a couple of thousand images and Google had millions and millions of images, and so instead I just generated tons and tons of synthetic data and that unblocked things. And we're seeing the same thing in the chemistry space especially. Where we have good theory. We actually know a lot about physics. We have the theory of quantum chemistry and quantum mechanics and we can create simulators out of that. We can approximate that and create more scalable molecular dynamic simulations. This gives the basis for a whole host of synthetic data. Then we have the models themselves, especially generative models, that can actually generate data that we can use scoring systems to help really enhance the information content of this data. But I think one of the big open spaces will be on what's called in vivo data. So data that you would normally measure on a real animal, something like a mouse or a rat.

我们知道这方面有一些历史数据，但你无法生成大量数据——实际上你根本生成不了。所以，寻找新的数据生成技术是一个巨大的机会。有一些了不起的人在做类器官芯片之类的事情——用芯片来测量你通常在真实动物身上测量的东西。我认为生物学和化学领域将会涌现出一大批数据生成技术的新突破。这也会对我们如何建模那个世界产生重大影响。

We know there's some historical data on that, but you can't generate tons of it—you can't really generate any at all. So there's a big opportunity to look to new data-generating technologies. There are some incredible people doing things like organoids on a chip—ways of starting to measure things that you would normally measure on a real animal, but completely on a chip. I think there's going to be a whole host of new breakthroughs in data-generating technology in biology and chemistry. That's going to have a big impact on how we think about modeling that world as well.

你们内部在做这方面的工作吗？还是希望其他参与者来填补这个空白？

Are you working on any of that internally, or are you hoping that other players fill in some of that gap?

在内部，我们在 Isomorphic Labs 实际上没有任何自己的实验室。但我们与许多其他公司合作。我们自己生成大量数据——大量专有数据。我们已经看到了惊人的影响。这非常合理。

Internally, we actually don't have any of our own labs in Isomorphic Labs. But we work with a whole bunch of other companies. We generate a lot of data ourselves—a lot of proprietary data. We've seen amazing impact of that. It makes a lot of sense.

所以有一种观点认为，建模分子的结构、功能以及功能的调节非常重要，但不一定是药物开发中始终的制约因素。你对此怎么看？

So there's a point of view that modeling structure of molecules and modeling their function and the modulation of function is very important, but not necessarily always the limiting factor in drug development. What's your point of view on that?

正如我之前提到的，药物设计非常非常复杂。在你进入药物开发阶段——也就是把这些设计真正用到人身上、进行临床试验之前——整个设计和开发领域存在很多瓶颈。药物开发是关于我们如何开始进行临床试验，如何测试这些药物在人体中的效果，如何既及时又安全地做到这一点。整个行业需要找出如何在这个领域进行创新，特别是随着我们预测这些分子如何与人体相互作用、毒性如何的模型越来越好，我们必须改变进行临床试验的方式，才能真正利用这些模型，最终将治疗方法交到急需它们的患者手中。即使在分子设计本身，也不仅仅是理解这些分子的结构，甚至不仅仅是理解这些分子如何改变这些蛋白质的功能。我们需要理解这些分子如何改变我们体内几乎每一种蛋白质的功能。因为如果我们把它做成药丸，它会遍布全身。这是毒性的主要原因：你设计了一种神奇的分子，完美地调节了你的特定靶点——这个靶点对你的疾病至关重要——但它也会影响其他东西。当然，你会做很多筛选来防止这种情况，但我们能预测得越多越好。从我的角度来看，真正令人兴奋的是，如果我们创建这些通用模型，理解这个分子如何与这个靶点相互作用，同时也理解与其他任何靶点的相互作用，那么为什么我们不能用同一个模型来理解这些分子如何与我们身体的其他部分相互作用呢？

As I touched upon before, drug design is really, really complex. And before you even get to drug development—where you take those designs and start putting them into real people, clinical trials—there are so many bottlenecks throughout this whole design and development space. Drug development is about how we start to approach clinical trials, how we should test these drugs out in people, how we can do this in a timely manner but still safe. There are a lot of bottlenecks there that the industry as a whole will need to work out how to innovate in that space, especially as our predictive models of how these molecules will interact with people, how toxic they will be, get better and better. We will have to change the way we approach clinical trials to really make use of that, ultimately to get therapeutics into the hands of patients who really desperately need them. Even in the design of molecules themselves, it's not just understanding the structure of these molecules, or even just understanding how these molecules change the function of these proteins. We need to understand how these molecules change the function of pretty much every single protein in our body. Because if we take this as a pill, it's going to go everywhere. That's the major cause of toxicity: you've designed this amazing molecule that perfectly modulates your specific target that is key to your disease, but it also affects other things. Of course, you do a lot of screening to protect against that, but the more we can predict that, the better. What's really exciting from my perspective is if we're creating these general models that understand how this molecule interacts with this target but also any other target, then why can't we just use that same model to understand how these molecules interact with the rest of our body?

那么 AlphaFold 3 现在为药物设计师带来了哪些可能性？你们内部是如何使用它的？

So what is now possible with AlphaFold 3 for drug designers? How are you using it internally?

AlphaFold 3 让我们的药物设计师能够理解他们的分子设计如何真正与这个蛋白质靶点——疾病的靶点——相互作用。我们的药物设计师可以对设计进行修改，然后立即看到这如何改变分子与蛋白质靶点的物理相互作用方式。这非常强大。在 AlphaFold 3 之前，你对此完全盲目。你实际上可能不知道你的分子如何与你的蛋白质相互作用。你只能依靠最好的直觉。也许在药物设计项目的某个阶段，你会将你的结构结晶出来，得到一个特定的设计。这意味着去一个真正的实验室，六个月后，如果幸运的话，得到一个解析出的 3D 结构，但即便如此，那也只是单个设计的 3D 结构，而不是你做的每一个修改。所以 AlphaFold 3 彻底改变了化学家进行设计工作的方式。但我必须强调，这离我们想要达到的目标还差得很远，因为这不仅仅是关于这些分子在相互作用时看起来像什么。我们实际上想知道这些分子与这个蛋白质结合的强度。我们想知道这些分子的其他属性。我们想理解这些分子与蛋白质相互作用的方式如何改变蛋白质的折叠或构象，如何改变蛋白质的功能，如何可能改变细胞的动态。有很多问题，这些都是我们正在研究的其他类似 AlphaFold 的突破。我们已经为此创建了令人难以置信的模型，我们的化学家正在设计过程中使用它们。

AlphaFold 3 gives our drug designers the ability to understand how their molecule designs really interact with this protein target—the target of disease. Our drug designers can make changes to the design and then see instantly how that changes the way this molecule physically interacts with the protein target. That's really powerful. Before AlphaFold 3, you would be completely blind to this. You wouldn't actually probably know how your molecule is interacting with your protein. You'd be using your best intuition. Maybe somewhere down the line in the drug design project, you would get your structure crystallized with a particular design. That means going out to a real lab six months later, if you're lucky, getting a resolved 3D structure, but even then, that's just the 3D structure of a single design, not every single change that you make. So AlphaFold 3 completely changes the way chemists can do this design work. But I would stress that that's nowhere near as far as we want to go, because it's not just about what these molecules look like in terms of interacting. We actually want to know how strongly these molecules interact with this protein. We want to know other properties of these molecules. We want to understand how the way these molecules interact with this protein changes the fold or conformation of the protein, how that changes the function of the protein, how it might actually change the dynamics of the cell. There are so many questions, and these are other AlphaFold-like breakthroughs that we're working on. We have created incredible models for that that our chemists are using in this design process.

有意思。那么你们内部在设计一些药物。你们专注于哪些靶点和项目？

Interesting. So you're designing some drugs internally. What targets and programs are you focused on?

我们有一个非常令人兴奋的内部药物设计项目。这些项目专注于免疫学和肿瘤学。我们取得了令人难以置信的进展，看到这些模型如何改变了我们在这些项目上进行药物设计的方式，真的非常令人兴奋。

We have a really exciting internal program of drug design projects. These are focused on immunology and oncology. We've been making some incredible progress there, and it's been really exciting to see especially how these models have transformed the way we're actually approaching drug design on these programs.

你们还与礼来和诺华合作，最近你们宣布扩大了与诺华的合作。你能分享一下这些合作是什么样的吗？

You're also working with Eli Lilly and Novartis, and recently you announced an expansion with Novartis' partnership. Can you share a little bit about what these partnerships look like?

是的，我们签署了这些初步的合作——两个合作，一个与礼来，一个与诺华——这太棒了。他们给我们带来了一些非常具有挑战性的问题。我认为这不是秘密，例如诺华带给我们的靶点是该领域和诺华已经研究了十年以上的靶点。所以这些不是那种“哦，我们试试看”的问题。这些是真正的难题。去年是令人惊叹的一年，无论是我们的内部项目还是这些合作项目，都真正看到了这些模型的效果。它们让我们能够真正发现新的化学物质，找出调节这些人们研究了很久的靶点的新方法。

Yes, we signed these initial partnerships—two partnerships, one with Eli Lilly, one with Novartis—that was fantastic. They brought some really challenging problems to us. I think it's no secret that the targets that, for example, Novartis brought to us are targets that the field and Novartis, for example, have been working on for 10 years plus. So these aren't sort of 'oh, we'll try things out' problems. These are real hard things. Last year was an amazing year, both for our internal projects but also for these partner projects, to really see how well these models are working. It's allowed us to really uncover new chemical matter, working out new ways to modulate these targets that people have worked on for a long time.

看到这项新协议扩展了艺术家合作，真是令人惊叹，我认为这证明了这些合作早期取得的成功。恭喜。我认为这是一个了不起的里程碑，尤其才一年时间。所以我想聊聊团队。你组建了一支非常优秀的团队，汇聚了人工智能、化学、生物学等多个领域的顶尖人才。你还引入了外行来挑战传统思维。能分享一下你是怎么考虑的吗？

It's been amazing to see this new deal which has expanded on the artist collaboration, which I think is a real testament to some of the success of the early days of these partnerships. Congratulations. I think it's an incredible milestone, especially just one year in. So I'd love to talk a little bit about the team. You've built a truly excellent team composed of the highest caliber talent across many different fields: AI, chemistry, biology. And you've also brought in outsiders into the field to help question traditional thinking. Can you share a little bit about how you thought about this?

人工智能用于药物设计这个领域存在的时间并不长。所以找到一个既是药物设计世界级专家又是机器学习或深度学习世界级专家的概率基本为零。我真的认为 Isomorphic 正在孕育一个新的科学领域，因为我们有这些真正生活在交叉领域的人。但既然我们招不到这样的人，我就思考如何把药物设计和药物化学的世界级专家与机器学习和深度学习的世界级专家聚集在一起，让这些了不起的人并肩工作。仅仅让这些优秀的人待在各自孤立的团队里是不够的。我们需要人们并肩工作，说彼此的语言，带着大量的同理心和好奇心去理解这门新科学，真正用自己的语言建立直觉。我们已经看到这种动态产生了非常惊人的成果：一个对化学或生物学一无所知的通才机器学习者开始理解药物化学家和药物设计者的问题。当我考虑为我们正在进行的研究招聘机器学习者、机器学习科学家和工程师时，我们团队中 60%、70%、80%的人之前没有化学或生物学知识，可能只有高中或大学水平。这实际上可以成为一笔真正的财富，因为你带着一点天真进来。只要你有好奇心，关键之一就是提出好奇的问题，提出这些愚蠢的问题，这让我们能够从第一性原理出发解决问题。它几乎让我们突破了以往经验的教条以及人们传统上处理这些问题的方式。我们可以从头开始思考。这是我们思考如何创造这些研究突破的很多心态。有点天真、高度好奇和高度自主性是非常好的事情。

The space of AI for drug design hasn't really existed for very long. So the chances of finding a world expert at drug design who's also a world expert in machine learning or deep learning is basically zero. I genuinely think about a new field of science that Isomorphic is breeding because we have these people who really live and breathe the intersection of this. But because we can't hire these people, I really think about how do we bring the world experts at drug design and medicinal chemistry and the world experts at machine learning and deep learning and get these incredible people sitting side by side. It's not just enough to have these amazing people sitting in their isolated teams. We need people sitting side by side speaking each other's languages, with a lot of empathy, a lot of curiosity to understand this new science, to really build intuitions in your own language. We've seen such amazing things come out of this dynamic where you really have a generalist machine learner who doesn't know anything about chemistry or biology start to come in and understand the problems of a medicinal chemist and a drug designer. When I think about even hiring machine learners and machine learning scientists and engineers for the research we're doing, I'd say 60, 70, 80% of the people on our team have no prior knowledge of chemistry or biology, maybe high school or university level. And that can actually be a real asset because you come in a little bit naive. As long as you're curious, one of the key things is asking the curious questions, asking these stupid questions, and that allows us to come at the problems from first principles. It almost allows us to break through the dogma of previous experience and how people traditionally approach these problems. We can think ground up from scratch. That's a lot of the mentality of how we think about creating these research breakthroughs. A little naive and highly curious and high agency is a very good thing.

去年 11 月，你还做了一个重大举措，推出了 AlphaFold 服务器，向学术界发布代码和模型权重。能分享一下为什么吗？

In November last year, you also made a very big move in launching the AlphaFold server, which releases code and model weights for academic use. Can you share a little bit about why?

AlphaFold 长期以来一直对学术和科学用途开放。在 AlphaFold 3 的最新突破中，确保科学界能够使用这一功能非常重要。AlphaFold 3 将对药物设计非常有用，已经如此，但它也对许多其他基础生物学领域和仅仅理解生物学有用。人们以非常有创意的方式使用我们的 AlphaFold 3 服务器和模型。所以确保非商业学术工作可以免费使用对我们来说非常重要。看到服务器被接受和使用，真是令人难以置信。

AlphaFold has a long lineage of being open for academic and scientific use. It was really important with this latest breakthrough of AlphaFold 3 that we make sure the scientific community has access to this functionality. AlphaFold 3 is going to be incredibly useful for drug design, already is, but it's also useful for many other areas of fundamental biology and just understanding biology. People are using our AlphaFold 3 server and model in very creative ways. So it's very important for us to make sure there is that free use for non-commercial academic work. It's been incredible to see the take-up of that and the use of the server.

我想聊聊未来。你能透露一下 AlphaFold 还有什么新动向吗？

I'd love to talk a little bit about future. Can you give us a tease of what else is to come with AlphaFold?

就结构预测这个问题而言，在我的脑海里，我想完全解决它。我认为 AlphaFold 3 是朝着这个方向迈出的奇妙一步，一个重大突破。但它不是 100%准确。在这个领域，100%准确甚至意味着什么？就像许多科学领域一样，当你开始突破边界时，你会发现问题又引出了更多问题。这就是做科学令人上瘾的部分。AlphaFold 3 就是一个很好的例子：当你开始获得这些能力时，你会发现还有更深层次的问题我们想要研究和接近。所以更好地、更准确地理解结构对我们来说总是有趣的，但也不仅仅是静态结构。AlphaFold 3 模拟了这些晶体结构，这些几乎是这些分子的静态结晶版本，以及这些分子如何相互作用。但实际上，我们体内没有晶体。这些分子在溶液中，它们四处移动，是动态的。所以理解这些系统的动力学也会非常有趣。

In terms of structure prediction as a problem, in my mind, I want to completely solve this. I think AlphaFold 3 is a fantastic step on the way, a significant breakthrough. But it's not 100% accuracy. What does even 100% accuracy mean in this space? Like with a lot of areas of science, as you start to push the boundaries, you see that the problem opens up into even more problems. That's the addictive part of doing science. AlphaFold 3 is a good example of that: as you start to get these capabilities, you see that there are even deeper problems we want to be working on and stepping towards. So understanding structure better and more accurately is always going to be interesting for us, but then it's also not just necessarily about static structure. AlphaFold 3 models these crystal structures, which are almost static crystallized versions of these molecules, how these molecules interact. But in reality, we don't have crystals inside of us. These molecules are in solution, they're moving about, they're dynamic. So understanding the dynamics of these systems is also going to be really interesting.

在人工智能生物学中，GPT-3 时刻是什么样的，我们什么时候能达到？

What does a GPT-3 moment look like in AI biology and when do we get there?

如果我想想 GPT-3，它实际上是一个生成模型，生成文本的东西。对我来说，GPT-3 时刻是跨越了这样一个边界：从'我们有文本生成模型，它们生成一些东西，看起来像文本，但我不相信它是人类生成的'到 GPT-3 开始成为第一个让你觉得'哦，这看起来像人类'的点。所以这个生成模型实际上是在重现它训练数据的分布。生成模型是拟合它训练数据流形的东西。当我想想这应用于生物学时，你可以想象这些生成模型在 GPT-3 时刻开始重现事物在现实中的实际样子。这非常令人兴奋，因为这意味着这些模型吐出的东西要么在世界上真实存在，我们可以验证，要么甚至可能发现世界上存在的新事物，或者它们可能存在于世界上。这意味着它们可能是我们可以设计、制造或创造的东西，实际上稳定、有效并存在于我们的物理现实中。

If I think about GPT-3, this is really a generative model, something that's generating text. The GPT-3 moment for me was crossing over that boundary between 'we've got generative models of text and they generate some stuff and it looks like text but I'm not convinced that it's generated by a human' and GPT-3 started to be that first point where you're like 'oh this looks like a human'. So this generative model is actually recreating the distribution of data that it's trained on. A generative model is something that fits the manifold of data it is trained on. When I think about this applied to biology, you can think about these generative models actually starting to recreate at that GPT-3 moment what things would actually look like in reality. That's quite exciting because that means these models are spitting out things that either they actually exist in the world and we can kind of validate that, or maybe even discover new things that exist in the world or they could exist in the world. Which means they could be things that we could design or manufacture or create that would actually be stable and work and exist in our physical reality.

我认为生物学中很酷的一点是，与语言不同——语言生成人类水平的东西时我们能理解，因为它源于人类——但化学和生物学中的许多问题，连我们自己都难以理解。所以当我们迎来那个 GPT-3 时刻时，我觉得它看起来不会那么像 GPT-3，而更像 AlphaGo 的第 37 手。我们开始看到一些超越人类理解、但确实存在于现实世界和物理现实中的东西，这超出了人类的认知范围。这将是令人震撼的。事实上，我们内部已经在生成模型中看到了这一点：我们创造的设计，人类药物设计师会说‘我不太确定，我更喜欢这个’，但你在物理现实中测试后，生成模型是对的，人类是错的。这太迷人了。

And I think the cool thing about this in biology is that unlike with language, where when it generates something at human-level quality we can understand it because it is human-derived, a lot of problems in chemistry and biology we even struggle to understand ourselves. So when we get to that GPT-3 moment, I think it will look a lot less like GPT-3 but much more feel like move 37 in AlphaGo. We're starting to see things that are beyond human understanding but that do exist in the real world, in our physical reality, but are beyond human comprehension. That's going to be mind-blowing. In fact, we're starting to see that internally with our generative models, creating designs that a human drug designer would say, 'I'm not so sure about that, I much prefer this.' And then you test it in physical reality and the generative model is correct and the human is wrong. That's fascinating.

我喜欢第 37 手这个类比，当模型开始展现出创造力并超越人类时。

I love the move 37 analogy when the model starts to see elements of creativity and past the human.

没错。第 37 手是 AlphaGo 对阵李世石时的一步惊人妙手。那是比赛的第 37 步，震惊了世界，震惊了围棋界，因为人类无法理解。它看起来像是一个错误。在人类数千年的围棋历史中，从未有人下过这一步。但当你展开整盘棋时，你会发现这正是让 AlphaGo 在那场比赛中击败李世石的关键一步。我们将从这些模型中看到大量类似的行为，尤其是当我们将它们应用于人类天生难以理解的领域，比如化学和生物学时。

Exactly. Move 37 was this amazing move during the AlphaGo games against Lee Sedol. It was the 37th move of the game and it stunned the world, stunned the Go world because it was uninterpretable by a human. It looked like a mistake. No one had ever played this move in the entirety of thousands of years of human history playing Go. And it turned out, as you unrolled the game, that this was the critical move that allowed AlphaGo to beat Lee Sedol in that match. We're going to see so much of that sort of behavior coming out of these models, especially when we're applying them to things outside of native human understanding like chemistry and biology.

是的，我很喜欢。这也是我们今天的关键问题。那么，我们什么时候能看到第一个 AI 生成的药物进入临床，以及一期、二期和三期试验？

Yeah, I love that. Also our punch line today. So when will we see our first AI-generated drug in clinic and also in phase one, two and three trials?

我们在药物设计项目上取得了惊人的进展。我想，当我们开始让大量 AI 设计的资产（这些分子）进入临床阶段时，我们该如何真正思考参与临床开发，以最快、最安全的方式将这些分子带给患者，因为存在大量未满足的医疗需求。所以，我在思考与监管机构互动的新方式，以及整合我们预测模型的新方法——不仅预测分子如何针对疾病起作用，还包括它如何与身体其他部分相互作用，可能引发的毒性类型。我认为有很多机会可以简化和加速这一过程。甚至可能彻底改变我们对人体临床试验的看法，因为我们的 AI 模型能够让我们更快、更有针对性地设计这些分子，并对其作用机制有更深入的了解。这将改变游戏规则。但作为行业，我们还有很长的路要走，才能真正弄清楚如何实现这种改变。

We're making amazing progress on our drug design programs. And I think about as we start to get a whole bunch of these AI-designed assets, these molecules, into clinical phase, how can we actually start to think about engaging in that clinical development to get these molecules to people as fast and as safely as possible, because there's so much unmet medical need. So here I think about new ways to engage with regulatory bodies, new ways to incorporate our predictive models for not only how this molecule works for the disease, but how it interacts with the rest of the body, the types of toxicity it may induce. I think there'll be a lot of opportunities to streamline and speed up this process. Maybe even completely changing the way we think about human clinical trials as our AI models become so we can design these molecules much quicker and in a much more targeted manner with so much more knowledge about how they work. That will change the game. But I think we've got a long way to go as an industry to really work out how that changes.

最后一个问题。随着 Isomorphic 的成功，以及可能整个领域的成功，传统制药世界会发生什么？

Last question. As Isomorphic succeeds and potentially as a whole field succeeds, what happens to the traditional world of pharma?

我认为，从某种意义上说，制药公司将使用 AI。我认为五年后，不可能存在不用 AI 设计药物的世界。这是必然的。就像做科学不用数学一样。AI 将成为生物学和化学的基础工具。它已经是了，至少在 Isomorphic 的世界里是这样。每个人都会使用它。所以不会是‘哦，这是制药还是 AI？’而是两者合二为一，整个行业都会适应这一点。

I think they become, in some sense, pharma will be using AI. I think there's no world where in five years' time you will be designing a drug without AI. That is an inevitability. It'll be like trying to do science without using math. AI will be this fundamental tool for biology and chemistry. It already is, at least in Isomorphic's world. Everyone will be using it. So it's not going to be 'oh, is it pharma or is it AI?' It's going to be one and the same in the sense that the whole industry will adapt to that.

太棒了。Max，非常感谢你今天加入我们。这是一次精彩的对话。

Amazing. Max, thank you so much for joining us today. This was a fascinating conversation.

是的，这是我的荣幸。谢谢。

Yeah, it's been a pleasure. Thank you.