本集简介
双语字幕
仅展示文本字幕，不包含中文音频；想边听边看，请使用 Bayt 播客 App。
欢迎收听FYI，这是一档关于四年创新的播客。
Welcome to FYI, the four year innovation podcast.
这档节目深入探讨科技驱动的颠覆性变革，因为投资创新，首先要理解它。
This show offers an intellectual discussion on technologically enabled disruption because investing in innovation starts with understanding it.
想了解更多，请访问arc-invest.com。
To learn more, visit arc-invest.com.
Arc Invest是一家专注于颠覆性创新投资的注册投资顾问公司。
Arc Invest is a registered investment adviser focused on investing in disruptive innovation.
本播客仅供信息参考，不应作为投资决策的依据。
This podcast is for informational purposes only and should not be relied upon as a basis for investment decisions.
其中任何内容都不构成ARC对任何服务或产品的明示或暗示推荐。
It does not constitute either explicitly or implicitly any provision of services or products by ARC.
节目中关于公司或证券的所有观点，仅代表ARC或嘉宾的个人看法，并非ARC的买卖或持仓建议。
All statements made regarding companies or securities are strictly beliefs and points of view held by ARC or podcast guests and are not endorsements or recommendations by ARC to buy, sell, or hold any security.
ARC投资管理的客户，可能持有本节目提到的证券。
Clients of ARC Investment Management may maintain positions in the securities discussed in this podcast.
你好。
Hello.
欢迎收听FYI，这是Arc Invest推出的四年一度创新播客。
Welcome to FYI, the four year innovation podcast from Arc Invest.
我是布雷特·温顿。
I'm Brett Winton.
今天我们请来了Nemo，他负责为我们追踪多组学领域中人工智能与药物发现的前沿动态。
With us today, we have, Nemo who covers, AI and drug discovery for us in the multiomics space.
嗨，Nemo。
Hi, Nemo.
最近怎么样？
How are you?
挺好的。
Great.
是的。
Yeah.
今天这期节目特别精彩。
We have a great show today.
我和Nemo聊了聊Insilico Medicine的创始人兼CEO Alex Zhavoronkov，这家公司是个用AI做药物研发和发现的平台。
Nemo and I, talked to Alex Zhavoronkovic, the founder and CEO of Insilico Medicine, which is a AI drug development and discovery platform.
Nemo，你给我们讲讲AI药物发现这个领域现在是个啥情况？
Nemo, lay out the AI drug discovery landscape for us.
就是，这事儿到底有啥前景？
Kinda how do how are what is the promise here?
这些公司都是怎么玩的？
How are these companies approaching it?
咱们在市场里到底看到啥了？
You know, what are we seeing in the marketplace?
那在你看来，Insilico在这个圈子里占个啥位置？
And then where does Insilico fit into kinda your understanding of the space?
对。
Yeah.
AI药物发现的大方向在于，AI特别擅长从数据里发现模式，只要数据量足够大，它就能给出靠谱的结果。
So the kinda the big picture kinda promise here with the AI drug discovery is that, you know, AI, it's it's pretty good in seeing certain patterns in the data or generalized, you know, if there's, again, if there's enough data and the like, that kinda gives us the the result.
说到药物发现的整个流程，我觉得AI最有潜力的地方，是在‘命中先导化合物’到‘优化先导化合物’这个环节。
And so, you know, when we think about in the direct kinda whole kind of space of the discovery, you know, where we could see the most kinda interesting kind of application is in this kinda hit to lead and the lead optimization kinda part of the drug discovery.
这个阶段就是你真正去设计药物、优化药物性能的时候。
So this is the part that you're actually trying to design the drug and kinda optimize the drug.
比如拿抗体举例，你得先找到能结合目标的抗体。
So for example, if you take, you know, antibodies as as as the example here, you know, so you need to kinda find the antibody that is buying buying to your target.
但光能结合还不够，这个抗体还得具备其他很多特性，才能真正成为一款好药。
But that antibody, in addition to buying to your target, also needs to kinda have certain other qualities to be actually a really good drug.
AI能帮你同时优化多个参数，不像传统方法那样，一个一个步骤慢慢试。
So AI can actually really help you do that kind of multi parametric optimization and kind of optimize the kind of different parameters at once versus you in a traditional kind of lab in traditional drug discovery doing each step once at a time.
另外，对于抗体和靶点的结合，只要你实验数据够多，模型就能从中学到规律，掌握更广的范围。
The second part here is that, you know, for the, again, for the antibody, for the binding, you know, if you run a big enough experiment, your model can learn from that experiment, you know, basically, the kinda the the bigger space.
下一次做新项目的时候，你就能更快地找到那个能精准结合目标的抗体。
And then for the next kind of campaign, you can be actually faster in developing that, finding that antibody that's binding the your target of interest.
所以这算是目前最常用的一个领域了。
So that's the kind of the the most kind of common, area where it's kinda used.
你知道的，他们用它，其实我们也能用它来做靶点发现和验证。
You know, they use it or we can also use it in the target discovery and validation.
在临床前阶段，现在能看到类器官的一些应用，但最普遍的还是前面说的这个部分。
In preclinical stuff, you'll see now what happens with with organoids, but this is the kind of the most common area.
对。
Yeah.
所以，药物研发的整个前端流程，其实可以这样看。
So, basically, the there's this whole front end of the drug development process.
在传统方式下，光是这一块儿就得花上五年半的时间。
It could you know, in traditional space, it might take you five and a half years to do it.
要让一个药，或者一堆候选药，一路走完这个流程、进入临床试验，中间得淘汰掉一大堆，最后能上市的，传统做法光是成本就得两亿八千万美元。
And, to get a drug, like, a drug or or enough drugs through that whole process into clinical trial with enough candidates that are gonna fail out into clinical trial that, to actually get a drug to market, it might cost you $280,000,000, traditionally to to just get there.
而AI药物发现的前景就是——
And the promise of AI drug discovery is, hey.
在进入人体试验之前，我们能用各种丰富的数据源，甚至让AI模型驱动自动化实验，来开发结合力更强、更有可能触及全新生物学机制的药物。
All of this stuff that we're doing before we go into human, we can use, you know, rich data sources, you know, potentially automated kind of experimentation driven by your AI models to, develop, you know, better binding, more potentially novel, like going after unique biology.
就算大家都在研究同一个靶点，你也能做出结合得更好的分子，更快、更便宜地把候选药推上市场——虽然这套方法还没被任何人完全验证过，但理论上它成功的概率更高，因为结合得更精准，而且花的时间和资源还更少。
But even if you understand the biology and everybody's going after it, you can develop something that gloms onto it better, and you can get to market, you know, faster, cheaper within a drug candidate that, you know, though this is yet to be validated all the way through by anyone really, but should be have a higher odds of actually kinda delivering the successful outcome that you want because it binds better, even though you spent less kind of, like, time and resources getting to market.
这就是AI的潜力所在。
That's the promise.
没错。
Yes.
就像我之前说的，临床前阶段主要分三块：药物设计、临床前研究这些。
And I I I would say that, you know, as I said, this kinda the the preclinical phase kinda has a three kinda target, designing the drug, and the preclinical stuff.
药物设计是最主流、最常被关注的方向。
Designing the drug is the kinda the kinda the, you know, the go to area.
靶点鉴定嘛，确实也有人在做。
The target ID, you know, I think that's there are people doing that.
而且说实话，有些低垂的果实就在那儿——比如直接挖掘现有的文献数据就能找到机会。
But and, yes, there I think there is somewhat, like, low hanging fruit that you can just mind the re like, literature that exists.
我觉得关键在于，如果我们不懂生物学，那一切都没法搞下去。
I think the right way there is really we don't understand if we're not for the biology.
所以真正有效的方法是整合多种组学数据，把扰动实验、单细胞分析和患者数据结合起来，才能发现新的靶点。
So it's really kinda multiomic orthogonal kinda datasets where you're kinda combining, you know, perturbation, single cell, and then combining it with the patient data to actually find the new targets.
说到进入临床前阶段，其实我们在做IND研究的时候，问题就来了。
To go to a preclinical, you know, side preclinical kind I mean, the the issue, you know, when we are trying to kinda make the IND studies.
你知道，我们通常在二维细胞培养里完成前两个阶段，但问题就在这儿。
You know, they're kinda again, the whole issue that there is not a failure is because, you know, we will do these first two phases, for example, in the two d cell culture.
接下来的临床前实验是在活体动物里做的，比如小鼠。
Then this preclinical stuff is in vivo, and we do it in mice.
这很明显嘛——你换了个系统，从细胞跳到小鼠，再从老鼠跳到人身上。
I mean, obvious and then you're you're changing your mind to another system, and then from there, you're going to humans.
所以临床前阶段在发现环节的失败率特别高，但我觉得问题不在别的，主要是模型本身不够好。
And so the preclinical phase can you know, that kind of part of the discovery has a huge drop, but I think it's more because the models are bad versus, you know, it's the something else.
就算你能在小鼠身上搞定癌症，但如果能跳过小鼠这一步，直接解决人的问题，那才叫真本事。
Mouse cancer even though you would have solved the human cancer if you were able to skip the mouse step.
所以这也是我们需要改进的另一个方面。
And so that's another area where we need to improve things.
这儿有个笑话。
There's a joke.
我以前的博士导师是一位特别有名的癌症小鼠研究专家。
So one of my PhD advisers, he's a very famous mouse cancer biologist.
我是说，他可是第一个做出癌症小鼠模型的人。
I mean, I think he made the first mouse to be to decide the cancer.
所以实验室里一直流传着一个笑话。
And so, I mean, it it it there was, a running joke in the lab.
我们能在小鼠身上治好癌症一百多次，可人类身上还是没辙。
Like, more than 100 times, we can cure the cancer in mice, but, you know, it's still in the in humans.
这中间的差距太大了。
So there's a huge difference.
对吧？
Right?
是的。
Yeah.
从整体来看，我们觉得药物上市的流程未来会这样演变：传统方式下，算上所有失败和出错的成本，一款药从研发到上市要花大约24亿美元，耗时十三年半。
So and then at an aggregate level, at least the way we think it's gonna evolve is that to get a drug to market traditionally costs you, including all of the things that fail and go wrong along the way, around $2,400,000,000 and takes you thirteen and a half years.
我们认为，而且有理由相信，这个数字现在其实已经算保守了——因为FDA现在正在推动更快的审批流程，有了AI，药物上市成本能降到6亿美元左右。
We think and there's reason to believe this is actually a conservative estimate now given how the FDA is enabling, you know, faster steps along the way, that with AI, you can get a drug to market for $600,000,000.
也就是说，成本降到四分之一，时间从十三年半缩短到八年，甚至可能更短。
So call it 25% the cost, in in, you know, eight years instead of thirteen and a half and so and perhaps even less.
所以整个行业的前景就是：花更少的钱、花更短的时间，把真正能帮到人的药推到市场上。
So that's the promise of the industry is that you spend less money and take less time to bring, you know, something that's gonna help people to market.
这样一来，投资回报率自然就高多了。
And so, therefore, your ROI just is much higher.
它会更接近科技公司的回报水平，而目前整个行业的研发回报率低得可怜。
It's, you know, closer to kind of a technology company ROI, whereas right now, industry wide return on r and d is pitiful.
大概只有4%。
It's around 4%.
所以现在公开市场上，像Recursion和AbSci这样的公司就在做这个方向。
So how does so an example companies going after this are Recursion and AbSci in the public markets.
还有一些没上市的公司，比如Insilico，也在跟进这个领域。
There's companies that are still private, including Insilico going after this.
那Insilico的做法，等会儿我们会讲到，跟我们之前看过的其他公司有什么不同呢？
How does Insilico's approach, as we'll learn about, differ from some of the other companies that we've seen and looked at in the space?
简单说，他们从头到尾整合了好几个不同的模型。
So, basically, what they have done, you know, they, kinda try to kinda put different models, from start to finish.
在靶点发现这个环节，他们把大量公开的数据集和科研论文都收集起来，然后把这些模型拼在一起，挖掘潜在的靶点。
And so for this kind of first piece of the kinda target ID, they kinda aggregate a lot of different publicly available datasets and also kinda research papers, publications available, and then basically try to kinda stitch the the different models together to kinda mine that area and find some of the potential targets that that they can go that they can go after.
接下来就是化学部分了。
And then after that, they had the kinda chemistry.
他们管这个叫Chemistry 42，或者类似的名字，是个生成式平台。
They call it, I think, chemistry 42 or something like that, which is their kinda generative kinda platform.
这个平台用的是多智能体自编码器，专门去找适合那个靶点的小分子药物。
And so which is gonna multi agent kinda autoencoders basically kinda trying to, you know, find the kinda molecule, which is the small molecule, for that for that particular kinda target.
他们在做这个的同时，也会优化一些其他对毒理研究很重要的因素。
And then kinda also when they are doing that, kinda trying to, optimize the, as I said, few few of the other things which are important for the for the tox studies.
所以这基本上是他们的主要方法。
So that's kinda the primary their primary approach.
这其实更偏向于全在计算机上完成，就像他们说的，验证工作都是后面才做的。
So it's really about much more being in silico, how they say it, and then they do the kinda validations later.
对。
Yeah.
相比之下，比如Recursion，他们的重点好像不太一样。
In contrast to, say, Recursion, which seems to focus yeah.
你具体说说两者的区别。
Give me the contrast.
对。
Yeah.
没错。
Yeah.
所以呢，说实话，我挺看好Recursion的，提前打个预防针。
So, like, you know and, obviously, I'm a fan of Recursion, so just, you know, disclaimer.
但说实话，Recursion的做法其实是把生物学和化学搞成一张张地图。
But, yeah, I mean, so Recursion basically kinda took you know, makes the kind of biology and chemistry maps.
核心想法就是，没错，生物学太复杂了，咱们得想办法把它捋清楚。
So whole idea is that, like, yes, we there's biology is very complex.
所以我们试着用大概50%的线索来画这张图，构建不同的生物背景，然后用CRISPR敲除基因，或者用小分子做干预。
And for us to understand it, we're gonna, you know, use the kinda try to map it in, like, 50% lines so trying to have kind of different kind of biological kind of context and then do basically different perturbations with either with CRISPR to remove the genes or we do it with a small molecule.
接着我们做各种表型检测。
And then we're gonna do readouts both kind of phenotypic.
一方面看细胞的成像，另一方面也抓分子层面的数据。
So we, you know, look at the kind of imaging of the cells, but also get the molecular one.
通过这种方式，我们才能真正搞懂生物学，进而发现新的靶点。
And like that, we are actually, you know, trying to really understand biology better and then actually come up come up with the new targets.
对。
Yeah.
我觉得吧，与其用所有学术文献来训练模型，不如直接用底层的生物学数据来训练，你觉得哪个更好？
And so the way I think about it is you could imagine imagine, is it better to train a model on all of the academic literature that's out there, or is it better to train a model on the underlying biology?
Recursion 的做法，其实更偏向于——
And Recursion's approach seems to lend itself more to, hey.
我们直接在生物学层面做自动化实验，让模型通过直接分析这些数据来理解生物学，而不是去翻一堆学术论文，有的靠谱、有的有问题，还得费劲筛选出好论文再拿去训练。
We're gonna in enable automatic experimentation on the underlying biology and have the model get an understanding of the biology by studying that data directly as opposed to we're gonna take a vast body of academic literature, you know, some of which is valid, some of which is flawed, and try to annotate it to, like, just identify the good papers and then train it on those papers.
当然啦，两家公司其实都会用这两种方法的组合，但真正投入资源和资金的方向，还是有区别的。
And, of course, both companies are using, you know, some combination of approaches, but it's where kind of capital allocation and resources are are really going.
对吧？
Right?
是的。
Yeah.
对。
Yeah.
另外呢，Insilico 他们也用一些公开的数据集。
And and just to be also, you know, Insilico, I mean, they also take publicly available datasets.
对吧？
Right?
但问题是，这些数据量都太小了。
But, like, the problem is that all of that is so small.
所以呢，Recursion其实也是做细胞培养，对吧？
And so, again, you know, like, each you know, it's Recursion also, like, it's a it's a cell culture.
对吧？
Right?
每种方法都有自己的问题。
Each has issues on its own.
对吧？
Right?
没错，确实存在一些问题。
So, right, like, obviously, there is some issue.
但我觉得，Recursion的做法更有潜力——尤其是如果能把它们和真实世界的数据结合起来，比如他们正在和Tempus这些公司做的那样，这样就能挖掘出更深层的生物学特征，然后去病人身上找这些标志。
But to me, the the for the recursion approach gives you more kinda you know, if you pair it with real world data, which they are doing with Tempus and the other ones, then you're actually gonna, I'm getting here very deep biology signatures, which I can go to patients and look for them there.
所以，我就可以去验证一下这个靶点。
And then, therefore, I can, you know, maybe validate the target.
相比之下，如果你去做文献检索，有时候也能发现一些现成的、容易摘的果子。
Versus kinda, you know, the lit and there could be also when you do the kinda literature search and all of that, know, there could be some along hanging fruit.
对吧？
Right?
所以，最近未来之家刚出了那个分子，你知道的。
So, I mean, the the future house just came out with, right, you know, the molecule.
那个更像是药物重定位，我觉得。
It is like more it it but that's more like, yeah, the the repurposing, I would say.
你懂我的意思吧？
You know?
但是
But
对。
Yeah.
我的意思是，现在行业里的标准实在太差了，所以用AI和机器学习技术，既挖掘现有数据，又生成新数据，这种聪明又积极的做法，完全有潜力让新药研发变得更高效。
I mean, I I think the way to think about it is the industry standard right now is so pitiful that there's a lot of, like, you know, basically doing an intelligent aggressive approach of devoting AI and ML techniques to both existing data and to help you generate novel data, should broadly be a way to be able to, you know, develop drugs more efficiently.
我觉得这个领域最让人兴奋的是，可能一两家公司会率先跑出来，形成数据生成和验证的飞轮效应，让他们的底层模型迅速壮大。
And there's a reason to believe what's provocative to me about this space is there's the potential that, you know, one or two players kind of gets a nose ahead, which then gives them this flywheel of data generation and kind of validation, that then allows their underlying model to grow very large.
这样一来，行业的整合方向就会转向这个创新引擎，而不是像现在这样，大公司靠的是庞大的销售网络。
And so then the consolidation in the space moves towards kind of whatever this innovation engine becomes as opposed to right now, the very big players in the space are basically have massive distribution.
现在的巨头本质上都是销售型公司，手里有肿瘤或囊性纤维化这些重磅产品线。
They're basically sales companies that take drugs and you know, and they have a franchise in oncology or they have a franchise in cystic fibrosis.
行业目前的整合，靠的就是这套模式。
And that's the mechanism by which the industry is consolidated.
但未来，真正推动行业整合的，可能是由AI驱动的底层发现平台。
There's the potential that, like, actually, the underlying discovery platform driven by AI becomes the consolidation mechanism of the industry.
你看看现在行业里的大公司，动不动就值几百亿美元，根本不是问题。
And if you look at the big companies in the industry, you know, these are hundreds of bill worth hundreds of billions of dollars easily for big companies.
所以完全有可能出现一个AI药物发现平台，估值高达几百亿美元。
And and so it doesn't it you know, you could end up with an AI drug discovery franchise that is worth hundreds of billions of dollars.
这事儿啊，确实有可能发生。
That's that's you know, could could occur.
我觉得这事儿其实挺容易预见的。
I think that's an e easy possibility to foresee.
我不是说非得是哪一种路径才对，但这里的机遇在于：谁能用AI把研发回报率提上去，谁就能用这个优势生成更好的数据，反过来又进一步提升研发效率。
Not to say that one or the other is necessarily gonna do it, but that's the, promise here, is you have a better return on r and d company that then uses that advantage to generate better data for itself, which then improves its return on r and d.
这样一来，它就能一直领先别人。
And so it kinda stays ahead of everybody.
咱们走着瞧吧。
And we'll see.
对啊。
Yeah.
而且你知道吧，很多小分子药物其实效果也没那么好。
And, you know, it could be also the kinda a lot of the small molecule stuff, you know, is actually not that great as as drugs.
但要是看生物药，比如细胞基因疗法，或者抗体药，这些更像是一种治愈手段，它们的靶向性和精准度，远超小分子药物。
But when you think about biologics, you know, is that the kind of cell gene therapies, which are kind of more of the cures or, like, antibodies, they are much more precise, you know, and specific than the small molecules.
而且，你想想细胞和基因疗法这个领域，你的药物本质上就是得把它送进去，然后让它自己发挥作用。
And then, you know, especially if you think about the, you know, cell and gene therapy space, right, You're basically you know, your drug there, like, there is a one you you have to deliver it, and then it has to kinda act on.
所以你实际上是把‘药物该去哪儿’和‘它该做什么’这两件事分开了。
So you're, like, basically decoupling kinda specificity, you know, where where the drug needs to act and what it needs to do.
小分子药物可就老有这问题了。
With a small molecule, you have that issue all the time.
对吧？
Right?
你懂的。
You know?
所以这事儿嘛，确实如此。
So that that's yeah.
所以我总觉得，生物药这一块——不管是抗体、细胞疗法还是基因治疗——才是真正的方向。
So that's where I feel that kinda more on the biologics side, which kinda takes all of these, you know, antibodies, cell gene therapy, you know, that's the way to definitely to go.
某种程度上，要实现真正的个性化医疗，就得靠生物手段才行。
Well, in some ways, it's it's, like, it's required for for truly personalized medicine to do it biologically.
对吧？
Right?
小分子药物就像到处乱插，哪儿都可能起作用，但抗体不一样——我这是专门给某个锁配的钥匙，就为了精准地拧开它。
A small molecule is something that, like, it's it's it plugs in all over your body in all kinds of different ways, right, as opposed to, like, I have this particular key that I'm designing for this particular lock that I know I specifically want to turn.
没错。
Exactly.
完全对。
Exactly.
其实这正是为什么抗体，还有我觉得小分子药物不太适合用生成式AI来设计的原因。
I mean and that's, like, why antibodies and and also I I think I think the small molecules are actually not that amenable for the AI generative AI.
你知道的吧？
You know?
为什么抗体特别适合呢？
Why antibodies are so amenable?
因为抗体本质上就是蛋白质。
Because an antibodies is kind of as proteins.
对吧？
Right?
AI所学习的解决方案空间本来就已经受限了，因为DNA序列和最终的抗体之间是有明确关联的，你知道的。
The solution space AI is learning on is already kind of constricted, and there's like a sequence, you know, the DNA sequence versus the kind of final final antibody, there is there is a kind of relation between that.
所以啊，你的机器学习AI其实能学会并生成新的抗体。
And so like that, you can your your machine learning AI can actually learn and kinda develop new ones.
但小分子就不一样了，你根本没这种东西，咱们能随便合成任何东西。
With small molecules, you don't have something like that, you know, because we we can make anything.
所以我对这一块特别怀疑。
So that's why I'm very skeptical about that side.
然后，当然了，还有基因编辑这一块。
And then, obviously, this you know, the the gene editing.
对吧？
Right?
我的意思是，mRNA也是这样。
I mean or the mRNA as well.
它的序列对我们来说是直接相关的。
Like, it's it's sequence is buying to us.
序列非常具体，基因编辑也是同样的道理。
Seek sequence is a very specific, the gene editing as well.
所以这要精确和明确得多，我觉得。
So it's just much more, yeah, precise and the specific, I would say.
那就好好享受和亚历克斯的对话吧。
Well, so enjoy the conversation with Alex.
我们一路聊到后面，深入探讨了长寿这个话题。
We get it and to the end, we get into some really meaty discussion about longevity.
所以千万别走开，尤其是如果你想像永生一样活着——当然我不是说人人都想，但很多人确实这么想。
So make sure you stick stick around for that, particularly if you wanna live forever, which I'm not sure everybody does, but I think a lot of people do.
好好听这场对话，也谢谢收听《四年创新播客》。
So, enjoy the conversation, and thanks for joining FYI, the four year innovation podcast.
你好。
Hello.
欢迎收听FYI，四年创新播客。
Welcome to FYI, the four year innovation podcast.
欢迎你，Alex Zh。
Welcome, Alex Zh.
我其实有点不敢念你姓氏，不过你屏幕上有，自己看吧。
I, you know, really hesitate to even try to pronounce your last name, but it's, you know, you can read it on the screen.
不过我以前还真念对过一次。
And I I have done it successfully once.
Alex是Insilico Medicine的创始人兼CEO，一位非常出色的科学家。
Alex is the founder and CEO of Insilico Medicine, a talented scientist.
Alex，跟大家简单说说你是谁，现在在做什么，Insilico又是个什么样的公司？
Alex, lay out kind of who you are, how you are where you are today, and where Insilico is, please.
当然可以。
Sure.
我现在碰巧在纽波特海滩参加一个会议，一边做报告，一边跟好几个合作伙伴碰面。
So today, I'm serendipitously in Newport Beach at a conference where I'm presenting, but also it has multiple partners here.
所以我们终于能聊点专业 stuff 了。
So we do get to speak shop.
我本来是搞计算机科学的，科班出身。
I am a scientist by training, computer scientist.
我在加拿大女王大学拿了前两个学位，后来进半导体行业，赚了点钱，终于能去追我这辈子最想干的事——延长人类寿命。
Did my first two degrees in computer science at Queen's University in Canada, worked in semiconductors, made enough money to pursue my lifelong dream of kind of chasing longer lifespans.
我所有的工作，全围绕着抗衰老研究。
So everything I do is focused on aging research.
别的事儿，我真不关心。
I'm pretty much I don't care about anything else.
我在约翰霍普金斯大学做了最重要的研究。
Did my great work at Johns Hopkins.
2014年，正好赶上深度学习爆发的前夜，我创办了Insilico，想用深度学习的力量去破解生命的奥秘，专门攻克衰老问题。
Worked for a number of biotechs 2014 at the cusp of the deep learning revolution started in silica to apply the power of deep learning to understanding life itself and going after aging.
我们公司其实是在英伟达GTC大会上诞生的，这事儿还挺有缘分的。
So we actually were born out of the NVIDIA GTC conference, which is kind of a nice nice touch.
从那以后，我们陆续提名了多个候选药物。
And since then, we've nominated.
我们一直专注做纯算法的AI公司，直到2019年才把软件推向市场，现在很多人都在用我们的AI。
So we operated as an al purely algorithm focused company till 2019 and then released our software to the market, so people use our AI.
我本人也用这套软件来发现新药，然后把药卖给制药公司，也就是授权出去。
I also use the same software to discover drugs and then sell those drugs to the pharmaceutical companies, so license them out.
这基本上就是我们的核心商业模式。
So that's kind of our main business model.
从2020年开始，我们已经提名了22个开发候选药物，这一步差不多就是进入人体临床试验前的最后一关。
And since 2020, we've nominated 22 developmental candidates, so that's kind of one notch before human clinical trials.
其中有10个进入了临床阶段，还有一个已经在特发性肺纤维化患者身上完成了二期a期疗效试验。
10 reached clinical stage and one has completed Phase 2a efficacy study in humans in idiopathic pulmonary fibrosis.
对此，我真的很自豪。
So I'm very deeply proud of that.
我们尽量把研究成果都发表出来。
We try to publish as much work as we can.
所以我们现在差不多每六天就发表一篇论文，都是发在不错的专业期刊上。
So we nowadays publish at the rate of about one paper every six days in good specialized journals.
我们也会开源一部分软件，让社区能用上我们的工具，也获得大家的反馈和验证。
And we try to open source some of our software to make the community try our tools and get community level validation.
我特别关注衰老问题，建议所有在听的朋友们，从这个角度去思考人生。
I'm very focused on aging, and I recommend all of you who who is listening to start thinking about life with that angle.
因为不管你这辈子挣了多少钱、取得了什么成就，最终都会慢慢消失，最好能彻底消失。
So because whatever you do in life, money you have, whatever you have achieved, it's going to gradually disappear and preferably completely disappear.
那我们为什么不团结起来，一起对抗这个最大的敌人呢？
So why don't we unite and fight this great enemy together?
Insilico 是2014年成立的，一开始就是想用AI做药物发现，最早其实是搞AI和生物学，研究长寿，后来才转向药物发现的。
So Insilico starts in 2014, you know, focusing on using AI in in drug discovery, or at first, it was just like AI biology and understanding longevity, and and then you moved into drug discovery.
是这么个转变过程吗？
Is that the transition that occurred?
我们最初其实是一家纯粹的算法公司。
So, originally, we started, as an purely algorithmic company.
所以我们当时专门用深度学习来处理制药行业里各种不同的任务和问题。
So we were applying deep learning specifically to many different tasks, many different problems in the pharmaceutical industry.
我们一开始是做靶点发现的，就是想弄清楚到底是哪个蛋白导致了疾病，又跟衰老有关。
We started as a target discovery company, so trying to understand which protein drives the disease and also is implicated on aging.
但通常呢，不是只有一个蛋白在起作用，而是一大堆蛋白共同作用。
And usually, there's just not one protein, it's it's a variety.
所以我们想搞明白这些蛋白在衰老和疾病过程中，到底是怎么相互影响、动态变化的。
So we were trying to understand the intimate interplay of those proteins in time, in aging and disease at the same time.
结果发现，那时候制药公司，甚至到现在，对靶点发现都没太大兴趣。
Turned out, at that time, the pharmaceutical companies and even today, they are not exactly interested in target discovery.
靶点科学其实特别重要，但大部分工作其实都是在学术界做的。
So target sciences is a very important field, but most of that is performed in academia.
可学术界手头没那么多钱，没法跟AI公司深度合作。
And academics do not have huge budgets to partner with AI companies.
所以我们只能转型，自己搞起了化学研发。
So we had to pivot and develop chemistry expertise.
所以我们开始和制药公司合作，帮他们针对自己的靶点、自己的蛋白，开发能做成好药的化学物质。
So we started working with pharmaceutical companies to work on their own targets, on their own proteins to develop specific chemical matter that could make a good drug.
大概在2015年、2016年的时候，我们开始用生成对抗网络。
And that's where we started utilizing in 2,015, 2,016 generative adversarial networks.
那时候我们就是因为这个出名的，后来又转到生成式强化学习，专门设计具有理想特性的分子。
So that's what we were famous for and then transitioned to generative reinforcement learning to design molecules with the desired properties.
说白了，我们不再是在干草堆里找针，而是直接造出一整批完美的针，再从中挑出最合适的那根。
So basically, instead of looking for a needle in a haystack, we started generating perfect needles and then selecting out of those needles, you know, the perfect one.
然后制药公司给我们抛来了一大堆问题，范围特别广。
And then pharmaceutical companies challenged us to a very large spectrum or a broad spectrum of problems.
有些是临床相关的，有些是生物标志物的发现。
Some of those problems were clinical, some of those problems were biomarker discovery.
我们就像深度学习大厨背后的帮厨，啥都得会点。
So we were pretty much like a sous chef for the chefs in deep learning.
后来我们意识到，既然手里有这么多现成的算法积木，能拼出整个药物研发的全流程，干脆就自己动手，推自己的药吧。
And then we realized, so then since we have this entire Lego, Lego blocks of algorithms that could be fit together in order to cover the entire drug discovery and development pipeline, we actually started thinking about pulling the trigger on our own drugs.
直到2019年，我们都没碰自己的药物。
And that's so until 2019, we did not touch our own drugs.
明白了。
Got it.
从2014年创立到现在，AI领域简直天翻地覆了吧？
And but it seems like from inception 2014 to now, gosh, has the AI landscape shifted dramatically dramatically?
我的意思是，‘注意力机制就是一切’嘛。
I mean, the, you know, attention is all you need.
Transformer架构是公司成立三年后才出现的，但这一下子打开了无数可能。
The the transformer the introduction of the transformer architecture happened, you know, I guess, three years after the start of the company, and that's unlocked so much.
以前搞图像生成，大家都用生成对抗网络，我还记得呢。
You know, generative adversarial networks, I remember that was the way to do image generation.
现在全都是扩散模型了。
Now that's all done on diffusion.
Insilico是怎么跟上这些新工具的节奏的？
How has Insilico reacted to kind of, like, adapting in these new tools?
那现在你该怎么定位呢？之前那个工具不是挺管用的吗？
And then where do you sit today relative to it's like, oh, you had a tool that was working?
在很多情况下，新工具的潜力明显更强。
Seems like in a lot of cases, it's like the the new tool is probably more potentially performant.
这到底是怎么发生的？
Like, how does that happen?
没错，完全对。
Absolutely.
现在我最大的挑战，就是平衡AI技术的飞速发展和药物发现那漫长得多的验证周期。
So nowadays, my greatest challenge is balancing the advances on the AI side with the kind of lengthy validation cycles on the drug discovery side.
对吧？
Right?
你可能几周就训练出一个超智能模型，但要验证它能不能生成真正有效的分子，却得花上好几年，还得等实验数据确认。
So it takes you sometimes a few weeks to develop a superintelligent model, but then it's gonna take you several years to see if it can produce molecules where the generation conditions are confirmed experimentally.
所以你得一直等到临床人体试验阶段才能知道结果。
So you need to actually wait all the way to humans.
所以你现在看到的临床试验里那些表现特别出色的分子，其实都是2020年、2021年、2022年我们早就做出来的成果。
So whatever you see right now in clinical trials that we have and some of those molecules are performing extremely well is the result of what we had in 2020, 2021, 2022.
当然，我们现在正在加快验证的节奏。
Now of course we are accelerating the validation cycles.
但从技术演进的大趋势来看，我发现咱们这个AI药物发现领域，并不是非得用最前沿、最顶尖的模型不可。
But in terms of the trend technology evolution, what I figured is that in our industry, in AI powered drug discovery, you not always you need to be at the absolute cutting edge with the most frontier model.
真正关键、最终起决定作用的，是经过实验验证过的模型。
What is important and ultimately important is experimentally validated models.
只要这个模型能把某个任务做得特别好，比如你让它预测某个细胞系或某种生物体的毒性，
So if the model can perform a certain tasks task extremely well, so basically you tell it, okay, predict toxicity in this cell line and in this organism.
它能根据你给的分子特性、生物假设，或者两者结合，准确搞定，那对这个具体任务来说就完全够用了，对吧？
If that model can do it really, really well, looking at the molecular properties that you are giving it or the biological hypothesis or both is good enough, right, for for this specific task.
如果你能证明它连续50次都干得漂亮，
If you can show that he's done it over, let's say, 50 times.
那就相当于你团队里请了个特别靠谱的专家。
It's like having a very good validated expert in your team.
专家是老是新，其实都没关系。
Doesn't matter whether the expert is old or young or whatever.
对吧？
Right?
只要能把活儿干好就行。
So as long as they can do the job.
我就特别喜欢丁晓平那句话：不管黑猫白猫，能抓老鼠的就是好猫。
I really like, you know, Ding Xiaoping's quote, it doesn't matter whether the cat is black or white as long as can catch mice.
所以只要你这模型真能搞定事儿，迟早你会用它——我可以告诉你，现在的情况是……
So if your model can really do the job, it will you you will employ it at some point in time because I can tell you that currently
你可能养了只黑白花的猫，结果老鼠泛滥成灾，它却只会抓老鼠。
You might have a black and white cat, and then you're overrun by rats, and it can only catch mice.
对吧？
Right?
但我很好奇，这种高度稳定可靠的模型，到底是内部用起来更有价值，还是在对外合作时特别关键？
And so but I think I think the I do is the particularly, like is it useful to have a very predictably performant model for internal purposes, or is that particularly useful in kind of, like, partnering?
这样你就能去找外部的人说，你看。
Because then you can go to somebody external and say, look.
我有这么多证据证明这东西真的管用。
I have all of this evidence that the thing works.
所以，比起说‘我有个特别牛的模型，你只能信我’，这种方式在药物授权的商业逻辑上更说得通。
And so then maybe it it fits into the kind of, like, business logic of kind of, like, licensing drugs in in a better way than saying, I have the most insane Ferrari of a model here, but you just have to trust me.
它肯定能行。
It's gonna work.
是的。
Yeah.
布雷特，你其实一开始就碰到了AI药物发现里最敏感也最重要的问题。
So, Brett, you actually originally touched on the most kind of sensitive and the most important topic in AI powered drug discovery.
对吧？
Right?
那你怎么跟得上呢？
So how do you keep up?
对吧？
Right?
我觉得咱们还在这儿聊着呢。
And I think that we're still on this subject.
你怎么跟上那些不断冒出的顶尖模型？
How do you keep up with the most capable models that are coming up?
对吧？
Right?
你可能连老鼠都比不过。
And you can be outcompeted by rats.
对吧？
Right?
我就想点出这个关键结论。
So I just wanted to make this kind of punchline.
所以，最关键的是，你得不断研发最前沿的模型，但同时也要有一批你信得过、经过验证的备用模型，作为后盾。
So it's extremely important to have a different the most frontier model that we can have constantly in development, but you need to have a long tail of models that you trust, that you can fall back onto, that are validated.
所以你有一大堆经过验证、值得信赖的模型，你的药物化学团队信得过，生物临床团队也信得过，连合作伙伴都放心用。
So you have this entire zoo of models that work, that are validated, that you trust, that your medicinal chemistry team trusts, that that your biology clinical team trusts, and also your partners trust.
对吧？
Right?
他们知道这玩意儿真管用。
They know it's it works.
所以他们随时能退回到上一代的可靠方案，但你始终在搞最前沿的那个。
So they can always fall back to something that is validated on previous generation, but you are always working on a frontier one.
我总是从SpaceX身上学到东西。
I always learn from SpaceX.
那是我特别佩服的公司，当然啦，我到处贴着马斯克的照片，每当我觉得自己太拼了，一看那张脸，立马意识到自己还在摸鱼。
So the company that I deeply admire, and, of course, you know, I have Elon's picture everywhere when I whenever I think that I am, you know, working too hard, I'm looking at this picture and I realize that I'm slacking.
你看看他们是怎么造火箭的。
So for like, look at how they are building rockets.
比如猎鹰系列，性能简直没得说。
So you have the Falcon series that works extremely well.
所以你把载荷送入轨道甚至更远，它都能成功。
So you get the payload to the to the to the to to orbit and beyond, and it works.
但与此同时，你还在搞星舰。
But at the same time, you're working on Starship.
对吧？
Right?
而且你已经在研发下一代星舰了。
And you are working on next generation Starship already.
它可能还没那么靠谱。
It might not be as validated.
对吧？
Right?
你不会拿它载人，但你照样得推进星舰项目。
You wouldn't put passengers on it, but you will be working on Starship.
你还在用猎鹰的经验，来优化星舰。
And you are using the learnings from Falcon in order to get a better Starship.
药物研发也是同样的道理，说到底，你得做出超高品质的药。
So that's exactly what we are doing in drug discovery because at the end of the day, you need to develop ultra high quality drugs.
这事儿可一点都不能出错。
This is where you cannot fail.
这就像是载着乘客的太空任务，只不过乘客多了无数倍。
It's kind of like space with passengers, but there are more passengers.
你不能伤害他们，还必须把‘载荷’安全送入轨道。
So you cannot harm them and you need to get the payload to orbit.
为此，你得用大量经过充分验证的模型架构。
So in order to do that, you need to have very validated model architecture with many models.
还得能用这些模型跑模拟实验。
You need to be able to run simulations using those models.
更要让整个行业都信得过这套模型。
You need to be able to ensure that the entire market trust this model.
所以这才需要公开发表成果。
So that's why you also publish.
但你总是在研发下一代星舰。
But you are always developing the next Starship.
一旦星舰成功了，它就成了猎鹰系列的一部分，证明你正在研发下一代星舰。
And once the Starship works, it becomes part of the Falcon series that's validated that you're working on the next Starship.
所以现在我们在英硅智能做的，我更愿意称之为‘制药超智能’。
So what we are doing at Insilico now, I like to use the term pharmaceutical superintelligence.
我的意思是，你只要提需求，系统就能直接给你生成药物。
And by that, I mean prompt to drug.
只要我能说清楚我想要什么，系统就能帮我做出来，然后我再开始验证流程。
So so if I can explain what I want, the system would make it for me and then I will start my validation cycle.
现在大概要花我这么长时间——不过你别拿这话当真，也别引用我这句话。
Currently, takes me around and please don't kind of judge me on that or quote me on that.
这只是一个大概的估算。
That's a ballpark estimate.
我刚才说的这些，你们千万别据此买股票、卖股票，或者吃任何药，拜托了。
Whatever I say now, don't buy, sell, or take any drugs based on what I say, please.
所以通常来说，计算部分搞定一次发射大概要三周左右，对吧？
So usually it takes the computational part takes us about three weeks, right, for any launch.
而要推进到二期，哪怕只是初步看到药物在人体里有点疗效，不搞那些夸大其词的宣传，也得花上五年左右。
And to get to phase two a complete, just to get, like, basic signal of efficacy in humans without, like, big big big claims, it takes you about five years.
这五年里，如果你能把计算时间从三周压缩到一周，甚至一天——我们也在朝这个方向努力——那剩下的实验验证环节，耗时和成本都远远超过计算部分。
So out of this five years, if you can get three weeks to to if you compress it to a week or a day, which we also will do, the rest of the validation cycle, the experimental work that you will need to do is vastly longer and vastly more expensive than the computational part.
所以你最好有个靠谱的系统。
So you better have a validated system.
三周、四周、五周其实都没关系，只要能跑通就行，毕竟这事儿 stakes 太高了。
It doesn't matter if it's three weeks or four weeks or five weeks, right, as long as it works because the stakes are very high.
对。
Yeah.
我有个后续问题。
I I had a follow-up question.
你刚才提到那些模型的时候……
So you were saying, you know, about the models.
关键是要信任这个模型，让它告诉你，在这个细胞系或者这个模型里，到底有没有毒性，或者毒性有多强。
It's about kinda trusting the model, and so kinda model showing you that in this cell line or in this model, you can you know, there's no toxicity or what is the toxicity.
但问题是，我们知道细胞系啊，还有大部分动物模型，其实都不太靠谱。
But, like, how do you deal you know, we know that cell lines as a and most of animal models, know, they're not good.
那如果模型在这种情况下给出预测，你该怎么确保它在临床试验和人体应用上是可靠的呢？
And so how do you like, if your model is predicting kinda in those instances, you know, how how do you ensure, like, for the kind of clinical trials and the and the and the and the human side?
是的。
Yeah.
要证明你的模型真有效，不只是你自己信，还得让别人信，通常你用的不是一个模型。
So the only way to prove to yourself and to the others that your model works and not and, usually, it's not just one model.
而是一组不同模型的组合。
It's an ensemble of different models.
这其实是个完整的流程。
It's a pipeline.
当然，我们现在正把它升级成多模态、跨物种、整合多种组学数据的超强模型，用上Transformer、扩散模型，还有其他好多架构。
Of course, we're turning it into now multimodal, multi species, multi omics, super capable models, transformer diffusion, many other architecture types.
归根结底，你还是得做实验。
At the end of the day, you need to perform the experiments.
你和一个分子的第一次见面，就是看它能不能结合。
And your first date with a molecule starts with binding.
对吧？
Right?
所以你先合成出来，测它的结合能力，这一轮下来大概要花两千五到五千美元。
So you synthesize it, you test it for binding affinity, and that first date will cost you about 2 and a half to $5,000.
对吧？
Right?
差不多得花两到三周时间。
And about, like, two to three weeks.
等你确认这分子确实能好好结合目标蛋白，而且不乱搭其他东西，就把它放进细胞里，看看它能不能干出点活儿，引发预期的反应。
Then you once you you establish that the molecule is a good binder and you test many other properties, so you see if it binds to your molecule protein target and nothing else, then you put it into cells, right, and see if it performs a certain function and results in a certain response in cells.
对吧？
Right?
展开剩余字幕（还有 480 条）
这就算你和这个分子的第二次约会了。
And that's your second date with a molecule.
在真正定下来之前，你得跟它约会好多次呢。
And then you have many dates before you commit to a committed relationship.
接下来你就带它去参加舞会了。
So then you take it to a prom.
对吧？
Right?
然后你把它放进小鼠体内，看看它是不是按你预想的那样被代谢掉。
And you put it into mice and see whether the mice metabolize this molecule the way you predicted it to to metabol to get metabolized.
看看有没有产生什么意外的代谢产物，会不会分解。
You see if it has any secondary metabolites, if it degrades.
再看看它会不会在动物体内跟一些你没想到的东西结合。
You see if it hits something that you don't expect in an animal.
这可就是你的难题了。
And that's your kind of problem.
对吧？
Right?
要走到这一步，平均下来通常得花上几个月，还得搭上几十万美元。
To get there, usually, it takes you about a couple months on average, and it will cost you a few $100,000.
对吧？
Right?
因为在那之前，你得先跟好几批分子“约会”过。
Because before that, you will have to take a few other molecules to dates.
对吧？
Right?
所以一般来说，每个项目我们大概会合成六到两百个分子。
So, usually, we synthesize about 60 to 200 molecules per program.
所以你得实际测试几个。
So you need to basically test a few.
这就像是你想结婚，但所有潜在对象都是选美冠军一样。
Usually, it's kind of like imagine if you want to get married, but all of your potential candidates are, like, miss universe contestants.
那你怎么确定呢？
So how do you know?
对吧？
Right?
所以你得再做久一点的验证。
So you need to do a little bit longer validation.
不好意思啊，用了这个比喻。
I'm sorry for using this analogy.
希望没人觉得不舒服。
I hope nobody is offended.
这还是我第一次公开这么讲。
It's the first time I use it publicly.
你得把这些模型带到PROMs去测试。
And you basically need to take those kind of models to PROMs.
我本人根本不约会。
And I I don't date.
就是觉得挺抱歉的。
It's just sorry.
这只是一个比喻。
It's just an analogy.
然后你得看看它在疾病里能不能行得通。
And then you basically need to see if it works in a disease.
所以你看，这才是最关键的挑战。
So you that there, this is the most important challenge.
关键是要搞清楚，到底该用哪种动物模型，或者人源化的模型，比如类器官，或者体外组织样本。
So basically figuring out which animal model or human human like model, organoid or maybe some ex vivo explant.
对吧？
Right?
你可以直接用真实病人的原代组织来做测试，或者跟有好模型的学术团队合作。
So you take primary tissue from the real patient and test something there or collaborate with an academic that has a really good model.
那到底哪个模型最能代表真实的疾病情况？
So which model is most representative of the real disease?
对吧？
Right?
如果我们对大多数癌症都有这样的模型，那这些癌症早就解决了。
So if we had that for, like, most cancers, we would already solve those cancers.
但目前这些模型没有一个是完美的。
But currently none of those models are perfect.
所以你通常会用两三个甚至四个模型，等证明你的分子在好几个模型里都有效了，才算是下定决心，正式推进。
So you usually do two or three or four and then you one once you demonstrate that your molecule worked in several of those models, you commit to a marriage, so to speak.
对吧？
Right?
这个决定性的选择就叫开发候选药物。
And that marriage is called developmental candidate.
之后，就进入正式的测试阶段了。
So after that, you just have the formal testing.
然后你就得跟着监管机构的要求，按他们的节奏往前走。
So then you move with the speed of traffic with what what regulators want you to do.
接下来就是正式的毒性测试。
So you do the formal toxicity testing.
这叫作IND申报前的研究。
It's called IND enabling studies.
通常是28天的GLP研究，也就是在两种动物身上做符合良好实验室规范的毒性试验。
Usually, that is twenty eight day GLP, so good laboratory practices, toxicity studies in two species.
这一般要花掉三百到三百万美元，还得耗上九到十二个月。
That usually puts you down about 1 and a half to $3,000,000, right, and takes nine to twelve months.
真希望能跳过这一步。
I wish I could have skip this step.
到目前为止，我们还没在这一步栽过跟头，但该走的流程一个都不能少。
So far, we have not failed in those studies, but you have to do that.
然后你才能提交IND申请，正式启动临床试验。
And then you apply for IND to start formal clinical trials.
顺便说一句，这可是你婚姻的第一年。
And by the way, that is your first year of marriage.
对吧？
Right?
你还没孩子吧？
You didn't you still don't have kids.
顺便说一句，我也没成家，还没孩子呢。
And by the way, I'm I'm I'm not a I don't have a family, so I don't have kids yet.
我真不知道那感觉是啥。
I don't know how it feels.
所以你别介意啊，我只是打个比方。
So I please don't be offended making an analogy out of my mind.
从零到确定开发候选物，这过程就像结婚一样，通常得花十三个月左右。
And that process from, like, zero to developmental candidate when you kind of get married usually takes us on, like, typical is around thirteen months now.
最短九个月，最长十八个月。
And shortest was nine months, longest was eighteen months.
至于你问的，我知道我走的是条慢路。
And to answer your question, I know I took a long route.
你怎么知道这些模型真的管用呢？
Like, how do you know that the models worked?
你其实就是慢慢建立起验证数据库。
So you basically build up the validation databases.
对吧？
Right?
你往发育候选阶段推的发射次数越多，就越能证明你的模型是有效的。
So the more launches you have to developmental candidate, the more evidence you generate that your models work.
到目前为止，我们已经做了22次了。
So so far we developed 22.
说白了，我们还是从SpaceX那儿学来的。
It's basically like, again, we learned from SpaceX.
SpaceX简直就是药物研发的最佳范例。
SpaceX is the best example for drug discovery.
你就相当于说：好，我搞了22次轨道发射，成功把载荷送上了天。
So you basically say, okay, well, I did 22 orbital launches and developed payload.
对吧？
Right?
所以现在我知道这个模型是经过验证的。
So now I know that this model is validated.
这些模型里有些还能更新，新的也可以加进去。
And some of those models so I can update, you can include the new ones.
但如果一家AI药物研发公司自称做AI药物研发，却拿不出五到十个来自模型的开发候选物，我是不会信它的，也不会信这家公司。
But if the AI drug discovery company, if they claim to do AI drug discovery, does not show you, let's say, five or 10 developmental candidates out of the model, I would not trust that model, and I would not trust that AI drug discovery company.
你可以看看英矽智能的管线，至少他们有两个项目在二期，或者一个在二期、几个在一期。
You can look at Insilico's pipeline and say, well, you have at least you have two assets at phase two or one asset at phase two and and a couple at phase one.
我们有一个项目正在同时进行两个二期试验。
So we have one program that is in two phase twos.
没错。
Yeah.
还有一个已经完成二期a阶段。
And one completed phase two a.
所以就像传统领域一样，考虑到失败率，能进二期得花上四亿美元左右。
And so kind of like traditional space, it might cost you adjusted for failures, you know, $400,000,000 to get into phase two.
我觉得英矽智能你……
I think Insilico You
你对药物研发可太懂了。
know your drug discovery.
是的。
Yeah.
英矽智能总共融了大概四亿美元。
Insilico has collectively raised around 400,000,000.
我不清楚你们具体总共融了多少，但大概就是……
I don't know exactly how much you've cumulative raised, but that's
快超过十亿美元了。
Just more on half 1,000,000,000.
对。
Yeah.
好的。
Okay.
明白了。
Okay.
所以是五亿美元。
So 500,000,000.
但至少有证据表明，你们已经把某个项目推进到二期了，我猜你们账上肯定还有不少现金。
But there's at least evidence that you've gotten something to phase two with you know, I'm sure you have some cash in the bank.
所以，加上大约四亿美元，还有正在推进的所有IND准备工作，整个管线加起来……
So with around, you know, $400,000,000 plus everything else that's in the pipeline, which is all of the IND enabling.
其实呢，我和内莫经常讨论这个问题，态度一直挺谨慎的。
So there is like, in in some ways, you know, Nemo and I are always talking about and kind of approaching skeptically.
现在AI发展太快了，又特别吸金，全世界的公司都被推着、 incentivized 去说：‘我们用AI了’。
It's like because when AI is moving so quickly and it's so, capital attractive right now that every company in the world, is being urged to and is incentivized to say, yes.
我们也在用AI。
We're using AI.
我们用AI来生产冰茶。
We're using AI to manufacture iced tea.
我们用AI干这干那，反正就是这么回事。
We're using AI to, you know, whatever.
现在任何生意要想融到钱、吸引眼球，都得挂个AI的名头。
Any business needs to have, like, an AI spin in order to raise capital and attract attention.
很多时候，这不过是给老生意加了层AI的糖霜而已。
And, you know, often that's like a layer of AI frosting on top of an existing business.
至少有一点可以肯定，英矽智能确实在更高效地把药物推向市场。
So within like, at least there's some evidence that Insilico is, like, bringing drugs to market more efficiently.
你怎么看那种巨型前沿模型，跟咱们这种拼拼凑凑的小模型之间的区别？
How do you think about kind of, like, the idea of, you know, a frontier monolithic model versus kind of, like, call it bits and piece models that you're piecing together.
你们的平台未来会往哪个方向走？打个比方，就像特斯拉的自动驾驶系统，一开始是机器学习、传统代码、软件2.0和AI组件混在一起，后来慢慢整合了。
And then in in what direction is your is your platform moving to, like, using another, Elon company analogy, you know, their autonomous their robotaxi software, was a whole series of some ML, some traditional coding, some more, like, software two point o or AI driven, components that over time they've consolidated into.
现在还没变成一个单一模型，但比以前更集中了。
It's still not a single model, but it's closer than it was.
快了，快到位了。
It's getting there.
是的。
Yeah.
对。
Yeah.
你们现在在哪儿？我猜整个行业都会朝这个方向走吧。
Where where are you I I assume that's the direction that the industry is going to go.
所以，第一，这真的会发生吗？
So one, is that true?
然后你们现在走到哪一步了？你们怎么判断该投入多少资源去加速这个进程，或者就按部就班地跟着走？
And then where are you on that journey, and how do you even track like, how how to allocate resources towards, like, accelerating that or just kind of pacing along at that advance?
没错。
Yeah.
问得真好。
Very good question.
这个机器人出租车的比喻特别贴切。
So and really good analogy with robotaxi.
因为它的理念和SpaceX非常相似。
So because it's very, very similar idea and also similar to SpaceX.
对吧？
Right?
这就是为什么我们都喜欢埃隆，还从他身上学东西。
Again, that's why we all love Elon and learn from him.
我发明了一个超酷的新概念，完全是我的原创。
So I've invented a really cool concept that is new and that's mine.
我管它叫‘排箫强化学习’。
So I call it the pan flute reinforcement learning.
简单说，就是用很多专门的小模型。
So you basically have many specialized smaller models.
现在我们已经有大约700个了。
Now we have around 700 of them.
我知道每个模型在特定场景下表现都特别出色。
I know that each one of them in certain settings works extremely well.
这就像是一个专攻某个小领域的专家。
So it's it's basically like a small expert in one small domain.
我们经常把它们堆在一起，来完成某个具体任务。
And very often, we just stack them together to achieve like a certain goal.
对吧？
Right?
你可以把这些小模型像乐高积木一样拼起来，也可以根据计算成本给它们排个序。
So basically you create a Lego building out of those small Lego blocks, but you can also take those small models and rank them by their computational cost.
对吧？
Right?
比如这个小模型，它就看看分子结构，说：这玩意儿像药。
You say, okay, well, this one small one, it basically just looks at the molecule and says, it's drug like.
看着就像个药，感觉没问题。
It's it looks like a drug, it looks like it's okay.
对吧？
Right?
所以这可不是垃圾。
So it's not trash.
就像那个‘热狗，不是热狗’的梗。
So basically hot dog, not hot dog.
你还记得《硅谷》那部剧吗？
Remember the Silicon Valley series?
但它的计算成本低到几乎可以忽略不计。
But it's it costs you like peanuts, nothing when it comes to computation.
接下来呢，就得看看这个分子在毒性方面有没有什么隐患，比如这里、这里、这里。
Then the next step would be, oh, does this molecule molecule have any liabilities from the toxicity standpoint here and here and here?
它有多大可能降解？
What's the probability it's gonna that is going to degrade?
对吧？
Right?
代谢稳定性预测怎么样？
What's the metabolic stability prediction?
要处理海量的数据。
Text many, many, many.
所以本质上就像做一支排箫，每加一个新模型，就把它往右放，计算复杂度也就跟着往上堆。
So basically create this pan flute, but every time you're including a new model, you are putting it to the right in terms of computational complexity.
对吧？
Right?
最后呢，你就搞出一堆又复杂又昂贵的模拟。
So at the end, you basically have very complex expensive simulations.
让我最后说一句。
And let me just finish.
想象一下，你有个超级智能的生成模型，训练数据涵盖了蛋白质组学、转录组学、各种组学，还有化学信息。
So now imagine that you have an superintelligent generative model that is trained on all kinds of data proteomics, transcriptomics, omics, chemistry.
我太喜欢这个了。
So I love this.
所以2022年2月，DeepMind发布了一个特别厉害的概念，叫Ghetto。
So in 02/2022, DeepMind published a really powerful concept called ghetto.
很多人没当回事，但其实它是个通用型智能体，一个模型就能干好多事儿。
A lot of people ignored it, but it's a generalist agent, one model that can perform many tasks.
它被忽视的主要原因是，单看每个任务，它表现得并不顶尖。
It was largely ignored because within each task it could not perform extremely well.
对吧？
Right?
你还是得靠专业模型来干精细活，但它能跨任务共享信息。
So you would still rely on specialist models to perform those tasks, but it could it had mutual information.
我当时就说，别干别的了，咱们停一停。
I said, guys, let's drop everything.
咱们就用这个‘排箫’思路，用我们自己那套经过验证的小模型组合，训练出超强的多模态模型。
We're gonna use this pan flute approach to train very capable multimodal models with our own small kind of zoo of validated models.
你就用手上所有的数据，去训练你的生成式模型。
So you train your generative model on all data you've got.
有时候数据可能没那么精良，但我们会去打磨它，教它怎么生成超高品质的分子。
Sometimes maybe it's not super high quality data, but then you massage it, you you teach it how to generate ultra high quality molecules.
对吧？
Right?
所以它其实就是跟那些小型专家配合，来生成类似药物的分子。
So basically it plays with the smaller expert to create drug like molecules.
然后它再去找计算成本更高的专家，学习怎么判断毒性。
Then it goes to a more computationally expensive expert and learns how to generate, how to understand toxicity.
接着它把懂蛋白折叠、对接、亲和力预测、以及具备目标特性的生成能力的模型都融合起来，再推进到三维结构，最后进入临床阶段。
And then it basically distills the model that understands folding, docking, affinity prediction, generation with the desired properties, and then you go to three d, and then you go to clinical.
等这套流程走完，你的多模态模型就变得特别强大，把整个研发流程里的所有专业经验都吸收进去了。
And after these kind of after this process is done, your multimodal model becomes super capable and basically distills the expertise from all your pipeline.
这正是我们目前最专注的方向。
And that is something that we are most focused on right now.
所以你本质上是从猎鹰升级到了星舰，而这个超级模型就是你的星舰。
So you're basically transitioning from basically Falcon to Starship, and that mega model is the Starship.
就是一个能搞定一切的模型，特别全能。
It one model that can do it all, very versatile.
不过现在我们通常把生物学和化学分开处理。
And, usually, now we do it in biology and chemistry separate.
我们有专门做生物的模型来干这活。
So we have biological models that do this.
对吧？
Right?
能力超强，能理解多种生物的生命机制。
So highly capable, that understand life in multiple species.
还有专门懂化学的模型。
And we have chemistry models that understand chemistry.
我们现在正尝试把它们合二为一，虽然大多数时候其实没必要——就搞一个超级智能模型，不再依赖那些小模型。
Now we're playing with the concept of uniting them, even though most of the time you don't need to, into one model that can do superintelligence without the need to rely on smaller models.
但在药物研发流程里，你的药物化学家们通常已经很熟悉，比如说，一些三维结构模型了。
But in a drug discovery cycle, usually your medicinal chemists who are very used to, let's say, some of your three d models.
他们总会回头用这些小模型对假设做一遍计算验证。
They would always go back and validate computationally some of their hypothesis using those smaller models.
对吧？
Right?
毕竟这比真做实验便宜多了，还能让他们心里踏实，知道模型靠谱。
Because it doesn't cost them as much as the real experiment and they get significant kind of peace of mind that that the model is gonna work.
你搞的是最前沿的下一代东西，但同时又有一大堆模型像员工一样随时待命。
So you are working on something that is super frontier next generation, but at the same time, have the zoo of models that can serve as employees.
对吧？
Right?
它们能干特定的活儿，也能当老师，教 bigger 的模型。
So they can perform a certain task or they can serve as teachers for larger models.
而且你知道，这些模型都经过验证了。
And you know that they are validated.
它们经历过好多轮实验打磨，你完全可以信赖，也能放心让它们去训练更大的模型。
They they they they they went through multiple experimental cycles so you can trust them, and you can trust them to train larger models.
所以这就是关键所在。
So that is the trick.
不过，这么想确实是对的。
It's the right it's the right way to think about it, though.
我特别喜欢这个排箫的比喻。
It's like it's almost like I I like the pan flute analogy.
你就像是在训练一个吹排箫的人，而排箫本身是一堆不同的管子。
Like, you know, you have, like, a bunch of just different tubes, you're training some kind of player of those tubes.
但这些管子里头，可能有不少其实是比较传统的机器学习模型。
But some of those tubes probably aren't I I mean, they're probably more traditional machine learning type models that are underneath.
对吧？
Right?
这简直就像工具使用——你有个协调代理，通过强化学习来掌握该用哪个工具，才能达成你想要的结果。
Like, this is it's almost like tool use where you have an orchestration agent that's, you are reinforcement learning about using the right tools to produce kind of, like, the outcome that you need.
你是怎么看待这个问题的？训练时用的数据组合又是怎样的？
How do you think about and what is the, like, data mix that's going into that training?
我们该怎么看待所需数据的丰富性？哪些是可能相关的其他数据源？还有，内部实验产生的数据，跟患者样本这类外部数据，该怎么区分和搭配使用？
How do we think about kind of the richness of data that you need, kind of the potential orthogonal data sources that you need to tap, and kind of, like, the, yeah, the data off internal experimentation versus data collected from, I don't know, patient samples or anything else?
说到底，我们该怎么理解这个问题？
Like, how do we think about that?
现在训练和测试用的绝大部分数据，其实都来自已经公开的实验成果。
So the vast amount of data that is used for training nowadays and also for testing is publicly available data from experiments that have already been published.
说来你可能不信，我刚开始做AI药物发现的时候，是靠自动分析美国的政府科研基金，后来扩展到全球。
And I kind of started cutting my teeth in AI powered drug discovery by, believe it or not, automatically analyzing government grants in The US, right, and then around the world.
我们追踪全球的科研经费，总额大约有2.5万亿美元。
So we track funding globally and we track about 2.2, $2,500,000,000,000 worth of funding.
然后我们会清洗这些数据，可能筛掉一半以上，最后剩下大概1万亿美元左右。
And then you kind of clean up this data, right, so you maybe bring it down to like, $1,000,000,000,000 or less.
接着把这些数据分门别类，比如生物、化学、文本、临床数据，再进一步细化，最后请领域专家来把关。
So you cut maybe, like, more than 50%, and then you put it into buckets.
再把数据分门别类，比如生物、化学、文本、临床数据，再进一步细化，最后请领域专家来把关。
So biology, chemistry, text, clinical data, further refine it, apply apply human experts.
我们早期，也就是2014到2019年，招了很多博士生、硕士生，专门清理数据、人工标注，有时候还得给相似的论文打分。
So in our early days, so 2014 to 2019, we hired many, many, many, like, PhD students, master's students that cleaned up the data, annotated you with the data using human intelligence, sometimes kind of scored these similar papers.
比如两篇讲同一个蛋白的论文，讲的完全是两码事。
So there could be two different papers published on the same protein telling different stories.
对吧？
Right?
因为实验结果不一样。
Because experimental results were different.
那你觉得该信哪个？
Which one do you trust?
所以第一步，得先把这些矛盾的地方挑出来。
So first of all, you need to kind of highlight those conflicts.
有时候，同一个蛋白在两个实验里表现得完全相反。
And sometimes, again, the protein moves in a completely different direction in two different experiments.
而且，实验条件其实差不多。
Right, in similar experimental conditions.
所以你得靠专家来验证这些实验数据。
So you need experts to validate this experimental data.
有时候他们做得不够好，但至少能把问题点出来。
Sometimes they don't do a good job, but they highlight the problems.
然后你把公开的文本数据跟公开的组学、化学数据库关联起来，拿这些数据来训练、清洗。
And then you link the publicly available text data to publicly available omics repositories, chemistry repositories, and train on that, clean it up.
我觉得，如果你连公开数据都用不好，还证明不了能从这些数据里做出可靠的PCC，那你根本算不上AI。
And that is the main so I actually think that if you cannot train on public data and demonstrate that you can deliver a PCC from from that data, you should you should not be an AI.
你就是个传统生物技术公司。
You are just a traditional biotech.
对吧？
Right?
所以要是缺数据，通常都是个不错的借口。
So if you need some additional data, usually it's a good excuse.
我到底缺了啥？
Oh, what am I lacking?
哦，我缺的是一台贵重设备，好让我拿去给投资人看看。
Oh, I'm lacking an expensive machine that I, you know, I I will show show to my investors.
这才是我跟别人不一样的地方。
That's how I'm different.
不是的。
No.
你看ChatGPT、Grok这些，它们到底有什么独家数据？
Like, look, chat GPT, grok, grok, like, what kind of unique data do they have?
对吧？
Right?
我的意思是，Twitter，也就是X，还有其他好多来源，都是公开的。
I mean, Twitter I mean, X and many other sources is publicly available.
只要你愿意，其实也能把这些数据收集起来，清理干净，训练出一个能理解宇宙的超级智能模型——至少在某种程度上。
So if you wanted to, you could actually collect that as well and clean it up and create a superintelligent model that understands the universe, right, to some extent.
再说了，你自己想想，你也没读过世界上所有的书吧。
And if you think about yourself even, right, so you didn't read all the books in the world.
对吧？
Right?
你只需要读那些好书，掌握关键流程的基本原理就行。
You just need to read the good books and understand the first principles of certain of certain processes.
这正是我们正在做的。
So that's what we are doing.
我们用的是公开数据，但后期你得有自己的黄金数据集，自己生成的，用来验证这些公开数据到底靠不靠谱，对吧？
We're using publicly available data, but later you need to have some golden datasets that you trust, that you generate at yourself in order to validate whether this publicly available data is good or not, Right?
也用来验证你的模型有没有效果。
And whether your model works.
而我们有个超大的自动化实验室，六个全自动联动的房间，24小时不停运转，几乎不用人工干预，专门生成验证数据。
And that's where we have a massive automated laboratory, six interconnected, fully automated rooms working round the clock with minimal human intervention that are generating validation data.
对吧？
Right?
不过，我再强调一下，这算是个额外的加分项。
But, again, I just want to emphasize this as a bonus.
对吧？
Right?
所以这个实验室能帮你发现新靶点、拓展适应症、做毒性分析，还能生成高质量数据，快速验证你的模型是否有效。
So that's that that that lab can help you identify new targets, do indication expansion, do toxicity analysis, generate really good quality data to see if your model worked, but and faster.
但跟公开数据比起来，那简直是天差地别。
But compared to the publicly available data, it's like, alright.
你得先打基础。
So you need to do the groundwork.
你得亲手打磨数据。
You need to grind the data.
如果你连公开数据都懒得好好处理，那根本没资格做AI药物发现公司。
And if you didn't grind the data available to you publicly, you should not exist as an AI drug discovery company.
我觉得反驳意见可能是这样：想象有一条河，你得过河。
I think the argument would be that that like, imagine that, there's a river that you need to cross.
第一个建桥过河的人价值巨大，而且只要你先过去了，别人就基本没动力再重复这事儿了。
You know, being the first person to build the bridge across the river is extraordinarily valuable and actually dissuades the second person from crossing the the river if you've crossed soon enough.
对吧？
Right?
所以啊，就算你知道怎么建桥，也得真动手去建，还得根据河床的地貌调整方案，桥才能稳稳建起来。
And so, the the, you know and so even knowing how to bridge build, like, you still need to build the bridge, and and kind of you still need to adapt to the terrain of the river bottom to successfully build the bridge.
还有啊，你最后用强化学习去优化你的引擎，让它能把药真正推到市场——这其实才是公司真正的护城河。
And, you know, and, like, the the the final, like, bit, the reinforcement learning that you're doing to your to your engine, that allows it to get drugs to market, you know, potentially then provides the protective mechanism for the business.
对吧？
Right?
或者你想想，要是你或者别人打算开源，但外面一堆公司都在喊：‘我们搞了个完整的生物模型，谁都能用’，那你怎么看未来的竞争优势？
Or or, like, how do you think about, yeah, competitive edge into the future if kind of you or others you've talked about open sourcing, but, you know, there's plenty of companies that are out there being like, hey.
我们打算推出一个功能齐全的生物模型，人人都能免费用。
We're gonna deliver, like, a fully featured biological model that everybody can use.
在这种情况下，英矽智能怎么保持领先？或者任何一家AI药物发现公司，到底该怎么打出自己的差异化优势？
How does Insilico stay ahead of the curve within that context, or how should any kind of AI drug discovery company try to, like, carve out competitive edge?
说到竞争和学习对象，我们只看 SpaceX。
So when we are looking at competition and also where to learn from, again, we learn only from SpaceX.
对吧？
Right?
所以我们本质上是想为药物研发打造一个太空探索版的SpaceX。
So we basically want to build SpaceX for drugs.
正如我之前说的，最终的考验是：你的药在临床中真有效吗？
And there, as I've explained already, the ultimate test is does your drug work in a clinic?
它真的能帮到病人吗？
Does it help the patient?
这时候你得在创新性、把握性、商业可行性之间找平衡，还有其他好多因素，但主要就这三点。
And there you're balancing novelty, confidence, commercial tractability, and many other properties, but mostly those three.
你得不断证明，你能从零开始开发出一个临床前候选药物，对吧？
And you need to demonstrate consistently that you can develop a pre clinical candidate, right, from zero.
不管靶点的新颖程度高低，都要能稳定地做到这一点。
So with various kind of levels of novelty of the targets and show consistency there.
对吧？
Right?
所以你得证明自己能把有效载荷送入轨道。
So you can you need to show that you can deliver payload to orbit.
为此，我们制定了一套评估标准。
So we developed a set of benchmarks for that.
过去五年，我们成功推出了22个开发候选药物，每一个都清清楚楚、有据可查。
We say we we we we like, last five years, 22 developmental candidates, clear clear cut.
其中很多成果都发表在《自然·生物技术》这样的顶级同行评审期刊上。
And many of those stories were published in top peer reviewed journals, like Nature by Technology.
我举两个从零到开发候选药物的例子。
I have two examples from zero to Developmental Candidate.
这些例子会告诉你整个过程是怎么做的，还有所有的实验数据。
It shows you everything how you do it and shows you the experimental data.
在这个行业里，赢得领先地位的唯一办法就是更快、更便宜、成功率更高。
So the leader the the way to achieve leadership in this industry is to just be faster, cheaper, with higher probability of success.
其实大多数公司PPT里都这么写，但真能拿出一个临床前候选药物的，没几家。
Actually, that's what most of the companies kind of have have have in their decks, but most did not deliver a single preclinical candidate.
所以那些跟SpaceX竞争的公司啊，什么 blah blah blah 的。
So it's kind of like folks that are competing with SpaceX, like blah blah blah.
听好了，我们真有火箭。
Look, we have a rocket.
哦，我们的发动机更厉害。
Oh, we have a better engine.
我们燃料更好，模拟系统也更优。
We have a better better fuel, a better simulation.
我们的AI也更先进。
We have better AI.
或者整个国家都在支持这项事业。
Or we have the entire country supporting this effort.
对吧？
Right?
然后你看看他们这些年到底发射了多少次，再对比SpaceX，你就明白这行是怎么回事了。
And then you look at how many did they launch over time and how many SpaceX launched and you you you understand the industry.
对吧？
Right?
所以如果你是个投资人，或者制药公司，或者员工，想评估哪家公司更厉害，现在唯一能看的指标就是：好吧。
So if you if if you are an investor, if you are a pharma company, if you are an employee, and if you are looking at, you know, how to rank the companies, the only way to rank them now is to say, okay.
那你们平均多久能出一个研发候选药物？
Well, how many like, what's your average time to developmental candidate?
因为这才是真正省时间的地方。
Because that's where you're saving.
对吧？
Right?
其他方面，你 basically 就跟堵车一样，慢悠悠地挪。
The rest rest, you are basically moving with the speed of traffic.
我完全明白你的意思。
I understand exactly what you're saying.
这就像是我们根本不知道那些火箭到底有没有进轨道。
And it's kind of like we don't know if the rockets have made it to orbit.
对吧？
Right?
这就像是，你其实是在说，应该根据你发射了多少次来评判，这确实有道理——至少你从每次发射中学到了东西。
It's almost like, you know, you there in some ways, you're arguing you should be judged on how many you launch, and there is validity to that and that kind of, like, the launch velocity is some at the very least, you're learning something from it.
而且，所有东西都在往上升，或者都在发射模式里，但我们根本不知道有多少真正进了轨道。
And, kind of, like, everything is going up, or or or in in launch mode, but we don't know how many have successfully made it to the orbit.
就像你之前提到的，制药公司通常不太愿意去碰全新靶点，因为他们指望学术界先做。
Like you mentioned earlier, that that kind of, like, pharma is not doesn't really like to basically buy novel targets because they rely on academia to do it.
在他们看来，这离商业化太远了。
Because from their perspective, it's like, that's so far away from commercialization.
他们怎么知道这个新靶点真的有用呢？
How are they gonna know this novel target's actually in a target that's useful to do?
所以他们宁愿让别人去冒这个险，或者让学术机构先把风险降低，他们才愿意进场。
They'll let somebody else take on that risk or an academic derisk it before they even get involved.
同样地，我看看英矽智能的管线，确实看到一堆项目都进入IND准备阶段了，但从外部来看，我其实搞不清楚具体情况。
And similarly here, it's kind of, you know, I can look at Insilico's pipeline and be like, well, yes, there's a lot that is, you know, in IND enabling studies, but I don't know from an outside perspective.
我怎么知道这些真的能解决那个问题呢？
How do I know those are, like, actually, you know, things that are gonna answer that.
我有个特别好的答案给你。
I have a great answer for you.
所以呢，
So for
要做到这一点，你得先把几个资产卖出去。
to do that, you need to sell a few of those assets.
一旦你走到开发候选阶段，要是你的质量特别棒，做得特别到位，尤其是在中等或低新颖性靶点领域，其实是有市场的。
So once you get, let's say, to Developmental Candidate, if your level of quality is amazing and if you did a really, really good job, usually, especially in the kind of moderate novelty target space or lower novelty target space, there is a market for them.
我就卖过几个，也授权出去过。
So I've sold I've licensed out a few.
我明明白白地授权出去了三个。
So I basically licensed out in a very clear cut way three.
我最出名的一个是，我把药推到了IND阶段，然后卖给了Exelixis公司——那是家特别硬核的药物化学公司，自己研发药物，靠他们推出的药年收入几十亿美元。
So my most famous one was I took the drug to IND and it sold, it got licensed to a company called Exelixis, one of the most hardcore medicinal chemistry companies that are develop discovering drugs themselves, they have, you know, billions of dollars in revenue from the drug they discovered and put to the market.
所以现在他们不仅自己研发，还从别人手里收购。
So now they are discovering themselves and also buying from others.
他们直接付了我们八千万美元，买下了一个刚达到IND阶段、才进入一期临床的分子，这可是一笔十亿美元的交易。
So they paid us $80,000,000 cash upfront for a molecule that just touched IND, touched phase one, and it's a, you know, a billion dollar deal.
而且我们还能拿分成。
Plus we have royalties.
另外，我们还把两个更早期的、处于临床前候选阶段的项目授权出去了，对吧？
And then we've licensed two more at earlier stages at preclinical candidate, right?
其中一个刚宣布，预付款就达两千万美元。
One just got announced $20,000,000 upfront.
这说明我们这些项目的质量确实是实打实的。
So it means that the level of quality of those launches are real.
对吧？
Right?
所以总的来说
So basically
这种对话是怎么展开的？
What is that conversation like?
通常来说，这么早期的药物很难对外授权。
Because usually it's hard to license out drugs at that earliest stage.
对吧？
Right?
我的意思是，从大药企的角度做净现值分析，一期临床的项目几乎不值钱。
I mean, you typically like, if you do the NPV analysis from a big pharma perspective, phase one is actually worth basically zero.
甚至可能是负值，毕竟从一期到上市的成功率太低，开发成本又高。
It might be a net cost, you know, given their rate of throughput from phase one to commercialization, the cost to bring something to commercialization.
当然，具体还得看你要进入的是什么市场之类的。
Again, depending on what, you know, market you're selling into, etcetera.
但总的来说，大药企其实就想躲开这些风险。
But net, you know, big pharma, in some ways, wants to just avoid all of that risk.
那你们是怎么跟对方谈成这笔交易的？
So how does that conversation happen between you and a license or where actually you manage to sell it?
你们是用一种能直接带来现金收益的方式把它卖出去的吗？
You manage to sell it in a, you know, cash accretive way?
他们到底看中了什么，才这么被吸引？
What is what is what are they seeing that that really, you know, attracts them?
所以我说过，创新程度得适中，甚至偏低一点。
So usually, that's why I said that the level of novelty must be moderate or moderate low.
我们在对外授权的时候，就瞄准那种有一两个领先者的领域。
So we we aim to be in a space when we are licensing out, we'll aim to be in a space where there is maybe one or two front runners.
不是人山人海的竞争，但大家都知道这个靶点。
So it's not a crowded space, but people know the target.
时间久了，他们对这个靶点自然就建立起信任了。
So they usually establish some trust in that target over time.
对吧？
Right?
顺便说一句，我们这些靶点都不是全新的。
So none of the targets are new, by the way.
所以整个行业都在围着三四十个目标打转。
So the entire industry is dancing around, like, thirty, forty targets.
这挺可笑的。
That is funny.
其他所有东西都不过是老套路的变种。
And everything else is kind of the variation on the on the old.
我们不喜欢玩这种游戏。
So we don't like that game.
我们更愿意进入另外三百个目标的领域，甚至更多，当我们对外授权时，那时候比的是质量。
We like to be in this, you know, additional 300 target space or maybe more when we are out licensing, and there you're competing on quality.
所以我们还得去向日本和德国的生物制药公司学习，怎么搞数据室。
So we even had to go and learn from, like, Japanese biopharma, from German biopharma how to do the data rooms.
对吧？
Right?
因为你的项目质量，得比药企自己内部习惯的水平更高才行。
Because the quality of your asset has to be higher than the pharmaceutical companies got used to internally.
因为他们内部可以这么说：‘这个实验我们没做，但这个假设有道理，觉得这个分子能成，那就往前推进吧。’
Because internally they can say, oh, well, we didn't do this experiment, but we trust this hypothesis and we think that the molecule is gonna work, that's why we're gonna progress it forward.
可一旦别人要买，尽职调查的严格程度可就高多了，对吧？
When they're buying, the level of level of due diligence is gonna be much higher, right?
因为买家通常有好几个选择，他们肯定想用合适的价格拿到最好的那个。
Because if they usually have a few choices and they want to get the ultimately best one for the good price.
而且价格谈判反而是最后才谈的事。
And the price negotiations is actually comes last.
对吧？
Right?
他们更看重你的数据室质量，价格反倒没那么重要。
So they don't care about the price as much as they care about the quality of your data room.
所以你要是做分子打算对外授权，就得确保它比那种‘我这儿有几篇论文’的水平高得多。
So when you are making molecules for sale for licensing out, you need to make them sure that they are better than the level of, you know, I have a few papers here.
对吧？
Right?
所以，通过技术手段，外界能从论文里看出你们数据室的质量。
So nature by technology can see the quality of our data rooms in some papers.
因此，当你在对外授权时，标准得比通过六位审稿人同行评审的要求还高。
So internally when you're licensing out, it has to be higher than the level necessary to pass the peer review with six reviewers.
对吧？
Right?
因为那时候会有三十位审稿人，个个都想把项目毙掉。
Because there you're gonna have 30 reviewers and every one of them wants to kill the program.
对吧？
Right?
所以数据必须做到绝对完美。
So it needs to be absolutely pristine.
所以我才说，这就像你正在发射火箭，而且还是给美国政府机构发射——他们可是要押上几十亿美元，赌一个极其昂贵的卫星。
That's why I'm saying that it's basically like imagine that you are doing a space launch, but you are also doing it to a US government agency that is going to bet, like billions of dollars on some very expensive satellite.
对吧？
Right?
所以他们得确保你真能走到最后。
So they need to make sure that you are gonna get there.
这正是你在对外授权时正在经历的事。
And that's what's happening when you're out licensing.
因此你必须保证数据质量绝对过硬。
So you need to ensure that the level of quality is absolutely great.
如果你在搞一些前所未有的全新靶点——我正好有一个，最近还有一篇《自然·医学》论文刚发表。
And if you are running with absolutely novel targets, right, that have never been in the clinic before, and I have one like that and here is, you know, a nature medicine paper that just came out.
你可以去看看临床结果长什么样，数据质量如何，我们是完全公开的。
You can read up on how the clinical readouts look like and the quality of that data and we disclose everything.
一旦你瞄准了全新靶点，就得做更多工作，因为药企一般想等你做完一期，最好能到二期再跟你合作，那时候大家都想抢这个项目，但接着你就面临另一个问题。
So once you are going after a novel target, you would need to do much more work, right, because pharma would want to partner with you after, let's say phase one or preferably phase two, so then they line up, they all want to the asset, but then you are facing a different problem.
你可能会想：嘿，我手头现金够用，还能再往前推一推，要是再推进一点，就能拿到更多患者数据，结果也会更扎实。
You're like, okay, well, I'm actually, I have enough cash to progress it even further and if I progress it even further, I'm gonna get much more valuable data from the patients and from the study.
我会把实验做到底，要是真成了，那回报远超我之前投进去的每一分钱。
I'm gonna validate to the end and if it works, it's billions of dollars I pay for the entire party many times over.
对吧？
Right?
重点来了。
And here is the punchline.
所以我的工作和梦想——再说一遍，这不是吹牛——就是打造一家能赚钱的生物科技公司。
So my job and my dream, so again, no claims, is to develop a profitable biotech.
如果你看看我们是怎么运作的，说实话，我们有时候已经快做到了，因为我们的收入……
So if you look at how we operate, well, I can say that we're sometimes we're almost there because our revenue
这想法真新颖。
is What a novel idea.
一家能产生现金流的公司。
A company that generates cash flow.
我简直不敢相信你说的。
I can't even believe you.
因为大家都习惯了亏钱。
So because people got used to losing money.
对吧？
Right?
所以我的工作，就像SpaceX的火星任务一样，就是要攻克衰老问题。
So my job, my kind of Mars mission for SpaceX is solving aging.
因为只要解决了衰老，其他所有问题就都能迎刃而解。
Because if you solve aging, you solve everything else.
你有足够的时间去搞定其他所有事。
You got enough time to solve everything else.
如果我们真能做到，整个社会都会被改变。
And we change the society if we if we do that.
但到目前为止，想攻克衰老这个领域，根本就没有一家成功的公司。
So but to solve aging, so far, there is zero zero successful companies in that field.
对吧？
Right?
所以目前没有任何药物在实验中真正有效。
So no drug works experimentally.
明白了。
Okay.
但我觉得这其实是个语义问题，现在咱们得聊聊衰老或者长寿了——毕竟，说到底，几乎所有疾病都跟衰老脱不了干系。
But see, this seems like a a semantic issue to me because and so now we're gonna branch into the aging bit or the longevity bit because isn't I mean, every disease is a disease of aging to some degree.
对吧？
Right?
比如心血管病，本质上就是衰老引起的。
I mean, cardiovascular disease is a disease of aging.
但艾滋病可不是衰老导致的。
Well, HIV is not a disease of aging.
对吧？
Right?
所以艾滋病就跟其他病一样，可能随时找上你。
So HIV can be like any disease
不过话说回来，这其实是衰老加上环境因素共同作用的结果。
Well, look, I mean, but it's aging plus environment.
我只是刚好接触到了错误的环境因素，比如那个病毒。
I just happen to be exposed to the wrong environmental variable, like that virus.
你懂的。
You know?
我体内其实有各种各样的病毒，但大多数都被我年轻健康的免疫系统压着，没啥事。
I mean, I have all kinds of viruses swimming around in me, and, you know, most of them, my body's, like, keeping under check because I'm, like, have a young, healthy immune system.
但等我免疫系统老了、退化了，那些病毒就变成要命的东西了。
But, like, as my immune system gets, like, older and more degraded, then those viruses become, like, deadly viruses to me.
我觉得癌症至少有一部分就是这么回事，或者某些癌症是这样。
And that's kinda at least some of what's going on with cancer, I think, or some cancers.
不是所有病都跟衰老有关。
Well, not every disease is a disease of aging.
抱歉。
Sorry.
我得纠正你一下，很多儿科疾病，尤其是那些特别痛苦的，其实是出生前就有的基因突变导致的。
I have to correct you because, like, a variety or, like, a huge variety of, like, pediatric diseases and some of the pretty painful ones are caused by, you know, the mutations that you get before birth.
我觉得咱们说的主要是那些更复杂的常见病。
I think, you know, we're just referring to kinda kinda more complex common diseases.
比如癌症，还有阿尔茨海默病、帕金森病这些。
So, you know, cancer on one side, like, our Alzheimer's, Parkinson's.
其实可能就是神经元在经历正常的衰老过程。
It could be just kinda the aging process actually happening in neurons.
这本身也是一种衰老相关的疾病。
You know, it's and it's it's the kinda aging disease as well.
还有心血管和代谢类的疾病也是一样。
And then, like, a cardiometabolic as well.
这就像你一辈子在用身体，总会有损耗。
It's like a pair of, you know, of using your organism during the years.
对吧？
Right?
是的。
Yeah.
我给你举个具体的例子吧。
Let me let me give you, like, a very tangible or specific example.
明白吧？
Right?
现在有几家基因编辑公司，正在瞄准一个能降低坏胆固醇的靶点。
There's, you know, there's a couple of gene editing companies going after a a target that reduces basically bad cholesterol.
那靶点叫啥名我一时想不起来了。
I can't remember the name of the target.
对吧？
Right?
而且——
And
是PCSK9。
there's PCSK9.
对。
Yeah.
对。
Yeah.
嗯，虽然不是那个，但另一个例子也差不多。
Well, not p there's another one, but, yes, that's an example of it as well.
你可以想象，如果这能成功，把这种基因突变推广到整个人群，或许能大幅降低心血管疾病的发生率。
You could imagine that, you know, depending on how this works out, maybe modifying that putting that mutation into the broader population actually reduces cardiovascular disease to a very broad extent.
这跟直接对抗衰老是两回事，还是说你认为它们其实是一条连续谱——我们只是在解决身体随着年龄增长自然出现的那些问题？
Is that separate and distinct from kind of, like, going after aging directly, or do you think of those as, like, on a continuum of we're just basically, like, solving kind of these problems that happen to bodies as they age?
这些算是不同的类别吗？
And, are these different categories of thing?
其实呢，它们既有相同的地方，也有不同的地方。
So it's it's kind of the same and different.
你提到的这个话题，其实非常关键。
So you touched on a very, very sensitive topic.
对吧？
Right?
比如PCSK9，这就是个绝佳的例子。
Because some things like, you know, PCSK nine, beautiful example.
对吧？
Right?
它很可能是个延寿靶点。
Most likely, it's a longevity target.
关键是，它能帮到很多高胆固醇的人。
And the thing is it's going to help a lot of people with bad cholesterol.
但对胆固醇正常的人有用吗？
But is it going to help people with good cholesterol?
问得好。
Good question.
对吧？
Right?
所以在人群层面，它可能像GLP-1一样有帮助。
So on a population level, it may help, like GLP ones, for example.
对吧？
Right?
GLP-1类药现在可火了。
GLP one drugs are very popular nowadays.
根本不用解释这是啥。
You don't need to explain what it is.
在我看来，它们会大大提升美国人的预期寿命。
They are going to significantly improve, in my opinion, on life expense expectancy in The US.
对吧？
Right?
因为这儿肥胖人口比例实在太高了。
Because here the percentage of the population that is obese is pretty is pretty high.
但对你我来说有用吗？
But is it going to help you and me?
大概没啥用。
Probably not.
对吧？
Right?
因为它会导致肌肉流失。
Because it causes muscle wasting.
还会增加患年龄相关性黄斑变性的风险。
It increases the chances of age related macular degeneration.
我们连所有副作用都还没摸清楚呢。
We don't even know all the side effects.
对吧？
Right?
顺便说一句，你的心脏也是肌肉。
And by the way, your heart is a muscle.
所以对你我这样的人来说，看你的体脂组成，我也不清楚你外面长什么样。
So probably for you and I, looking at your body composition, I don't know how you look like outside.
我就是开个玩笑。
I'm making a joke.
但很可能，这对你根本没用。
But most likely, it's not gonna help you.
对吧？
Right?
你只会掉肌肉。
It's only going to you're gonna lose muscle.
所以我们是在找那些也能帮到你的目标，对吧？
So we are looking for targets that would help you as well, right?
我们的想法是找出推动某些衰老过程的根本机制。
So the idea is to identify the fundamental mechanisms that drive certain aging processes.
现在我们用的叫作‘衰老特征评估’，对吧？
Currently what we have is called hallmarks of aging assessment, right?
有一些特定的生物学类别，既跟衰老有关，又在推动衰老。
So there are certain buckets, certain types of biology that are at the same time associated with aging and driving aging.
大概十五年前，老年学领域提出了一套框架，用来分类不同的衰老机制。
There is a framework in by gerontology that has been proposed only about fifteen years ago that kind of classifies the different aging mechanisms.
我们通常的做法是，利用所谓的‘衰老时钟’来识别蛋白质靶点。
And what we usually do, we identify protein targets utilizing what we call aging clocks.
有一种针对人类和动物衰老的模型，能发现某些蛋白质特别相关——它们对预测你的实际年龄或生物年龄很有参考价值。
So there's a model of human and animal aging where we see that the protein is specifically implicated in well, it's it's it's being very it has some predictive value in predicting your chronological age or biological age.
然后我们会把这些和衰老时钟相关的蛋白质，对应到那些衰老标志上。
And then we also look map those proteins that are implicated in aging clocks onto those hallmarks of aging.
看看这些蛋白质到底涉及多少种衰老机制。
So we see in how many buckets of aging this protein plays.
我们还会把它们和各种疾病关联起来。
And then we also map it onto diseases.
我们现在有三千多种与年龄相关的疾病，会问：这个靶点最可能在哪些病上起作用？
So we say, okay, well now we have three thousand age related diseases, where is it likely to work this target, right?
所以你本质上是在找一把‘瑞士军刀’，希望它能覆盖尽可能多的场景。
So you're basically looking for a Swiss army knife that it may work in as many as possible.
然后你会问：哪一种疾病和衰老的关系最紧密？
And then you say, okay, well, which disease is most most age related?
对吧？
Right?
所以我们其实能很快地测试这种药，对吧？
So we can actually test this drug quickly, right?
因为那些病人往往活不了多久。
Where people die very quickly.
我特别关注的一个疾病领域是特发性肺纤维化。
So one of my favorite disease areas is idiopathic pulmonary fibrosis.
这病一般发生在晚年，患者平均年龄在六十五六十八岁左右，本身就属于高龄群体。
It hits late in life, so usually average age of your typical patient would be 65, 68, and so they're reasonably on the older side.
所以他们剩下的寿命不多，生活质量也大打折扣。
And the so they don't have a lot of life to live and the quality of life is diminished.
这样一来，临床试验能跑得快一点，还能顺便收集蛋白质组生物标志物，看看你的药有没有抗纤维化的效果，同时用同一套数据判断它是不是真能逆转衰老时钟。
So you can run clinical trials a little bit faster and also collect, let's say proteomic biomarkers to study the anti, like, fibrotic signature as an anti fibrotic effect of your drug and at the same time, use the same data to see if your, drug is reversing the aging clocks.
我觉得身体里那些东西，虽然我仍认为是衰老的副产物，但它们在全身范围内其实有共性。
Imagine there are, like, a number of I would still argue they are artifacts of aging that are appearing in the body, but they're common across the body in some way.
而且因为这些问题在全身普遍存在，当某个部位的退化达到一定程度，就会引发特定疾病。
And then because they are happening common across the body, there are certain diseases that are triggered by enough degradation in that area.
比如特发性肺纤维化。
So idiopathic fibrosis of the lungs.
肺纤维化就是一种，当身体里的某个机制被触发，情况就会急转直下，变得非常严重。
Pulmonary fibroid fibrosis is one that it basically, like, crosses a tipping point in people's bodies where something that turns on, it's really bad.
所以如果你能瞄准这个你认为根本性的过程，
So then if you can target kind of you're you're really targeting what you think is that underlying process.
那你不仅能延缓原本由衰老驱动的疾病爆发，还能同时评估：你是否真的能减缓，甚至逆转这个衰老进程。
And so then you're putting off this disease that otherwise would be triggered by the kind of, like, aging vector, and simultaneously, you're getting a measure of whether or not you can flatten or hopefully even, like, turn around the aging vector and and and turn it in reverse.
这想法真够大胆的。
Provocative.
你不觉得人体其实是由一堆各自独立演化、却又彼此咬合的系统组成的吗？这些系统很可能都演化出了差不多的故障率。
And don't you think that it's like the body is just composed of a, like, a a whole series in of, independently evolved kind of interlocking systems that have probably evolved to, like, a similar kinda fail rate.
所以，真会有一个共同的机制在推动衰老时钟往前走吗？还是说——
And so that there's not you know, it's kinda like you is it really likely that there's a shared common mechanism that is, like, kind of, like, driving the clock forward rather than, hey.
就像你说的，这个系统在某个故障率下会出问题，那个系统也是在某个故障率下崩溃。
It's like, you know, this system breaks down at a certain, like, fail rate, this other system breaks down at a certain fail rate.
所以你解决了其中一个，但其他问题还是拖着你的后腿。
And so you'd solve one, but you still have the others holding you back.
其实我完全同意你的看法。
So I agree with you actually here 100%.
这正是衰老发生的方式。
So it is how aging transpires.
对吧？
Right?
你的身体里头，其实有好多好多不同的时钟，各自在不同的部位、不同的组织层级上，以不同的速度滴答走着。
So you basically have many, many different clocks ticking at different rates in different parts of your body at different level of organization.
对吧？
Right?
从分子层面、细胞层面，到系统层面、器官层面，甚至心理层面，都是如此。
At the molecular level, at the cellular level, at the systems level, at the organ level, even at the psychological level, by the way.
对吧？
Right?
信不信由你，现在的你和13岁时相比，很可能已经完全是另一个人了。
So believe it or not, you are a different person now than you were, let's say, when you were 13, most likely.
顺便说一句，我确实发表过几篇相关论文。
And by the way, I did I I I I I published a couple papers on that.
我真开发过一个深度学习模型，用来追踪心理年龄的变化，而这种变化其实是可以逆转的。
Like, I really did develop a deep learning model that tracks psychological aging, and that's something that you can reverse.
我给你们所有听众一个建议：如果你想在听这个播客的时候变年轻，那就坚定地相信自己一定能活到120岁。
I can basically can for all your listeners, if you wanna get younger while you're on this podcast, just religiously convince yourself that you are going to live to a 120.
做得到，完全做得到，保持乐观，这样你要是60岁，就已经走了一半的路了。
Achievable, achievable, optimistic, and suddenly you are, you know, if you are 60, you're halfway there.
这真的会彻底改变你的人生。
So and and it it will change your life.
对吧？
Right?
你会表现得像个更年轻的人。
You'll you'll you'll act act as a younger individual.
但别难过，很多人虽然靠这种方式逆转了年龄，却没变得更开心，因为路还长着呢。
But don't feel sad if a lot of people who reverse their age this way, they kind of are not as happy because there is a long way to go.
不过，等等，抱歉。
And but here here sorry.
不好意思。
Sorry.
说正经的，当你想找那些能影响、从而减缓或逆转衰老的共同机制时，我觉得确实存在一些。
Jokes aside, so when you are trying to find this common ground, common mechanism that you can influence in order to slow down or reverse some aging processes, I think there there are some.
我相信是有一些共同机制的。
I I believe that there are some mechanisms that are common.
不过它们可能不会带来特别惊人的效果。
They are likely not to give you super dramatic benefit.
对吧？
Right?
所以你可能能多活个两三年。
So you might be able to get, like, two to three years.
如果你用刚才那个比喻的话——
So some like, if if if you are using kind of your analogy.
对吧？
Right?
咱们来打个比方，就当是辆车。
So imagine that we're talking about the car.
如果你把刹车系统升级一下，或者加固那些容易坏的部件、修好损伤，哪怕其他部分比如发动机一直没好好保养，
So if you improve the improve your brakes, right, or improve something that is very easily breakable by reinforcing the material or fixing the damage, maybe the rest of your system, your engine can go for and not properly maintained.
这车也能跑上好几百年。
It could go for hundreds of years.
你的大脑就是这样的。
That's your brain, for example.
对吧？
Right?
关于 Bayt 播客
Bayt 提供中文+原文双语音频和字幕，帮助你打破语言障碍，轻松听懂全球优质播客。