Lab-in-the-Loop at Compute Speed — Ian Quigley, Leash Bio ·

00:00

My co-founder Andrew Blevins and I were at Procursion looking at a small molecule on a screen

我和我的联合创始人 Andrew Blevins 当时在 Procursion，看着屏幕上的一个 small molecule

00:04

And then a medicinal chemist walked by as this guy named Mike Cooker easy and Mike was like, hey

然后一个 medicinal chemist 路过，也就是这个叫 Mike Cooker easy 的家伙，Mike 说，嘿

00:08

Looks like a Stuart trivert molecule and then we are like, oh no, you can tell just by walking by

看起来像是一个 Stuart trivert molecule，然后我们就想，哦不，你光从旁边走过就能看出来

00:14

Maybe like who made this?

也许就像，这是谁做的？

00:15

That's horrifying and if the models can identify the chemist by looking at the molecule alone

这太可怕了，而且如果 models 光看 molecule 就能识别出是哪个 chemist

00:21

Then maybe the models can infer chemist intent which means maybe they don't learn anything at all

那也许 models 就能推断出 chemist 的意图，这意味着也许它们根本什么都没学到

00:30

We downloaded all chemicals from all chemists

我们下载了所有 chemists 的所有 chemicals

00:33

It chemble which is the biggest our repository of these things and their activities every chemist who published more than I think it was 30 papers and

它是 ChEMBL，是这些东西和它们的 activities 最大的 repository，每个发表超过——我想是 30 篇论文的 chemist，而且

00:40

600 molecules see it on molecules and all the stuff

600 个 molecules，在 molecules 上就能看到它，还有所有这些东西

00:43

It's 1815 chemists and then we tried to class fire to predict a chemist from looking at the molecule alone

是 1815 位化学家，然后我们试着用 classifier 只看 molecule 本身来预测化学家

00:49

And we got a top five accuracy of 60 percent

然后我们得到了 60% 的 top five accuracy

00:52

We chose to represent the molecules not with any chemical structures or formulas at all

我们选择表示 molecules 时完全不用任何 chemical structures 或 formulas

00:57

But just a vector of the probability of creator. So now it's this 1800 element

而只是一个 creator 概率的 vector。所以现在它就是这么一个 1800 维的

01:02

You know vector and it's got some probabilities who made it and you can use that representation and then train a model to predict activity from

你知道，vector，里面有一些概率说明是谁做的，然后你可以用这个 representation 去 train 一个 model，根据

01:09

Kempel and it does as well as if you use chemical formulas

Kempel 预测 activity，而且效果和用 chemical formulas 一样好

01:13

We really think that it's learning a lot of sociology. It's not learning ground truth when you train models on the on the exhaust you might be in trouble

我们真的觉得它学到了很多 sociology。它学的不是 ground truth；当你用这些 exhaust 去 train models 的时候，你可能会遇到麻烦

01:21

Welcome to late in space AI for science podcast. I'm RJ Hanuki, CTO of mere omics and I'm joined by my cohost Brandon Anderson

欢迎来到 late in space AI for science podcast。我是 RJ Hanuki，mere omics 的 CTO，和我一起的还有我的联合主持人 Brandon Anderson

01:30

He develops RNA therapeutics at a time of AI super excited to have Ian Quigley in the studio with today. This is a this episode

他在 AI 时代开发 RNA therapeutics，今天能把 Ian Quigley 请到演播室，真的超级兴奋。这是这一集

01:37

It's going to be a treat very excited

这会非常精彩，超级期待

01:39

Ian is the co-founder and CEO of leash bio

Ian 是 leash bio 的联合创始人兼 CEO

01:42

Which in my personal opinion is one of my favorite young tech bios at the last like few years

在我看来，这是我过去这几年最喜欢的年轻 tech bio 公司之一

01:48

You know at a high level

你知道，从宏观来看

01:50

leash is combining epistemic rigor data collection at like a crazy scale and

leash 在把 epistemic rigor 和疯狂规模的数据收集结合起来，而且

01:56

Has has a strong commitment to open science and has an incredible series of blog posts a lot of which we'll cover today

还非常坚定地投入 open science，并且有一系列超棒的博客文章，今天我们会聊到其中很多

02:02

But which I think really tells you how to do machine learning in a very hard to main in kind of the right way

但这，我觉得其实真的能告诉你，怎么以一种很难维持、但某种意义上正确的方式去做 machine learning

02:10

I think I'll have some fun. I know dogs hopefully can

我想我会玩得开心。我知道 dogs 希望可以

02:13

Trickle in through the episode, but yeah, just start with the thanks for being here and you know, what's your is there anything about us that you want to tell us before you get started

在这期节目里慢慢渗透进来，但，对，就先从谢谢你来这里开始，然后，你知道，你的——关于我们，有没有什么你想在开始之前告诉我们的？

02:22

Oh, no, thanks. That was very kind. Yeah, we we found at leash historically

哦，不用，谢谢。你太客气了。对，我们，我们至少从历史上看发现

02:27

After being confronted by a particular problem of trying to predict small molecules binding to protein targets

是在遇到一个具体问题之后，就是试图预测 small molecules 与 protein targets 的结合

02:32

And then finding that the data that was out there was kind of wanting and so we went out and built this thing at a reasonable scale to kind of try to go after it

然后发现当时现有的数据有点不够，所以我们就出去，以 reasonable scale 搭了这个东西，试着去攻克它

02:39

Yeah, yeah, so reasonable scale

对，对，所以是 reasonable scale

02:42

You're founded upon the idea of doing something called a DNA encoded library

你们是建立在做一种叫 DNA encoded library 的东西这个想法之上的

02:46

Every person I've ever met who's touched Adele has come out severely traumatized

我见过的每一个接触过 Adele 的人，最后都留下了严重的心理创伤

02:52

I don't know any person who has ever said does that's a great idea. Let's do hundreds or thousands of these things

我不认识任何一个说过“这真是个好主意，我们来做几百上千个这种东西吧”的人

02:57

So why what and also please explain what Adele is before you get started. Yes

所以为什么、是什么，还有，在你开始之前请先解释一下 Adele 是什么。好

03:03

So when you're going after a small molecule drug discovery

所以当你去做 small molecule drug discovery 的时候

03:07

The ultimate goal is to get a small molecule or a dinky chemical

最终目标是让一个 small molecule 或者一个不起眼的 chemical 去 bind 到那个你认为在某种疾病里造成问题的 protein 上

03:11

to bind to the protein that you think is causing the problem in a disease

然后从那里继续推进，而这些 protein 里有些很容易找到能粘上去的 chemical 材料，有些则更难

03:14

And then go from there and some of these proteins are pretty easy to find chemical material that sticks to you and some are harder

做这件事的一种方法，是把很多不同的 chemical 放在不同的试管里

03:20

One way to go about this is to have a lot of different chemicals in different tubes

一种做法就是，把很多不同的 chemicals 分别放到不同的试管里

03:24

And you then dunk the protein in one tube at a time and it kind of takes forever

然后你就一次一个 tube 地把 protein 浸进去，感觉永远都做不完。

03:29

And then a different solution that was proposed some years ago now was maybe if you had a lot of chemicals in the tube at once

然后几年前有人提出的另一个方案是，也许可以一次往 tube 里放很多 chemicals。

03:35

And then one tube and then you've got a small DNA barcode that's physically attached to each small molecule

然后就只用一个 tube，然后每个 small molecule 上都物理连接着一个小的 DNA barcode。

03:42

You dunk your protein in the tube kind of like a fishhuck you pull out different ones that bind

你把 protein 浸到 tube 里，有点像鱼钩，把能 bind 的不同分子拉出来。

03:47

And then you sequence the DNA of those and so the idea is that you get then pool this large number of molecules

然后你对这些的 DNA 做 sequencing，所以思路就是你把这一大批分子 pool 到一起。

03:52

And then you wouldn't have the logistical overhead of trying to go across large numbers of tubes or individual wells and plates

这样你就不用承担那种要跑遍大量 tubes 或者一个个 wells 和 plates 的 logistical overhead。

03:58

And so what's happened downstream of that is that you've got chemical libraries that if you do one tube at a time

所以这件事 downstream 带来的结果就是，你有了 chemical libraries，而如果一次一个 tube 来做的话……

04:03

The biggest ones in the world are in the low single digit millions with Del's the largest one that has been produced

世界上最大的那些也就是低个位数的百万量级，其中 DELs 是已经产出的最大的一类。

04:09

Is 40 trillion molecules

是 40 万亿个 molecules

04:11

There's some problems and they got that big

有些问题，它们变得那么大

04:12

But that is the kind of scale you could try to attempt with this technology

但这就是你可以用这种技术去尝试的那种 scale

04:16

So the way you

所以，你

04:18

Create these things is you you use this combinatorial chemistry where you have these building blocks

创造这些东西的方式是，你，你用这种 combinatorial chemistry，里面有这些 building blocks

04:24

And then you you build the library piece by piece, but you build it combinatorially

然后你，你一块一块地构建这个 library，但你是用 combinatorial 的方式构建它

04:28

So that is correct. Yes. Yes. And so to unpack that a little bit

所以这是对的。是的。是的。那稍微展开讲一下

04:33

If you've got you know some number of millions of molecules what this really means is it is a combinatorial mixture

如果你有，你知道，某个几百万个 molecules 的数量，这真正意味着它是一个 combinatorial mixture

04:40

Of building blocks and the number of building blocks is going to be a lot smaller than the final product of all of those

关于 building blocks，building blocks 的数量会比它们最终组合出的产物数量少得多

04:45

So you might be squared or cubed right

所以可能是平方，也可能是立方，对吧

04:48

And some different examples

再举几个不同的例子

04:50

So it's not uncommon for there to be maybe thousands or tens of thousands of building blocks that within generate a chemical library of many millions of molecules

所以，可能有几千甚至几万个 building blocks，用它们生成一个包含数百万个 molecules 的 chemical library，这并不少见

04:57

And this is

而这

04:58

Restrictive in some ways because you know you're stuck with the smaller number of building blocks

在某些方面是有局限性的，因为你知道，你就只能用这更少数量的 building blocks

05:02

But one thing that's really nice about it for machine learning is that you've got all these different combinations of things that look very similar to each other

但对 machine learning 来说，它特别好的一点是，你有所有这些不同的组合，而这些组合彼此看起来非常相似

05:08

Except for maybe one or two changes

只是可能有一两处变化

05:10

And so that you know, I'm sure going to get into this later

所以你知道，我之后肯定会详细讲这个

05:12

But that does really enable

但这确实能促成

05:14

comparisons so that you can identify oh, it's this part of the small molecule that drive in the binding

比较，这样你就能识别出，哦，是 small molecule 的这个部分驱动了 binding

05:19

Because I see a lot of variations in this other part that's not affecting the behavior too much

因为我在这个其他部分看到很多 variations，而这些 variations 并没有太影响 behavior

05:23

And that turns out to be incredibly useful

而事实证明这非常有用

05:25

The other aspect of this is how you do the readout which is the DNA sequencing

另一个方面是你怎么做 readout，也就是 DNA sequencing

05:30

This has been a big theme on the podcast that we've had many

这一直是播客上的一个大主题，我们有很多

05:32

Guests who have had some variation of we do something crazy

嘉宾都说过某种版本的“我们会做一些疯狂的事”

05:36

And then we convert the reading into a DNA sequencing

然后我们把读取结果转换成 DNA sequencing

05:40

Readout so how does this work and why is this so useful

readout，所以这到底是怎么运作的，为什么它这么有用？

05:44

Right

对

05:45

Yeah, I would definitely say that it is a party line at least that if you're going to go off and try to do

是啊，我肯定会说，至少这算是大家的共识：如果你打算跑去做

05:50

Life science stuff and couple it with machine learning approaches

life science 的东西，并把它和 machine learning 方法结合起来

05:54

There's really only really only one scalable technology in town and that technology is taking pictures

真的真的只有一个 scalable 的技术可用，而这个技术就是拍照

06:01

Sequencing DNA is another way of taking pictures

DNA sequencing 是另一种拍照的方式

06:04

And so it's basically pictures and DNA sequencing

所以基本上就是图像和 DNA sequencing

06:06

And if you can shovel your challenging problem

而如果你能把你那个棘手的问题硬塞进

06:10

Into that using one of those approaches as a readout

那个里面，用其中一种方法作为 readout

06:13

So for you know like recursion takes actual pictures of cells and a bunch of other groups do that

所以你知道，像 recursion 就是拍细胞的实际图像，还有一大堆其他团队也这么做

06:17

Like in cedar does that and then almost everyone else with a spatial omics stuff or other approaches

比如 cedar 也这么做，然后几乎所有搞 spatial omics 这类东西或者其他方法的人也都这样

06:22

They'll do some mixture of pictures and you know and sequencing which is also pictures

他们会把图像和，你知道，sequencing 混着来，而 sequencing 其实也是图像

06:26

Yeah, there's be some proteomics people out there who gonna you know disagree with me

对，外面肯定会有一些搞 proteomics 的人，你知道，会不同意我

06:29

But I still argue the scale is not there yet hot take number one

但我还是坚持认为 scale 还没到位，第一个暴论

06:35

Yeah, and one of the really cool things about sequencing is that it is dirt cheap

对，而且 sequencing 最酷的一点就是它便宜到爆

06:40

You can you know extremely scalable very

你可以，你知道，非常 scalable，非常

06:43

Well established technique

成熟的技术

06:45

It is it is very very cheap. It is getting cheaper all the time

它真的非常非常便宜。而且一直在变得更便宜

06:50

There was a little while where it kind of plateaued in terms of the price dropping

有一小段时间，它在降价这方面算是有点停滞了

06:54

But now a bunch of new companies have started up and they're driving it down again

但现在一堆新公司冒出来，又把价格压下去了

06:57

And so like we very much look at like the trajectory of secrecy cost dropping and it has implications for how we think about data collection in the scale that we're doing

所以我们非常关注 secrecy cost 下降的轨迹，这对我们怎么看待我们在做的这个 scale 下的 data collection 很有影响

07:05

Yeah, it's really nice. It's cool. So you have these components that you

是啊，这真的很好。很酷。所以你有这些 components，你

07:10

You are mixing together to form different this combinatorial mixture of molecules

把它们混合在一起，形成不同的，这种 combinatorial mixture of molecules

07:16

And each of those components is barcoded. Is that right? That's correct

而且那些组件每一个都带 barcode。对吧？没错

07:20

And so they and then they so they kind of bind together because they're

然后它们，接着它们，所以它们算是会 bind 在一起，因为它们...

07:26

These proteins change or these amino acid chains that will

这些 proteins 会变化，或者说这些 amino acid chains，它们会

07:30

Click on to the ends of each other and fold together and then they

Click 到彼此的末端上，然后 fold 在一起，然后它们

07:34

Either bind or don't bind to your target. So first of all, how do you know that they bound to your target?

要么 bind 到你的 target 上，要么不 bind 到你的 target 上。所以首先，你怎么知道它们 bind 到了你的 target 上？

07:41

Yeah, so first of all

对，所以首先

07:42

It is very much like this sort of fish analogy where if you were to so right now the target

这非常像那种钓鱼的类比，就是如果你要，所以现在这个 target

07:48

You know some protein or a protein complex that you're going to try to find chemical material to bind to

你知道，某个 protein 或者一个 protein complex，你要试着找到能 bind 到它上面的 chemical material

07:52

It's attached to a solid support

它会被连到一个 solid support 上

07:56

And so some folks do it a variety of different ways

所以有些人会用各种各样不同的做法

07:58

But the way that we usually do it is it's attached to magnetic bead

但我们通常的做法是，把它连到 magnetic bead 上

08:02

And so you dunk this bead in this complex mixture of all these small molecules

于是你把这颗 magnetic bead 浸到所有这些 small molecules 组成的复杂混合物里

08:07

And in a very easy case, let's pretend that the protein has a very distinctively shaped hole in it

在一个很简单的情况下，假设这个 protein 上有一个形状非常独特的洞

08:12

You dunk it in there and then the chemicals in that pool some tiny fraction of them have the appropriate shape and electronics

你把它浸进去，然后那一池 chemicals 里，有一小部分具有合适的 shape 和 electronics

08:18

That they're going to bind

它们就会结合

08:20

Into this very distinctively shaped hole

进入这个形状非常独特的洞里

08:21

So then you pull this protein back out and you only pull out the things that are in that hole

所以然后你把这个 protein 再拉出来，而且只拉出在那个洞里的那些东西

08:25

You rinse it a couple times in some way

然后你用某种方式把它冲洗个几次

08:28

And then you sequence the DNA of those things that come out that are in that hole

然后你对从那个洞里出来的那些东西的 DNA 做 sequencing

08:31

Okay, and then when you say sequence of DNA you're basically amplifying the barcodes that are that is correct

Okay，然后当你说 DNA 的 sequence 的时候，你基本上就是在 amplify 那些 barcodes，对，没错

08:38

And so and then you can

所以，然后你就可以

08:40

So amplifying any PCRs and they like that

所以就是 amplify 一些 PCR 之类的

08:42

You do a little bit of PCR and then yeah, then you toss it into the sequencer

你做一点点 PCR，然后，对，然后你就把它扔进 sequencer 里

08:46

And then one thing that you're probably thinking right now is okay

然后你现在可能正在想的一件事就是，好吧

08:49

Well, we need to compare it to something to get a good measure of relative enrichment

嗯，我们需要把它跟某个东西比较一下，才能很好地衡量 relative enrichment

08:53

And so you do have to engineer a bunch of good negative controls to say oh well if I just stuck the bead in there by itself

所以你得设计一堆好的 negative controls，来说明：嗯，如果我只是把 bead 单独放进去会怎样

09:00

Or are there pan-protein

还是说存在 pan-protein

09:02

Stickingness signals that are going to pull out certain chemical libraries chemicals from the library

还有那些 stickingness signals，它们会从 library 里拉出某些 chemical libraries chemicals

09:06

And so you've got to think pretty hard about all those things and make a reasonable comparison to the thing

所以你必须认真思考所有这些东西，并且跟那个东西做一个合理的比较

09:10

You're actually trying to to get binders for

你实际上是想为它拿到 binders 的

09:12

For the things that bind you sort of have these

对于那些能 bind 的东西，你算是会有这些

09:16

Combinations of barcodes that all sort of it okay if this

barcodes 的组合，它们差不多都——如果这个没问题的话

09:21

This one bound then I have like these how many barcodes are we talking about for one

这个一 bound 之后，然后我就有这些，呃，对于一个，我们在说多少个 barcodes？

09:26

One-protein question

one-protein 的问题

09:28

For one small molecule

针对一个 small molecule

09:30

So there's a couple things one thing if we want to get into the kind of a PCR-based weeds and so for the audience here

所以有几件事，一件事是，如果我们想深入讲讲那种 PCR-based weeds，那，对这里的听众来说

09:36

If you're making lots and lots of copies of DNA

如果你在做出很多很多 DNA copies

09:39

There's little subtle biases in the DNA that might make them amplify faster than others

DNA 里会有一些很细微的 biases，可能会让它们比其他的 amplify 得更快

09:44

And so you want to control for that

所以你要 control for 这一点

09:45

So you put in a little bit of unique sequence in there

所以你会在里面放一点点 unique sequence

09:48

So when you have lots of copies you can then collapse it so that they all came from this one that you know very much

所以当你有大量 copies 的时候，你就可以把它 collapse 起来，这样它们就都来自你非常清楚的这一个

09:53

You know, it's very much just from this one event

你知道，它基本上就是来自这一个 event

09:56

And then there's a bunch of other things about barcodes that we worry about at least

然后关于 barcodes，还有一堆别的我们会担心的事情，至少是这样

10:00

Which is that when you are trying to stick proteins in there some of those proteins are going to stick to DNA because that's their job

也就是，当你试着把 proteins 塞进去的时候，其中一些 proteins 会粘到 DNA 上，因为那就是它们的工作

10:06

Right there's lots of proteins whose job it is to interact with the genome and cells and turn things on or turn things off or affect them in some other way

对，有很多 proteins，它们的工作就是和 genome 以及 cells 相互作用，把东西开启或关闭，或者以其他方式影响它们

10:14

And then you've got this assay where you've got a ton of DNA and you're throwing them in there

然后你又有了这个 assay，里面有大量的 DNA，而你把它们扔进去

10:18

Yeah

嗯

10:19

And so that could be risky

所以那可能挺有风险的

10:21

And that can cause false positives right very much so yeah a lot of people could play in about this

对，那确实会导致 false positives，非常是，是啊，很多人可能会在这件事上加入进来

10:26

This is kind of to the earlier comment about folks having their dull nightmares

这有点像之前那个评论，说人们会做自己那些无聊的噩梦

10:30

This is one of those problems

这就是那类问题之一

10:31

There's a variety of ways to address it

有很多种方法可以解决它

10:34

One thing that we do is we have large numbers of barcodes that each encode

我们做的一件事是，我们有大量的 barcodes，每一个都 encode

10:39

One particular type of small molecule

一种特定类型的 small molecule

10:41

Because here's one small molecule that's great

因为这里有一种很棒的 small molecule

10:43

Now we have 32 or 48 different barcodes all encoding this one small molecule

现在我们有 32 或 48 个不同的 barcodes，全都在 encode 这同一个 small molecule

10:50

Oh I see right and so then you know so like mallet you know there'll be you know

哦，我明白了，对，然后就是，你知道，所以就像 mallet，你知道，会有，你知道

10:54

I single example one single molecule has got this barcode and then it's neighbor which is the same

一个例子，一个 single molecule 有这个 barcode，然后它旁边的那个也是一样的

10:59

Chemical formula will have a different barcode and then so on so if there's a differential there then you know that you

Chemical formula 会有不同的 barcode，然后以此类推，所以如果那里有差异，那你知道你

11:04

That is right that you're binding to the DNA instead of the that's right

没错，就是你 binding 到 DNA，而不是那个，没错

11:08

That's right and so and so for like the the DNA protein aficionados in the audience

没错，然后，然后对于观众里那些 DNA protein 爱好者来说

11:13

One really nice thing about this is that because there's DNA in there

这件事一个特别好的地方是，因为里面有 DNA

11:16

You can see if there's a single DNA encoding barcode that's like way binds way better than everything else

你可以看到，如果有一个单独的 DNA encoding barcode，它 binding 得比其他所有东西都好得多

11:24

Even though it's all the same small molecule

尽管它们全都是同一个 small molecule

11:26

And then you can look at the sequence that you haven't reached there

然后你可以看看你还没到的那段 sequence

11:29

And you can recover the described DNA binding preferences of those proteins

然后你可以把那些 proteins 被描述过的 DNA binding preferences 恢复出来

11:34

It's like here's a transcription back here

就像这里有一个 transcription back

11:36

It binds this series of nucleotides that is very well described

它会 bind 这一系列 nucleotides，这个被描述得很清楚

11:40

You know it's like six or eight of them

你知道，大概就是六个或者八个

11:41

And then you look at the

然后你去看那个

11:43

You recover the DNA sequences back from the six parameters like oh great

你从这六个 parameters 里把 DNA sequences 恢复回来，就像，哦太好了

11:46

We found the consensus binding motif, right? So I think our protein probably got made right you have this

我们找到了 consensus binding motif，对吧？所以我觉得我们的 protein 很可能做对了，你有这个

11:52

barcode that you've

那个 barcode，你已经

11:54

Identified and you've kind of gotten rid of the false positives using the technique that you just had

识别出来了，而且你用刚刚那个方法差不多把 false positives 清掉了

11:59

and so now I

所以现在我就

12:02

Know that for a given

知道对于某个给定的

12:05

combination so for a given small molecule that I mean I bind to my target and

combination，所以对于某个给定的 small molecule，我是说，我能 bind 到我的 target，而且

12:11

And it's but don't you have all these

然后就是——但你不是有所有这些

12:14

DNA tags barcode tags like hanging off your molecule does not impact the yes

DNA tags、barcode tags 这样挂在你的 molecule 上，不会影响那个……是的

12:20

They are happy they are happy here. This is another complaint that people bring up about delts

它们挺开心的，它们在这里挺开心的。这是人们关于 delts 提出的另一个抱怨

12:24

There's a couple different answers here. I would say but like the the very like 10,000-foot philosophical answer

这里有几个不同的答案。我会说，但就像那种，那种非常宏观的哲学层面的答案

12:31

Is that there are two different or there are multiple different reasons why you might want to roam with these experiments

也就是说，可能有两个不同，或者说有多个不同的原因，让你想在这些实验里四处探索

12:36

One reason is that you're trying to get a small molecule material

一个原因是，你在尝试得到一种 small molecule material

12:40

That behaves a certain way that you're going to progress into drug programs

它会有某种特定的表现，然后你会把它推进到 drug programs 里

12:43

And so if that is your main goal then you want to make sure that you're getting some kind of fidelity

所以如果那是你的主要目标，那你就想确保自己能获得某种 fidelity

12:48

Between having the DNA barcode on there and then having it off there

也就是在把 DNA barcode 放在那上面，和之后把它从那儿拿掉之间

12:51

And that's true

这是真的

12:52

And when we try to pursue interesting chemical material at least like you do have to go through the work of

而当我们试图去追求有趣的 chemical material 时，至少就像，你确实得去做……的工作

12:57

Manufacturing the chemical off DNA

基于 DNA 制造化学分子

12:58

You've got to run them in these functional assays. There's a certain level of attrition

你得把它们放到这些 functional assays 里跑。会有一定程度的 attrition

13:02

For things that are easy in our hands about 90% of them work for things that are hard is about 50

对于在我们手上容易做的，大概 90% 都能成功；难做的，大概是 50%

13:06

So there's some loss other people have different experiences

所以会有一些损失，别人的经验可能不一样

13:09

But I would also say that if you're thinking about doing this for machine learning training purposes

但我也会说，如果你是在考虑出于 machine learning training 的目的来做这件事

13:13

All these chemicals with their DNA barcodes are in the tube you ran the experiment

所有这些带 DNA barcodes 的化学分子，都在你跑实验的那个管子里

13:18

Some of them bound and some of them didn't and that is real signal that you were physically measuring from the world

有些结合了，有些没结合，而这是你从真实世界里实际测量到的真实 signal

13:22

And so then you just condition your models on the existence of this barcode

然后你就只需要根据这个 barcode 是否存在来 condition 你的 models

13:26

And you need to keep moving on with your model training

而且你得继续推进你的 model training

13:28

And so in our hands when we run experiments

所以在我们这边，当我们做实验的时候

13:31

So make fresh protein batches make a new batch of the chemical library over many many months of time

所以要做新的 protein batches，并在很多很多个月里做一批新的 chemical library

13:37

We get back the same hits in the same order for the most part

我们大体上会以相同的顺序拿回相同的 hits

13:40

And so it's extremely reproducible and robust

所以它非常 reproducible，也很 robust

13:42

And so in terms of like ground tree signal we feel quite good about it

所以就 ground tree signal 而言，我们觉得还挺不错的

13:45

Okay

好

13:46

You could be blocking the pocket or something with your

你可能用你的……堵住了 pocket 或者什么的

13:48

You could

你可以

13:49

That'll be a false negative there

那就会是个 false negative 了

13:51

But you don't because you know that the DNA is there it did not bind

但你没有，因为你知道 DNA 就在那儿，它没有 bind

13:55

The way that you're thinking about it is like oh I'm thinking about how the molecule behave missing the DNA

你的想法就像是，哦，我在想这个 molecule 在没有 DNA 的时候会怎么表现

14:00

But we know it's there and we don't pretend that it's not

但我们知道它在那儿，我们不会假装它不在

14:02

Oh I see so it does your hit include the DNA then

哦，我明白了，那你的 hit 里包括 DNA 吗

14:06

Yeah like when we train models we put we very much put it in a placeholder

对，就像我们训练 models 的时候，我们会把它——我们非常明确地把它放进一个 placeholder 里

14:09

Uh that that marks like the position of the DNA

呃，那个，那个就像标记了 DNA 的位置

14:12

If you're going to

如果你打算

14:14

Synthesize this for

为…… synthesize 这个

14:16

The drug we leave that

我们留下的那种药

14:18

That's where you get the nutrition rates right

这就是你把 nutrition rates 弄对的地方

14:20

And then you've got to make it and sometimes they behave differently

然后你还得把它做出来，而有时候它们表现会不一样

14:23

Yeah yeah

对对

14:24

But some practice I got it

但有些练习我搞懂了

14:28

You know actually those are really good questions

你知道，其实那些真的是很好的问题

14:30

Yeah yeah

是啊是啊

14:32

There are lots of ejections and like there's one

有很多 ejections，而且就像，有一个

14:34

These are many of the reasons why folks are trying to do the thing that we're trying to do.

这些就是为什么大家要去做我们正在尝试做的这件事的许多原因。

14:38

Yeah, I was going to say, going back to my original statement, I again have never met someone who

对，我正想说，回到我最开始说的，我还是从没见过一个人

14:43

was worked on Dells and said, we should do more of these. So, yeah. So, I mean, I think, I mean,

在 Dells 工作过，然后说，我们应该多做点这些。所以，对。所以，我是说，我觉得，我是说，

14:51

hopefully we will get to, I think, why I'm actually excited that this route seems to be working.

希望我们能讲到，我觉得，为什么我其实很兴奋，因为这条路子似乎行得通。

14:58

Hopefully the audience will agree with me before this is over. In fact, actually, maybe even a

希望在这结束之前，听众会同意我的看法。其实，实际上，也许甚至可以说一段

15:03

personal anecdote, we have a mutual friend who was, I think, an angel investor in Leish.

个人轶事，我们有个共同好友，我想，是 Leish 的天使投资人。

15:08

And when he told me he was an angel investor in a Dell company, I jokingly said, okay,

当他告诉我他是一家 Dell 公司的天使投资人时，我开玩笑地说，好吧，

15:12

well, that's why you'll never see again. I'm like, here in front of our entire audience to say,

那么，这就是为什么你再也见不到了。我就说，现在当着我们所有观众的面说，

15:18

I was wrong. I wish I had listened. I'm curious about, first of all, why are Dells so hard?

我错了。我真希望我当初听了。首先，我很好奇，为什么 Dells 这么难？

15:27

I think we just covered a lot of these. Like, what in the historical community of people doing Dells

我觉得我们刚刚已经讲了很多这些了。比如说，在历史上做 Dells 的那群人里，

15:33

was making it so dramatic as an experience. And then why did you think you could overcome that?

是什么让这段经历变得如此戏剧化？然后，你为什么觉得自己能克服这一点？

15:40

One thing I do want to say is that when Dells were invented, it was definitely at a time when

我确实想说的一点是，Dells 刚被发明出来的时候，绝对是在一个

15:48

sequencing was kind of more expensive. And it really felt like taking doing sequencing experiments

sequencing 还比较昂贵的时期。而且当时真的感觉，做 sequencing 实验，

15:53

was kind of like a big logistical and financial swing. And then when you built the technology

有点像是一次后勤和财务上的大动作。然后当你把这项技术做出来的时候

15:59

at the time and then pointed at the problems that it was pointed at, it was very much,

在当时，然后指向了它当时所指向的那些问题，它非常...

16:03

we want to find new chemical material by needed targets that are very challenging.

我们想通过那些非常有挑战性的所需靶点，找到新的化学物质。

16:07

And so that brings up a couple kind of tactical issues. One of them is that

所以这就带出了几个有点战术性的问题。其中一个问题是

16:14

classically, folks don't screen large numbers of protein targets. They're usually interested

传统上，人们不会筛选大量的 protein targets。他们通常感兴趣的是

16:18

in finding chemical matter for something that's like really hard and they have maybe tried a bunch

为某种特别难搞的东西寻找 chemical matter，而且他们可能已经试了一堆

16:24

of other techniques and they haven't worked out. So, okay, let's try Dell screen because we don't

其他技术，但都没成功。所以，好吧，咱们试试 Dell screen，因为我们没有

16:27

have anything. Then they'll go off and make the protein that they're going to try to screen.

任何东西。然后他们会去做他们准备尝试筛选的那个 protein。

16:32

And they'll put a lot of work into it because it feels like kind of a lift to do a Dell screen.

而且他们会在这上面投入很多工作，因为做一次 Dell screen 感觉挺费劲的。

16:36

And so maybe they might spend two to six months growing big batches of the protein that they care

所以也许他们会花两到六个月，培养大批他们关心的 protein

16:41

about, doing a bunch of purifications, maybe they're going to run some functional assays to convince

然后做一大堆 purifications，也许还会跑一些 functional assays 来说服

16:45

themselves that oh, it's really got to be this good off for us to go off and run this screen.

自己：哦，它真的必须好到这个程度，我们才会去跑这个 screen。

16:51

And then when they run the screen, they're going to try to maximize their chances of getting

然后等他们跑 screen 的时候，他们会尽量提高自己拿到

16:55

pretty good chemical material. And so then the incentive drives into a place where we want,

相当好的 chemical material 的机会。于是激励就会把人推向一个我们想要的局面，

16:59

we want a big Dell. We want a lot of diversity. And so if you go looking at folks who are offering

我们想要一个大 DEL。我们想要很多 diversity。所以如果你去看那些提供

17:04

CRO services or internal Dell groups a lot of times, they'll push into these very large Dell's

CRO services 的公司，或者内部的 DEL 团队，很多时候，他们都会推你去搞这些非常大的 DEL

17:09

and hopes. And that's their motivation. And it makes a lot of sense. What we have done that's

还有希望。这就是他们的动力。这很说得通。我们所做的事，非常

17:13

very, very different is we screen tons and tons and tons of proteins. And we are less driven by

非常不同，就是我们 screen 海量、海量、海量的 proteins。而且我们不太受

17:19

trying to find individual binders for a particular protein target that might be of high interest for

为某个特定 protein target 寻找单个 binders 这件事驱动，而这个 target 可能对

17:25

a program. We just want to get tons and tons of active chemical material across large

某个 program 来说高度关注。我们只是想获得海量、海量、海量的 active chemical material，覆盖大量

17:29

numbers of proteins. And so there we can identify how our Dell's behave very, very well because

的 proteins。这样我们就能很好地看出我们的 DELs 表现如何，因为

17:37

every one of our chemical libraries gets screened against many hundreds of targets.

我们的每一个 chemical library 都会被 screen 到数百个 targets 上。

17:41

So you're characterizing the chemical library. The chemical library better.

所以你们是在 characterize chemical library。把 chemical library characterize 得更好。

17:45

We know very, very well. And so we start to encounter lots of these motifs that are quirk

我们非常非常了解它。于是我们开始遇到很多这些 motif，它们很 quirk

17:50

to the chemistry that we designed or maybe the manufacturing process. Like, oh, well,

到我们设计的 chemistry，或者也许是 manufacturing process。就像，哦，好吧，

17:54

even, I mean, a great example is we have one chemical library that we've had made a couple times.

甚至，我的意思是，一个很好的例子就是我们有一个 chemical library，我们已经让人做过好几次了。

17:58

And even though it's the same chemical library, it does behave differently. And we did not know

而且即使它是同一个 chemical library，它的表现也会不同。而我们当时不知道

18:02

that until we screened hundreds of proteins against it. And then you have to stare at those results

直到我们用数百个 proteins 对它做了 screen。然后你必须盯着那些结果

18:06

and say, oh, well, now we're seeing these motifs show up a little bit more. And then as you

然后说，哦，好吧，现在我们看到这些 motifs 出现得稍微多了一点。然后随着你

18:10

characterize the library more and more, it really helps you evaluate and troubleshoot even your protein

越来越深入地 characterize 这个 library，它真的能帮你评估和排查甚至你的 protein

18:15

production process. And so the one reason that we felt that this is pretty tractable overall

production 流程。所以我们觉得这件事整体上相当 tractable 的一个原因是

18:21

is that I and a couple of the people, the company come from real like genomics methods development

就是我和公司里的几个人来自真正的 genomics methods development

18:27

backgrounds as compared to chemistry backgrounds. And so we knew what it might take to tweak experimental

背景，而不是 chemistry 背景。所以我们知道要 tweak experimental 可能需要什么

18:33

protocols to really get good signal out of these funky assays. And I would say it was that plus

protocols，才能真正从这些古怪的 assays 里拿到好的 signal。我会说，是这一点，再加上

18:39

many years at recursion as where I was last with my co-founder, this kind of Andrew Levenson,

在 recursion 的很多年，那是我上一次和我的联合创始人——这位 Andrew Levenson——共事的地方，

18:43

a number of other people on our team. And we all made, you know, honestly like a lot of mistakes

还有我们团队里的其他一些人。我们都犯过，你知道，老实说，很多错误

18:48

in high dimensional biology, bumping around in the dark, figuring out what different modes,

在 high dimensional biology 里，在黑暗中乱撞，搞清楚有哪些不同的 modes，

18:53

failure modes there might have been that would allow models to cheat on, you know, the high

failure modes，可能会让 models 在，你知道，high

18:56

dimensional imaging data that recursion produces. And so we had this real eye for thinking about, oh,

dimensional imaging data 上作弊，这些是 recursion 产出的数据。所以我们真的练出了这种眼光，会去想，哦，

19:01

well, you know, there's these sources of noise this way. We're going to chase those down and, oh,

嗯，你知道，这边有这些 noise 来源。我们要把这些追查下去，然后，哦，

19:05

can you predict, you know, the instrument that this was run on and if you can, that's really bad.

你能不能预测，你知道，这个是在哪个 instrument 上跑的，如果你能，那就真的很糟。

19:09

And so we incorporated all that thinking into going after this particular assay. And we had a feeling

于是我们把所有这些思考都融入到了攻克这个特定 assay 的过程里。而且我们有一种感觉

19:13

that folks hadn't really used stared at it, you know, in that amount of detail before. Yeah,

就是大家之前并没有真的用过、盯着它看过，你懂的，没到那种细节程度。对，

19:19

this goes back to the Ron Alpha, noetic episode where I found, yeah, I found, there's some real

这要追溯到 Ron Alpha, noetic 那期节目，我在那里发现，对，我发现，有一些非常真实的

19:29

spiritual parallels between both of your companies, this like real epistemic rigor checking

你们两家公司之间的精神层面的相似之处，这种真正的 epistemic rigor，检查

19:34

very carefully everything you're doing, doing things in different ways and sort of like semi-overlapping

非常仔细地检查你们在做的一切，用不同的方式做事，而且有点 semi-overlapping

19:40

ways. And I mean, I'm just curious, yes, we're on this, yeah, very data focused, like at scale.

的方式。我的意思是，我就是好奇，对，我们正在做这个，对，非常 data focused，就像 at scale 那样。

19:46

Automation. And yeah, and I'm curious, like, was there something like at recursion, like was

Automation。然后，对，我好奇的是，像是在 recursion，是不是有什么，就像是

19:52

this just like quarter recursion's culture that is sort of, it is like, yes, yes, especially in the

这就像是 quarter recursion 的文化，算是，就像，对，对，尤其是在

19:58

early years. I would say that there are four that I think, I think there's four X-like recursion

早年的时候。我会说有四家，我觉得，我觉得有四家 X-like recursion

20:06

mafia companies, and they all kind of rhyme. So there's us, there's Noetic with Ron Alpha,

mafia 公司，而且它们都挺像的。所以有我们，有 Noetic，还有 Ron Alpha，

20:11

there's Envated with Vishwakaluru and Augustaal and Hanukwarden and all these great people of

有 Envated，里面有 Vishwakaluru、Augustaal 和 Hanukwarden，以及所有这些很棒的人，来自

20:15

Davily. And then there's also known medicine that was run by Katie Risskele, they got picked up

Davily。然后还有 known medicine，是由 Katie Risskele 经营的，他们被

20:19

by pace a couple years ago. And so all these companies are basically looking at different interesting

pace 几年前接手了。所以这些公司基本上都在看不同的、有意思的

20:25

biology problems with extremely noisy measurements. And then they pull in people from recursion to

biology 问题，带着极其 noisy measurements。然后他们会从 recursion 拉人进来，来

20:30

figure out, like, oh, we're pretty good at noisy measurements. Let's try to see if we can get

搞清楚，比如说，哦，我们挺擅长 noisy measurements 的。我们试试看能不能让

20:33

consistent a lot of this particular corner. You know, it's a for us, it's a small molecule chemistry

这个特定角落的很多东西都变得 consistent。你知道，这——对我们来说，这是 small molecule chemistry

20:37

problem for invata, it's it's a mixtures of natural products, they use it, they do do a lot of,

对 invata 来说，问题在于，它它是 natural products 的混合物，他们会用，他们会做很多，

20:44

like, you know, kind of mass spec kind of stuff. And then there's Ron and then there's Katie Risskele's

就那种，你懂的，mass spec 那一类的东西。然后还有 Ron，还有 Katie Risskele 的

20:49

group. So yeah, we're all very similar. So let me let's pause for just a second. What

团队。所以对，我们都很像。那让我——我们先停一下。你们

20:55

do you guys actually do? Sorry, I'm so excited about this. I just jumped in. Yeah, we are a

到底做什么？抱歉，我太激动了。我直接插进来了。对，我们是一家

21:05

machine learning small molecule company. And so what that means is we go off and we collect

machine learning small molecule 公司。所以这意味着我们会去收集

21:12

massive numbers of small molecules by need to protein targets. And we use that to drive

大量 small molecules，来满足 protein targets 的需求。然后我们用这些来推动

21:17

advancements and machine learning understanding of this kind of stuff. And we also push, you know,

进展，以及 machine learning 对这类东西的理解。我们也会把，你懂的，

21:22

small molecule material into partners and move it forward towards, you know, towards patients.

small molecule 材料推进给合作伙伴，然后把它往前推，你懂的，朝着患者。

21:27

Just for the audience, it's not for me. What, what is a small molecule? Why do we care? How is it

这是讲给听众的，不是给我自己听的。

21:32

different than other therapeutic? This is great. Nice. This is great. So when you're trying to

那，什么是 small molecule？

21:37

drug something, I talked a little bit earlier about how you've got a protein of interest in your

我们为什么关心这个？

21:42

body and you're going to try to interact, you know, affect its behavior in some way. There are a

它和其他 therapeutic 有什么不同？

21:46

couple different things that are currently on the market that people use. One thing that's common

这太好了。

21:51

is antibodies which are made by cells. They're large molecules. And by large, they're big enough

不错。

21:59

that they have a really hard time getting inside cells. They also have a really hard time being

这太好了。

22:04

in a pill. So if you have like antibody treatment or one of these protein treatments like GOP,

所以当你想 drug 某个东西的时候，我之前稍微讲了一下，你体内有一个 protein of interest，然后你要试着去和它互动，你知道，以某种方式影响它的行为。

22:09

like GOP1 treatments, you're going to have to have kind of an ejection. It can affect certain

就像 GOP1 疗法，你得有某种 ejection。它会影响某些

22:14

proteins that are easy for them to get to. But there's other ones inside the cell that drive a lot of

那些它们很容易接触到的 proteins。但细胞里还有其他一些，会驱动很多

22:18

disease that are quite challenging for those. It is kind of more appealing in a lot of ways to

疾病，而这些对它们来说相当有挑战性。从很多方面来说，更吸引人的是

22:22

instead have this big thing that's a protein to have a little tiny chemical. And tiny chemicals,

反过来，不用这个大的、是 protein 的东西，而是用一个小小的化学分子。而小小的化学分子，

22:27

small molecules here are things that can be in pills. They can go inside your cells and affect

这里说的 small molecules 就是可以放进药片里的东西。它们能进入你的细胞并影响

22:31

behaviors. They do make up most of drugs. If you've got any pills in your house, they're almost

行为。它们确实构成了大多数药物。如果你家里有任何药片，那它们几乎

22:37

certainly small molecules. And of the different machine learning drug discovery groups that are

肯定都是 small molecules。而在不同的 machine learning drug discovery 团队中，那些

22:42

going after treatments in a general sense, we have made, we as in all of life scientists in the

从广义上在寻找治疗方法的，我们已经做出了，我们是指所有生命科学家在

22:48

21st century, have made pretty good progress on being able to predict and enable the behaviors

21世纪，已经在能够

22:54

of protein-type drugs. So you've got a bunch of antibody colonies and like isomorphic

protein-type drugs 的。所以你有一堆 antibody colonies，还有像 isomorphic

22:59

does this or like chai or noblan. There's all these groups of very good, very smart people pushing

在做这个，或者像 chai 或 noblan。有所有这些非常优秀、非常聪明的人组成的团队在推进

23:04

on these particular problems. And I think part of that is because we understand or have enough

这些特定的问题。而且我觉得部分原因是，我们了解，或者有足够的

23:09

data for protein behaviors that make it reasonably effective for these companies to do these

关于 protein behaviors 的数据，让这些公司能够相当有效地做这些

23:14

designs. Arthesis is that those capabilities are really not here for small molecule approaches.

设计。Arthesis 是，那些能力对于 small molecule approaches 来说其实还不存在。

23:20

And so that was one reason why we built this company as we thought that we could go after

所以这就是我们创办这家公司的原因之一，因为我们觉得我们可以去追求

23:23

a meaningful way. And small molecules, one of the key characters, they go everywhere in your body.

以一种有意义的方式。而小分子，作为关键角色之一，它们会遍布你全身。

23:29

So that's good and bad. They can penetrate into the cell and they can also get through

所以这有好有坏。

23:36

defenses of tumors and cross the blood-brain barrier more easily. But at the same time,

它们能穿透 cell，也能穿过 tumors 的防御，还更容易穿过 blood-brain barrier。

23:42

they do all those things and therefore they can be more prone to toxicity and other side effects.

但与此同时，它们能做到所有这些，所以也更容易出现 toxicity 和其他 side effects。

23:47

Very true. So the kind of path through the regulatory barriers for small molecules can be a little

确实如此。

23:54

bit more fraught than those for antibodies. And so antibodies we know pretty well right now about

所以 small molecules 走 regulatory barriers 这条路，可能会比 antibodies 更棘手一点。

23:58

how they behave in the body and what they're likely to cause problems thing and what they're not.

而 antibodies，我们现在已经相当了解它们在体内如何表现，以及它们可能会引起哪些问题、哪些不会。

24:03

And for small molecules, there's a bunch of well-described problems that some of them might

而对于 small molecules，有一大堆已经被充分描述的问题，其中一些可能会影响到。

24:08

affect. They might cause liver problems, for example. They might cause heart arrhythmias.

比如，它们可能会引起 liver problems。

24:12

And as a medicinal chemist, a person whose job it is to develop these molecules to get them

而作为一名 medicinal chemist，一个工作就是开发这些 molecules 的人，要把它们从在试管里看起来还行的东西，变成你有足够信心给到人类的东西，这里面有很多这样的坑，他们得做很多工作来避开这些坑。

24:18

from something that looks okay in a tube to something that you would feel confident enough to

还要确保我的 molecule 不会引起 heart arrhythmias。

24:22

give to human being, there are a lot of these pitfalls and they have to do a lot of work to avoid

然后如果我改一下它，让它不再引起这些问题，这会不会影响它去做我想让它做的事？

24:27

these pitfalls. And make sure that my molecule doesn't cause heart arrhythmias. And then if I change

你们的 models 如何帮助解决这个更大的问题——开发这些，然后也许还有那种 downstream 问题？

24:31

it to stop the causing those problems, does it affect it doing the thing that I want it to do?

所以一种思考方式是，这些问题中的一部分，这类 off-target toxicities，会是由你的 small molecule 驱动的，

24:36

How do your models help with this larger problem of developing these and then maybe with sort of

你们的模型如何帮助解决这个更大的问题，即开发这些分子，然后也许处理

24:44

these downstream problems? So one way to think about it is that some number of these problems,

这些下游问题？所以一种思考方式是，这些问题中的一部分，

24:50

these sort of off-target toxicities are going to be driven by your small molecule,

这些所谓的脱靶毒性，将会由你的小分子驱动，

24:54

by need to other proteins in ways that you might be able to predict right now. And so that's

以你现在也许能够预测的方式与其他蛋白质发生需求关系。所以这就是

24:58

something that we are very much pushing forward to is, oh, let's also make those proteins,

我们正在大力推进的事情，哦，让我们也制造那些蛋白质，

25:03

let's also screen those proteins or their chemistry. It seems useful if you have a giant assay of

让我们也筛选那些蛋白质或它们的化学性质。如果你有一个巨大的检测库，

25:08

that's right, different proteins in a library that you can... Yeah, it is not at all an exaggeration

没错，一个包含不同蛋白质的库，你可以……是的，这完全不夸张

25:13

to say that it is our aspiration to have screened every protein. And then also move on to those

说我们的愿望是把每一个 protein 都 screen 过。然后还要继续推进到那些

25:19

outside of humans, both for reasons to get variation in the proteins, but also to identify

人类之外的，既是为了获得 protein 的 variation，也是为了去 identify

25:24

maybe you don't want to run a mouse model, maybe you should run a rat one because your chemical

也许你不想跑 mouse model，也许你应该跑 rat 的，因为你的 chemical

25:28

material behaves differently in those different species because those proteins might be slightly

material 在不同的 species 里表现不同，因为那些 proteins 可能稍微

25:31

different. Interesting. Right, so you want to do that stuff. But yes, going after large numbers of

不一样。有意思。对，所以你想做那些事。但没错，去搞大量

25:36

these things allows you to think, all right, well now I've got my chemical material and then I know

这类东西能让你想，好吧，现在我有我的 chemical material，然后我知道

25:40

of all these chemicals which ones and what variations are ones likely to bind these things will

所有这些 chemicals 里，哪些以及哪些变体可能 bind 这些东西，会

25:44

cause other toxicity problems. Got it. Yep. And then downstream of that, there's a bunch of other

造成其他 toxicity 问题。明白了。对。然后在那之后，还有一堆其他

25:49

assays that lots of people do that are more kind of emergent behaviors where it's maybe not a

assays，很多人会做，它们更像是 emergent behaviors，可能不是单个

25:54

single protein, but it's how well it gets across cell membranes, how interaction a more complex way

protein，而是它穿过 cell membranes 的效果如何，如何以更复杂的方式发生 interaction

26:00

that you require a cell to evaluate or it might be like a cardiac cell, for example, if it was

需要你用一个 cell 来评估，或者比如可能像 cardiac cell，如果它

26:04

harder with me, a stuff. And so those things, you know, right now there are lots of groups that are

对我来说更难，之类的东西。所以那些东西，你知道，现在有很多团队都在

26:08

pushing on technologies to measure these at the cellular level, which is one level of biological

在推进一些技术，想在 cellular level 测量这些，这比我们现在关注的东西在 biological

26:13

abstraction away from what we're focusing on right now. But we like to think that models have a

abstraction 上隔了一层。但我们倾向于认为，models 如果对

26:17

good understanding of the biophysics themselves can be useful and maybe fine tuned on smaller

biophysics 本身有很好的理解，会很有用，也许未来可以在更小的

26:21

datasets in the future. Popping back one more maybe in getting a little into the model wide

datasets 上 fine-tuned。再往回跳一点，也许稍微聊聊 model-wide

26:26

delts. Delts really are, at least in our view, the most scalable technology for measuring small

DELs。至少在我们看来，DELs 确实是测量 small

26:33

molecule interactions because you can put many millions of them in a tube. Now one thing that

molecule interactions 最 scalable 的技术，因为你可以把数百万个放进一根管子里。

26:39

that we haven't touched on in terms of like these chemical libraries is that when you make them

现在有一件事

26:44

in a pooled process, there are probably going to be some that react imperfectly and that the

我们还没提到，就是这些 chemical libraries：当你把它们

26:50

label translation of what you think you have might be, it might be a little bit off. And so we've

你觉得自己有的东西的 label translation，可能有点偏。所以我们已经

26:54

designed a lot of controls into constructing our chemical libraries to mitigate this. When you're

把很多 controls 设计进了我们构建 chemical libraries 的过程，来缓解这个问题。当你

26:59

using these, you know, the word might be arrayed libraries or libraries that are in one chemical

用这些的时候，你知道，那个词可能是 arrayed libraries，或者是一次一种 chemical 的 libraries

27:04

at a time at a tube or in a well and a plate, when folks study it to figure out how good those

比如在 tube 里，或者在一个 well 和一个 plate 里，当人们研究它，想弄清楚那些

27:10

labels are, like you hope that they're good, but when folks examine this, they're typically off by

labels 到底有多好，你当然希望它们是好的，但当人们检查这个时，它们通常会偏

27:16

20 to 30%. It's an enormous mislabeling problem. We think it's probably comparable to mislabeling

20 到 30%。这是个巨大的 mislabeling 问题。我们认为它可能和 mislabeling

27:21

a delts maybe worse and delts themselves are just enormously, enormously scalable. And maybe I can,

一个 delts 差不多，也许更糟，而 delts 本身又极其、极其 scalable。也许我可以，

27:27

I'll tell a quick story that I'm sure that we'll get into at some point anyway. We ran a

我讲个简短的故事，我相信我们之后总会聊到。我们跑了一个

27:32

Kaggle competition a couple of years ago and we decided to generate the data for the Kaggle

几年前有一个 Kaggle competition，我们决定为这个 Kaggle competition 生成 data，然后把它发布了。

27:37

competition and we released it. There's existing, there's a database that was at the time two

当时已经有一个 database，那是两年前叫 Binding D.P. 的 database，它是最大的 small molecules finding to protein targets 的聚合集合，相比你的 cells 或者别的什么。

27:41

years ago called Binding D.P., which was the biggest aggregated collection of small molecules

它大约有两百万，稍微不到三百万个 physical interactions。

27:46

finding to protein targets as compared to your cells or something else. And it was about two,

于是我们发布了一个大约四十亿个 physical interactions 的 dataset。

27:50

a little under three million physical interactions. And so we released a data set that was about four

而这些都是由一个人在我家地下室里测出来的。

27:55

billion physical interactions. And it was all measured by a single person in the basement of my house.

是啊，早期的 blog posts 真的很好笑。

28:01

Yeah, the early blog posts are really funny. You're like, literal picture, you have an, it's not

你就像，真的放了一张照片，你有一个，它不完全是车库 startup，但它是，它是在地下室里。

28:06

quite a garage startup, but it's, it's in a basement. It's a basement apartment with like a seven

那是个地下室公寓，大概有个七……

28:10

foot ceiling. Yeah. It looks claustrophobic. Yeah. I've never seen a biotech startup started in a

英尺高的天花板。是啊。看起来有点幽闭恐惧。是啊。我从没见过一家 biotech startup 是在

28:17

basement. Oh, we don't. Okay, so the, so this Kaggle competition I think is just a fascinating story

地下室里起步的。哦，我们不是。好吧，所以这个，所以这个 Kaggle 比赛我觉得就是个特别吸引人的故事

28:27

because you had such high hopes going into this. We did. You were like, and I think this is one of

因为你一开始对这件事抱了特别高的期望。我们确实抱了。你当时就像，而且我觉得这是其中一种

28:33

those cases of like very nice, episemic rigor where you didn't just provide, oh, like, you know,

那种非常漂亮的 epistemic rigor 的情况，你没有只是提供，哦，就像，你知道，

28:40

here's a train test bullet with random split or something. You, you thought carefully about how to

这里有个 train test bullet，带着随机 split 或者什么的。你，你仔细思考了要怎么

28:44

organize this Kaggle competition, how to run it, and how to kind of evaluate fairly. Can you

组织这个 Kaggle 比赛，怎么运行它，以及怎么比较公平地评估。你能不能

28:50

talk about that process, the decisions you made, and then also the result?

谈谈那个过程、你做的决定，还有结果？

28:54

Yes, happy to do that. So we were not by a long shot the first group to try to evaluate whether

是的，很乐意。所以我们远远不是第一个尝试评估是否……的团队

29:02

you could use computers to make reasonable predictions or small molecules, binding to chemicals,

你可以用计算机来合理预测 small molecules 与化学物质的结合，

29:06

and then test those predictions. There are lots of papers that are doing this, like this

然后测试这些预测。有很多论文都在做这件事，比如这个

29:09

are real classic paper that came out of X, Kim and Google some years before before we found

是真正的经典论文，来自 X、Kim 和 Google，几年前，早在我们发现

29:13

at least that does this. We had run a lot of our own experiments on public data, and we had

至少有一篇能做到这一点。我们在 public data 上跑了很多自己的实验，而且我们

29:19

suspected that no one was really very good at it. And so we wanted to put together this contest,

怀疑其实没人真的很擅长这件事。于是我们想组织这个竞赛，

29:24

you know, and, and design the evals in particular in a way that could really test it across a range

你知道，而且，而且尤其是把 evals 设计成能真正测试它，覆盖一系列

29:29

of eval difficulties, because we were thinking about, okay, we're going to have a lot of people,

不同难度的 eval，因为我们在想，好吧，我们会有很多人，

29:34

hopefully a lot of people compete in this. Some are going to be very good, some are probably

希望很多人来参加这个。有些人会非常厉害，有些人可能

29:37

going to be beginners. We want to make sure that there are some nice easy wins, then like a bunch

会是初学者。我们想确保有一些轻松容易的小胜利，然后像一堆

29:42

of stuff in the middle that's like kind of hard, and then some things at the very, at the very top

中间的那些东西，算是有点难的，然后还有一些东西在非常、非常顶端的

29:46

end of what would be difficult here, what we imagined difficulty might be. And so to kind of unpack

困难部分的最顶端，也就是我们想象中难度可能到的地方。然后为了稍微拆解

29:51

that, we've talked a little bit already about how these libraries are combinatorial. So you're

这一点，我们已经稍微聊过这些 libraries 是 combinatorial 的。所以你会

29:55

going to have some smaller number of building blocks, but a large number of combinations.

得到数量较少的 building blocks，但会有大量的组合。

30:00

We got our hands on a chemical library from a wonderful sponsor, this guy named Shaji Liu,

我们从一个很棒的赞助商那里拿到了一份 chemical library，这人叫 Shaji Liu，

30:04

who runs a shop called Alphamon Shanghai, and the chemical library was about 130 million molecules.

他经营一家叫 Alphamon Shanghai 的店，而这份 chemical library 大约有 1.3 亿个 molecules。

30:11

And the center part of the molecule, you can imagine it kind of like a hub with three spokes,

而 molecule 的中心部分，你可以把它想象成有点像带三根辐条的中心枢纽，

30:17

and then the center hub was the same for every molecule. That's a, it's a particular chemical

然后 center hub 对每个 molecule 来说都是一样的。

30:21

motif called a trisine. And so what we did, we said, okay, so we got the center hub, then we got

那是一个，它是一种特定的 chemical motif，叫做 trisine。

30:26

these three spokes for the evaluation data set. Let's have a couple different ways to split this.

所以我们做的是，我们说，好吧，我们有了 center hub，然后我们有了这三根 spokes，用于 evaluation data set。

30:32

One way is that we could say, all right, the chemicals might have one shared building block,

我们弄几种不同的 split 方式吧。

30:37

one shared spoke, and then the other two spokes are different. Okay, that's one thing we could do.

一种方式是我们可以说，好吧，这些 chemicals 可能有一个共享的 building block，一个共享的 spoke，然后另外两个 spokes 是不同的。

30:42

So let's do that for some, and that was the easy stuff. Another one we can say is, all right, well,

好，这是我们可以做的一件事。

30:46

maybe we're not going to have any shared building blocks, but we do know that if they're going to come

所以我们对一些就这么做吧，那是最简单的部分。

30:50

from this chemical library, there's going to be the shared hub. And so we can imagine that some of

另一种我们可以说的是，好吧，那么，也许我们不会有任何共享的 building blocks，但我们确实知道，如果它们要来自这个 chemical library，那就会有共享的 hub。

30:54

the spokes might be similar, but they won't be exactly the same, and they won't be shared. So we

这些 spokes 可能相似，但它们不会完全一样，也不会被共享。所以我们

30:58

think this is going to be a little bit harder. That sounds great. And then for the third most

觉得这会稍微更难一点。听起来很棒。然后对于第三个最

31:02

difficult eval, we had a completely different chemical library that we had designed at least.

难的 eval，我们有一个完全不同的 chemical library，至少是我们设计出来的。

31:07

There was no common hub, no common building blocks, different attachment chemistries,

没有共同的 hub，没有共同的 building blocks，不同的 attachment chemistries，

31:10

different everything. And so we thought, okay, we're going to do, we're going to have these three

所有东西都不一样。于是我们想，好吧，我们要做，我们要有这三种

31:14

and different proportions. We're not going to tell anyone that's what we did. We're just going to say

而且比例不同。我们不会告诉任何人我们就是这么做的。我们就说

31:18

here that evals. And let's go see how it goes. And so we lost the kind of competition. It was two

这里有那些 evals。然后看看会怎么样。结果我们输掉了那种竞赛。那是两

31:24

years ago. And you guys can look it up if you want. We've written a lot about it. And then after 2000

年前。你们想的话可以查一下。我们写了很多相关内容。然后在 2000 之后

31:28

teams competed, the winner on did, you know, okay, on the stuff that where they were shared building

团队们竞争了，赢家做到了，你知道，好吧，就在那些他们有 shared building blocks 的东西上

31:35

blocks, depending on the target, we saw some boosts and performance for where there were no shared

取决于 target，在没有 shared building blocks 的地方，我们看到了一些 boosts 和 performance，但核心 hub 还是一样，tri-seam 还在。

31:39

building blocks, but the core hub was still the same, the tri-seam was still there. And then when you

然后当你去做最有挑战性的 eval 时，没有人表现得比 chance 更好。是啊，那个赢家大概就是 1,200 左右，只是一个提交了 Jupiter notebook 的随机的人，对吧？

31:43

go to the most challenging eval, no one did better than chance. Yeah, the like winner was like the

对。所以赢家是这位 Victor Schlappoff，他搞出了一个稍微不同的 approach。

31:48

like 1,200, just a random person who submitted a Jupiter notebook, right?

而且他的东西很好。但再一次，像其他所有人一样，没有比 chance 更好。

31:54

Yeah. So the winner was this kind of Victor Schlappoff, who put together a slightly different

前五名里有三个是那个 noisy tutorial notebook。

31:57

approach. And his was good. But again, like everyone else, no better than chance. Three of the top

但说清楚一点，他们通常是在 Kaggle competition 里的。

32:02

five were the noisy tutorial notebook. But just to be clear, they're usually with Kaggle competition.

五个是那个 noisy tutorial notebook。但说清楚一点，它们通常是跟 Kaggle competition 一起的。

32:09

There's like a public eval for you to train to test your algorithm, I guess. And then there's

就好像有个 public eval，让你训练来测试你的 algorithm，我猜。然后还有

32:16

the private one that's held out. That's right. And in the competition runs that and all you know

那个 held out 的 private 版本。没错。然后在比赛里跑的就是它，你只知道

32:20

is a score. Yep. So was your library included in the public eval? Okay. Yeah. It was hidden back

一个 score。对。所以你的 library 被包含在 public eval 里了吗？好。是啊。它被藏起来了，

32:26

to be to make it spicy, right? Okay. Yeah. And it was spicy. So what did you take from this?

就是为了让它更刺激，对吧？好。是啊。而且它确实很刺激。那你从中学到了什么？

32:33

What was the lesson? And I mean, first of all, what did you hope to get from this? You kind of

教训是什么？而且我是说，首先，你本来希望从这得到什么？你大概

32:37

talked about that. But then how did that sort of affect your thinking about like how do you proceed

聊过那个了。但后来那又是怎么影响你的想法的，比如你接下来怎么推进

32:42

forward? Yeah. Yeah. We had a couple hopes. And some, and some were fears. One thing I will be

往前？是啊。是啊。我们有几个期待。而且有些，有些是担忧。有一点我会

32:48

kind of about is, you know, we had released this data set that was a thousand times bigger than

有点在意的，就是，你知道，我们发布过这个 data set，它比……大一千倍

32:51

anything anyone had ever seen. And we gave it about a 5% chance that recent advancements and kind of

任何人见过的任何东西。我们给了它大概 5% 的概率，认为近期的进展以及那种

32:57

like co-folding approaches would mean that it would obviate our company entirely. Like we didn't

像 co-folding 这样的方法会意味着，它会让我们的公司彻底变得没必要存在。就像我们根本没

33:01

happen. No, didn't happen. No, but we were like, you know, there's like a small chance. Like we

发生过。不，没发生。不是，但我们当时想，你知道，还是有那么一点小概率的。就像我们

33:05

are really messing up here. And that didn't happen. We suspected more likely that folks weren't very

真的搞砸了。结果那并没有发生。我们更怀疑的是，大家其实不太

33:12

good at it and that the data might help. And so we're hoping to see that. One thing that anyway,

擅长这件事，而数据可能会帮上忙。所以我们希望能看到这一点。还有件事，反正，

33:16

you'd want to know that. You'd want to know that before he's in your life. Absolutely.

你会想知道这件事。在他进入你的生活之前，你会想知道这件事。当然。

33:21

Absolutely. And like another thing that we wanted to know is there's lots of different ways to

当然。还有，我们想知道的另一件事是，有很多不同的方式可以

33:25

represent molecules and give them to computers. You could use these things called smiles that do it

表示分子并把它们交给计算机。你可以用这些叫做 smiles 的东西来做这件事

33:31

some way. You could do 3D graphs. There's like weird spherical, like spherical convolutional nets.

某种方式。你可以做 3D graphs。还有一些奇怪的球形，比如 spherical convolutional nets。

33:37

People try to kind of goofy things to try to represent molecules in a useful way. And there'll be

人们会尝试一些有点搞怪的东西，想用有用的方式来表示分子。而且会有

33:41

lots of people in the field who will argue that like, oh, I think this way is better than this

这个领域里很多人会争论说，哦，我觉得这种方式比这种

33:44

way or whatever. And we did really encourage competitors, especially those who had a chemistry

方式更好，或者随便怎样。我们确实很鼓励参赛者，尤其是那些有化学

33:49

background, like let's try like some of these different representations. And so let's see how it

背景的人，比如咱们试试这些不同的 representations。然后看看它

33:52

goes. And we were grateful to find that like the thing that we were doing, which is just basic

会怎么样。我们很感激地发现，我们当时做的那个东西，也就是基础的

33:55

smiles, kind of beat everything else. Like it wasn't, it wasn't like a resounding thumping,

SMILES，基本上打败了其他所有方法。不是说，并不是那种压倒性的碾压，

33:59

but it was better enough for us to continue to using that approach. And smiles just, if by

但它足够好，让我们能继续沿用那种方法。而 SMILES 只是，如果按

34:04

recall correctly, is just basically a sequence of a textual sequence of strings. It doesn't encode any

如果没记错的话，它基本上就是一个 textual sequence，也就是 strings 组成的序列。它不编码任何

34:10

3D structure. It does. It's all emergent. It is a graph traversal translated into a text string.

3D structure。它确实会。全都是 emergent 的。它是一个被转换成一个 text string 的 graph traversal。

34:18

And so you can start the graph at different points. And so there's what they call canonical,

所以你可以从不同的点开始走这个 graph。所以就有他们所说的 canonical、

34:22

non-canonical smiles, depending on where you start. But yeah, it's an extremely simple

non-canonical smiles，取决于你从哪里开始。但 yeah，它是一个极其简单的

34:26

representation. And so yeah, we were thinking that, you know, we wondered if some of these fans,

representation。所以 yeah，我们在想，you know，我们想知道这些 fans 中的一些，

34:32

your ways of representing molecules might beat at least this particular problem. And they super

你们表示 molecules 的方式，至少可能在这个特定问题上胜出。而它们超级

34:35

didn't. So was your takeaway that you should go full force into this? Was this

没做到。所以你的 takeaway 是，你应该全力投入这件事吗？这

34:39

encouraging to you? Or was this discouraging you to do like, what was the collective emotional

让你受鼓舞吗？还是说这让你很泄气，像，大家整体的情绪是

34:46

reaction at least asked after that? I always, I've always wondered that. I mean, he kind of wrote

reaction 至少在那之后问了吗？我一直，我一直很好奇这个。我的意思是，他算是写了

34:49

about, you had several follow ups, but I'm just curious here. Yeah. Yeah. This is a great question.

关于……你有几个 follow-up，但我只是好奇。是啊。是啊。这是个好问题。

34:56

So like we, we are extremely bitter lessened, piled over here. And so our thought was, some other

所以就像我们，我们极其 bitter lessened，堆在这里。所以我们的想法是，一些其他的

35:02

was kind of delight, which is like, wow, this is much harder than we thought. That's kind of great.

有点让人惊喜，就好像，哇，这比我们想的难多了。这还挺棒的。

35:07

That's kind of great. Because if this problem is like all the other ones that we're aware of,

这还挺棒的。因为如果这个问题跟我们了解的所有其他问题一样，

35:13

if you can get pretty good signal out of a scalable measurement regime,

如果你能从 scalable measurement regime 里得到相当不错的 signal，

35:16

then you should just try to go get that. Yeah. Yeah. No, I mean, for us, it was,

那你就应该直接试着去拿到它。对。对。不，我是说，对我们来说，这……

35:20

it was pretty validating like, oh man, like we, this is an opportunity here.

这挺有验证意义的，就像，天哪，我们，这里有个机会。

35:24

Especially if you're scalable measurement regime is hard and you have specialized knowledge

尤其是如果你的 scalable measurement regime 很难，而且你得有专业知识

35:29

to do that. Absolutely. Absolutely. It was like, oh man, like this is at the time to like

才能做到。那当然。那当然。那感觉就像，天哪，这时候就该，像是

35:33

pour gas on this. Wow. Yeah. I mean, that's awesome. So then you were just like, okay,

给这件事火上浇油。哇。是啊。我是说，那太棒了。所以你就想，好吧，

35:39

let's full force. We're going to just start collecting data, collecting data, collecting data.

咱们全力开干。我们就开始收集数据，收集数据，收集数据。

35:43

Correct. Yeah. And so you did that for a while. And then what came next? And maybe first,

没错。对。然后你这么做了一段时间。那接下来呢？也许先说说，

35:50

also, what were the lessons you learned in the process? And then building to kind of your,

还有，你在这个过程中学到了什么经验教训？然后再慢慢讲到你的，

35:56

your first useful, like, yeah. Yeah. Yeah. There are a couple of things that were from the

你第一个有用的东西，怎么说呢，对。对。对。有几件事是来自那个

36:02

process that we thought were very, was very, very useful. One thing that we realized is that the

过程，我们觉得非常，非常非常有用。我们意识到的一件事就是，

36:07

competition was based on three protein targets that were fairly well studied. It became clear to us

比赛是基于三个研究得相当充分的 protein targets。我们越来越清楚，

36:12

that we really wanted to, you know, for lots of reasons, but including these that, you know,

我们真的很想，你知道，原因有很多，但也包括这些，你知道，

36:16

getting performance across novel proteins, probably meant that we'd have to do a whole lot more.

要在 novel proteins 上都做出表现，可能就意味着我们得做多得多的事情。

36:21

So, okay, so that, you know, we were always motivated to do that, but that really underscored

所以，好吧，所以，你知道，我们一直都有动力去做，但那真的凸显了，

36:24

like, oh, man, we're going to have to go. We're going to have to make a lot of proteins, aren't we

就像，天哪，我们得上了。我们得做很多 proteins，对吧

36:27

here? And so that was part of it. Part of it was thinking that we had this very large number of

这里？所以那是一部分原因。还有一部分原因是，我们想着我们有这么大量的

36:33

small molecules, like this hundred million compound library. That didn't help as much to us as

small molecules，比如这个一亿规模的 compound library。那对我们的帮助没那么大，不如

36:38

better characterizing a smaller one. And so that was also like a strategic choice that we made.

更好地 characterizing 一个更小的库。所以这也是我们做的一个战略选择。

36:43

This is actually quite important because you have a fixed budget. And so you can either, let's say,

这其实相当重要，因为你的预算是固定的。所以你要么，比如说，

36:48

sequence or sequence, it's been more money on sequencing, getting higher resolution for what

sequence 或者 sequence，就是把更多钱花在 sequencing 上，拿到更高 resolution，

36:53

you already have for a smaller set, or you can get more molecules. And they hope that the noise

针对你已有的更小的集合，或者你可以获得更多 molecules。而他们希望 noise

36:59

sort of averages out or something. That's correct. Like having a fixed budget really does affect

差不多能 average out 之类的。没错。就像预算固定真的会影响

37:03

like your strategic choices. And for us, we wanted to make sure that we're getting a lot of

你的战略选择。而对我们来说，我们想确保我们能得到很多

37:06

variation across both axes, as compared to like one axis and like a handful of proteins.

在两个 axes 上的 variation，而不是像只在一个 axis 上，以及只有少量 proteins。

37:11

I wish two axes. So that would be like the number of chemicals against the number of proteins.

我希望能有两个 axes。所以那就像是 chemicals 的数量对 proteins 的数量。

37:15

Right. And so you can imagine like, oh, I'm going to do three proteins in a billion chemicals,

对。所以你可以想象，哦，我要在十亿个 chemicals 里做三个 proteins，

37:20

or maybe I should do like thousands and single the gemillions, you know, or some other combination

或者也许我该做像几千个，然后 single the gemillions，你知道，或者别的什么组合

37:26

of those. Single the gemillions being probably one of the tiniest cells ever constructed.

那些的。Single the gemillions 可能是构造过的最小的 cells 之一。

37:29

It's pretty small. The smallest one I've seen is 50,000 molecules.

它挺小的。我见过最小的一个是 50,000 molecules。

37:32

Oh, okay. Yeah. I guess you can make a custom dull if you want to do like SAR.

哦，好吧。对。我猜如果你想做类似 SAR 的话，你可以做一个 custom dull。

37:35

You can do whatever you want. Yeah, yeah. Yeah, it's a, well, you have the technology just do it.

你想怎么做都行。对，对。对，这是个，嗯，你都有技术了，直接做就行了。

37:40

Yeah, yeah. You do building blocks. Yeah. Yeah, yeah. You can do whatever you want.

对，对。你做 building blocks。对。对，对，你想怎么做都行。

37:43

And so that's the nice thing is you have enormous control over, over like, you know,

所以好处就是，你能对……对，就像，你知道，有极大的控制权，

37:47

both like the library size and characteristics and like this barcode issue that I talked about.

比如 library 的大小和特征，还有比如我提到的这个 barcode issue。

37:51

You can expand that kind of stuff. And we've put a lot of thought into, you know,

你可以扩展那类东西。而且我们花了很多心思，你知道，

37:54

kind of bells and whistles in our own delts to truly boost, you know, kind of like a performance

就是一些花里胡哨的功能，在我们自己的 delts 里，真正去提升，你知道，有点像 performance

37:58

and signal over noise and all this stuff. But getting back to the kind of like the cargo

以及 signal over noise 之类的所有东西。但回到那种，像是 cargo

38:02

competition lessons, you know, was also clear to us that models could think about memorizing

competition 的教训，你知道，对我们来说也很清楚，models 可能会考虑 memorizing

38:10

individual building blocks. And that was boosted performance that was something that seemed

单个的 building blocks。而那提升了 performance，这一点似乎

38:13

pretty clear. It seemed pretty clear that they might be able to get an idea through probably

相当清楚。看起来相当清楚的是，它们也许能获得一个想法，可能是通过

38:19

through memorization that if you have a very similar building block, I might behave similarly.

memorization，如果你有一个非常相似的 building block，我可能会表现得类似。

38:22

So that also some of those performance metrics. And then also that, you know,

所以那也提升了一些 performance metrics。然后还有就是，你知道，

38:26

when I said that it's pretty far away, they, I mean, at the current regime, at that regime,

当我说那还差得很远时，他们，我是说，在当前的 regime 下，在那个 regime 下，

38:30

they simply could not do it. And when we release these results, basically saying, hey, you know,

他们根本做不到。然后当我们发布这些结果时，基本上就是在说，嘿，你知道，

38:35

no one did better than chance. We did get some low back where folks were like, well,

没人做得比 chance 更好。我们确实收到了一些 low back，有人就说，好吧，

38:40

did anyone good compete, you know, in your del competition? Like maybe it was just bad.

有没有厉害的人参赛，你知道，参加你们的 del competition？就好像也许只是太糟糕了。

38:44

Maybe it's all first year grad students. And so as a result, we just released all of the data

也许参赛的全是一年级研究生。所以结果就是，我们干脆把所有 data 都

38:49

entirely. And so if anyone out there thinks they're awesome, then like please show us,

完全公开了。所以如果外面有谁觉得自己很厉害，那就，拜托，给我们看看，

38:52

because like we want this problem solved. But we haven't seen anyone do create, you know,

因为我们就想解决这个问题。但我们还没看到有人做出来，你知道，

38:57

perform great on that stuff yet. And we say all the data you have, including all the raw read

在那上面表现得很棒。而且我们说，你们有的所有 data，包括所有 raw read

39:02

accounts and everything. So, so you can go through and say, oh, it's not just, you know,

账号之类的所有东西。所以，所以你可以从头到尾看一遍然后说，哦，这不仅仅是，你知道，

39:06

processing error. Like you can go from the raw data and use it. You could take our, yeah,

processing error。比如你可以从 raw data 出发并使用它。你可以拿我们的，对，

39:10

you could take our binary labels. You could do it yourself. All the evals, everything. It is,

你可以拿我们的 binary labels。你可以自己做。所有的 evals，所有东西。它，

39:15

it is all on Polaris, which is like a nonprofit group that was run by recursion. It's there for

它全都在 Polaris 上，Polaris 就像一个由 recursion 运营的非营利组织。它在那里供

39:19

anyone to use. Please use it. You know, like we, we intentionally released it as, you know,

任何人使用。请使用它。你知道，就像我们，我们是有意把它发布出来的，你知道，

39:23

hoping that community would would make good use of it. Do you think though the structure of this

希望社区会，会好好利用它。不过你觉得，这个

39:26

data set was fundamentally limited in a way that maybe without some external augmentation,

data set 从根本上受到了限制，也许没有一些 external augmentation，

39:31

you're not going to be able to do any better than what people have pretty much already done.

你就没法做得比大家基本上已经做到的更好。

39:35

I think if you had really perfect structural understanding or like really good structural models,

我觉得如果你真的有特别完美的 structural understanding，或者就特别好的 structural models，

39:39

then you ought to be able to do okay on this, right? Because if we, we all have models that really

那你应该能在这个上面做得还不错，对吧？因为如果我们，我们都有 models，真的能理解，真正理解

39:43

understand, truly understand like the physics of the interactions between the amino acids of

small molecules 的 amino acids 之间的 interactions 的 physics，

39:47

the small molecules, then it doesn't really matter what the distribution of the small molecule

那么 small molecule training set 的 distribution 会是什么样，其实就不太重要了，如果它们真的很擅长这个的话。

39:51

training set will look like if they're really good at this. You know, and I think when we see,

你知道，而且我觉得当我们看到，当我们看到 structural model predictions，然后有人说，

39:54

when we see structural model predictions and folks say, oh, well, you know, I've got this time split

哦，那个，你知道，我在 PDV 上有这个 time split，比如说，你知道，我，你知道，

39:59

on the PDV, for example, and you know, I, you know, our models never seen these small molecules

我们的 models 从来没见过这个 protein 的这些 small molecules，

40:03

of this protein before and then hopefully we can, you know, kind of use this understanding to

然后希望我们能，你知道，算是用这种 understanding 来...

40:07

make reasonable folds. Like that's very much in the spirit of being able to make the predictions

做出合理的 folds。就像，这非常符合能够做出 predictions 的那种精神

40:11

on this stuff because no one has seen these models, especially our internal molecules like no one

在这些东西上，因为没人见过这些 models，尤其是我们内部的 molecules，就像没人

40:14

has seen those before. That's cool. So Kaggle competition, you went ahead and you learned, you learned

之前见过那些。那很酷。所以 Kaggle competition，你就去做了，然后你学到了，你学到了

40:20

some lessons, you went full force data collecting, regroup, committed to scaling. Yep. All right. So

一些教训，你全力投入 data collecting，重新整顿，下定决心做 scaling。是的。好吧。那么

40:27

then you, I think something like six months later, you released, or you had a blog post about your

然后你，我想大概六个月后，你发布了，或者说你写了一篇博客文章，关于你的

40:33

model, Hermes, which is, yeah, can you tell us about that? Sure. Some of the learnings there, and I

model，Hermes，它是，呃，能跟我们讲讲那个吗？当然。里面的一些收获，还有我

40:39

think both, and especially how you evaluated that, I think, and all the care you which went into that

觉得两方面都是，尤其是你如何 evaluate 那个，我觉得，还有你为那个 model 投入的所有用心

40:46

model. Hermes is a model that is the original brainchild of the data scientist, at least named Mike's

Hermes 是一个 model，它最初是那位 data scientist 的原创点子，至少是那个叫 Mike 的 data scientist 的

40:50

Klein Soster, Max Klein Soster, and also Sean Francis Lyon and, and Andrew Levins, of course,

Klein Soster、Max Klein Soster，还有 Sean Francis Lyon，以及当然还有 Andrew Levins，

40:54

the CTO there. What the thinking was originally was that we had a bunch of data, and we were thinking

那里的 CTO。最初的想法是，我们有一堆数据，而且我们在想

41:00

about how to split it in different like training regimes. And so we wanted to come up with a model

怎么把它拆成不同的、类似 training regimes 的东西。于是我们想做出一个 model，

41:05

that was really, really dead simple, so that we could do lots of potential experimental iterations.

它特别特别简单，这样我们就能做很多潜在的 experimental iterations。

41:09

It wasn't the purpose of it, it wasn't to be a state of their models. Yeah, you know, we want to

这不是它的目的，不是要成为他们 models 的那种 state。对，你知道，我们就想

41:12

try stuff fast. And so we got Hermes, and then the basic architecture of Hermes is two different

快速试东西。于是我们做出了 Hermes，然后 Hermes 的基本 architecture 是两个不同的

41:18

beddings. One of them is a evolutionary scale modeling at ESM embedding. That's a protein embedding

embeddings。其中一个是 evolutionary scale modeling 的 ESM embedding。那是一个 protein embedding，

41:25

that comes from a group that was in meta, and then went off to start the road company, and then they

它来自一个最初在 meta 的团队，后来他们出去创办了 road company，然后他们

41:28

got, and they got picked up by a transographer more recently. And so we had Alex Reeves on the

对，而且它们最近被一个 transographer 接手了。所以我们就请了 Alex Reeves 上

41:32

podcast. Yeah, you want to hear more. Yeah, yeah. If you want to hear more. Yeah, thank you, Alex,

podcast。对，你想听更多。对，对。如果你想听更多。对，谢谢你，Alex，

41:39

if you're listening. So we did that for the protein side, and then we had a different embedding called

如果你在听的话。所以我们在 protein 这边做了那个，然后我们有一个不同的 embedding，叫做

41:44

Kenberta, which is basically like a next token predictor chemistry, kind of a bit of this side.

Kenberta，基本上就像是一个 chemistry 的 next token predictor，有点像这一侧的东西。

41:49

Something which predicts miles, or yeah, just predicts, yeah, smiles tokens. And so these

就是那种能预测 SMILES 的东西，或者说，对，就是预测 SMILES tokens。所以这些

41:53

are very, very two simple beddings, and then we just put them together with cross attention,

是非常非常简单的两个 embedding，然后我们只是用 cross attention 把它们拼在一起，

41:57

and that was it. And then we used it to predict a binary label. Simple as possible.

就这样。然后我们用它来预测一个 binary label。尽可能简单。

42:00

It is incredibly simple, very pretty small. And then we trained some of our own data on it.

它简单得不可思议，非常小。然后我们在它上面用我们自己的一些 data 训练了一下。

42:05

Actually, not the simplest. You could have done a MLP after those, right, instead of cross-stition.

其实，也不是最简单的。你本可以在那些之后接一个 MLP，对吧，而不是 cross-stition。

42:11

So why attention? The thinking with the attention was had to do with some other experiences that

所以为什么用 attention？关于 attention 的想法，其实和我们有过的另外一些经验有关，那些经验

42:19

we'd had, kind of with graph, convolutional, like graph models for protein structure. And one

我们有过，算是跟 graph、convolutional 有关的，比如用于 protein structure 的 graph models。还有一个

42:25

thing that we'd noticed is if you've got these sort of graphs, and there's a couple models that do

我们注意到的一件事是，如果你有这类 graph，而且有几个模型会做

42:29

this, it's other work that we had done, we had scrambled the positional coordinates of those,

这个，那是我们做过的其他工作，我们打乱了那些的 positional coordinates，

42:33

and performance for some of those models more or less stay the same. And that made us think that

而其中一些模型的 performance 或多或少保持不变。那让我们觉得，

42:37

attention was super critical. And so we ended up just making sure that the model was at least

attention 超级关键。于是我们最后就是确保这个模型至少

42:42

in our, at least what we hope, just that it would be looking at the different parts of the

在我们，至少是我们所希望的，就是它会去看不同的部分，那个

42:44

proteins and small molecule structures that could be useful here. So this is, I think, oftentimes

在这里可能会有用的 proteins 和 small molecule structures。

42:48

known as the Picasso problem in image generation, where we'll say CNN's, for example, would,

所以我觉得，这个问题在 image generation 里常常被称为 Picasso problem；比如说 CNN，曾经有一阵子很出名，你可以把一张图片的 features 打乱，像 Picasso 那样，比如把鼻子翻过来，或者把眼睛转一下之类的。

42:54

or famous for a while, where you could scramble just the features of a picture, like Picasso,

而 CNN 还是会认出来，哦，这是 Picasso，因为它只是抓到了单个的 local features，而不是某种 global structure。

43:00

like flip the nose around or just rotate the eyes or something. And the CNN would still

而这个问题在 small molecule drug discovery 里已经存在一段时间了，它常常会抓到 local motifs，然后对 topology 完全没有全局理解。

43:06

pick, oh, this is a Picasso, because it just picked up on individual local features, rather than

所以尤其是像 Adele 这种东西，你有这种 combinatorial

43:11

some sort of global structure. And this has been a problem in small molecule drug discovery for a while,

某种 global structure。而这在 small molecule drug discovery 里已经是个问题有一阵子了，

43:16

where oftentimes it picks up on local motifs, and then it doesn't have any global understanding

它常常会捕捉到 local motifs，然后它并没有任何 global understanding

43:22

of the topology. So especially for something like Adele, where you have this combinatorial

对 topology 的。所以尤其是像 Adele 这样的，你会有这种 combinatorial

43:26

structure of individual building blocks, you could imagine the model learning each building block

各个 building block 的结构，你可以想象模型学习每个 building block

43:30

independently, and then essentially just decomposing into the sum of three building blocks or something,

独立地，然后本质上就只是分解成三个 building blocks 的和之类的，

43:36

or I mean, is a simple concept. Yeah, yeah. So cross-attention in some sense forces certain

或者我是说，这是个简单的概念。对，对。所以 cross-attention 在某种意义上会迫使某些

43:42

molecules to interact with like certain residues in our protein, or at least maybe not forces,

molecules 去跟我们 protein 里的某些 residues 相互作用，或者至少也许不是迫使，

43:47

but provides the opportunity to think about it. It's an inductive bias, yeah. Yeah.

而是提供了去思考它的机会。这是一种 inductive bias，对。对。

43:51

Kind of what you're getting at is, you know, we were thinking about how to evaluate this. And so

你想说的有点像是，你知道，我们当时在想怎么评估这个。所以

43:56

we had done some of the similar tricks that we did over with the cow competition where we had

我们做了一些类似的技巧，就像我们在 cow competition 里做的那样，当时我们有

44:00

training sets on data that we generate internally, and then we'd look to chemical libraries that

基于我们内部生成的数据的 training sets，然后我们会去看 chemical libraries，那些

44:04

were further away. But they also we wanted to think about splits to kind of probe Hermes behavior.

更远。但他们也，我们也想考虑一些 splits，来某种程度上 probe Hermes 的行为。

44:11

And so because Adele's, I have this combinatorial approach, you could do the same thing where you might

所以，因为 Adele's，我有这个 combinatorial 方法，你也可以做同样的事，你可能会

44:16

share certain building blocks. And so we do that. And of course, performance is better, which we

共享某些 building blocks。于是我们就这么做。当然，performance 更好，而我们

44:19

suspect is from memorization. But the other thing is you can imagine that there are building blocks

怀疑这是来自 memorization。但另一件事是，你可以想象，有一些 building blocks

44:24

that are very similar to one another that if you're just trying to split on the identity of those,

彼此非常相似，如果你只是试图根据它们的 identity 来 split，

44:28

and then you might have things that are close enough that that memorization or some other,

然后你可能会遇到一些东西，它们足够接近，以至于那种 memorization 或某种其他

44:33

you know, cheap mode might, you know, convert predictive power. And so we came up with a scheme

你知道，廉价模式可能会，你知道，转化 predictive power。于是我们想出了一个方案

44:38

where we then did basically clustering of these different building blocks. We had building

然后我们基本上对这些不同的 building blocks 做了 clustering。我们有 building

44:44

block clusters on similarity. And that way we would just avoid entire clusters from being

根据 similarity 对 block 进行 cluster。这样我们就能避免整个 clusters 被

44:49

overlap between test and train. And so when you do that, then, you know, the performance does drop.

test 和 train 之间的 overlap。所以当你这么做的时候，你知道，performance 确实会下降。

44:55

And so one thing that you might be alluding to is we have a, we use lots of, lots of reasonable

所以你可能暗示的一件事是，我们有一个，我们用了很多，很多合理的

45:01

baselines here that we're really worried about. And so for almost everything that we do, we'll use

baselines，这些是我们真的很担心的。所以对于我们做的几乎每一件事，我们都会用

45:05

like XG boost, right? You should do something like that just to see how it's going. Yeah, yeah, yeah,

比如 XG boost，对吧？你应该做点类似的事情，看看进展如何。对对对，

45:09

right? Yeah, every, yeah, the classic young machine learning practitioner, they've fed six

对吧？对，每个人，对，经典的年轻 machine learning 从业者，他们喂了六

45:14

butts off doing the favorite model. I did someone walks by. It does like X, you know,

个屁股都掉了，做最喜欢的 model。我做了，有人路过。它就像 X，你知道，

45:18

XG boost rate. Yeah, XG boost rate of four is really like, oh yeah, yeah, that does

XG boost rate。对，XG boost rate of four 真的就像，哦对，对，那确实

45:22

beat like six months of work. Oh, yeah, yeah. Yeah. And so we, and so we did these different

超过了大概六个月的工作量。哦，对，对。对。然后我们就，我们就做了这些不同的

45:28

like split regimes to try to evaluate how well things do relative to like XG boost performance.

像是 split regimes，来试着评估这些东西相对于 XG boost performance 做得怎么样。

45:34

And so one thing that was really gratifying to see is that you see XG boost drops deadly

所以有一件事真的让人很欣慰，就是你看到 XG boost 掉得特别惨

45:40

as you get further and further away with these building block cluster approaches.

当你用这些 building block cluster approaches 越走越远的时候。

45:43

It's like, if you share two building blocks, it's a certain level. If you only share one, it gets worse.

就像，如果你共享两个 building blocks，它就在某个水平。如果你只共享一个，它就会变差。

45:48

If you share one, one, you know, only in none, it's worse than if you do the building block cluster

如果你共享一个，一个，你知道，只有在 none 的时候，它比做 building block cluster 还差

45:52

things, even worse than that. And then we compared that to the Hermes performance, which was pretty steady

这些东西，甚至比那还差。然后我们把它和 Hermes performance 做了比较，它相当稳定

45:56

across all these different regimes. I see. Yeah. So it made us think like, oh, yeah, you know, you've

在所有这些不同的 regimes 下。我明白了。对。所以这让我们想，哦，对，你知道，你已经

46:01

really got to get pretty far out of distribution for XG boost to really tank. Yeah, and that's

真的得跑到相当 out of distribution，XG boost 才会真的崩掉。对，而且这

46:05

something that we haven't, you know, we haven't, it's in a couple plots and some work that we publish.

是我们还没，你知道，我们还没——它就在几张图和我们发表的一些工作里。

46:08

We haven't yelled about it too much. I thought you didn't have a block post like XG boost is all

我们没怎么大声嚷嚷这件事。我还以为你没有一篇像“XG boost 就是你要的全部”那样的 blog post。

46:13

you. Yeah, running the splits is something that we've put a lot of thought into trying to figure out

对，跑 splits 这件事我们花了很多心思，想弄清楚

46:20

like, okay, how do you stop your models from really, you know, memorizing your room or anything?

比如说，好吧，你怎么阻止你的模型真的，你知道，把你的房间之类的东西背下来？

46:24

This might be a good time for a little bit of a detour, but referencing, I know,

这可能是个稍微岔开一下话题的好时机，不过提到这个，我知道，

46:28

there want to be a blog post, which I found absolutely hilarious for my own personal experience.

有一篇 blog post，以我自己的个人经历来说，我觉得它超级好笑。

46:32

I have seen machine learning models fake like pick up on so many weird detractors. Uh-huh. And

我见过 machine learning models 会，像，抓住好多奇怪的 detractors。嗯哼。而且

46:38

you're one, you're your blog post about molecule ownership or authorship was a guy that got a really

你是其中一个，你关于 molecule ownership 或 authorship 的 blog post 讲的是个哥们儿，他有个真的

46:45

fun story. Yeah, several hands post. Yeah. Yeah. So many years ago, uh, my co-founder Andrew

很好玩的故事。对，several hands post。对。对。好多年前，呃，我的联合创始人 Andrew

46:50

Levin's and I were at recursion looking at a small molecule on a screen. And then a medicinal chemist

Levin 和我在 Recursion，看着屏幕上的一个 small molecule。然后一位 medicinal chemist

46:55

walked by as this guy, uh, this guy in my cookery see, uh, and Mike was like, hey, looks like a

走过，就是这位，呃，这位在我 cookery see 里的家伙，呃，然后 Mike 说，嘿，看起来像个

47:01

Stuart trivert molecule. And then we are like, oh, no, you can tell just by walking by and be like,

Stuart trivert molecule。然后我们就想，哦，不，你光走过去就能看出来，然后说，

47:06

who made this? That's horrifying. And then that haunted Andrew in particular for like many years.

这是谁做的？太可怕了。然后这事尤其困扰了 Andrew 好多年。

47:11

Like it was like six or seven years. And we joked, uh, you know, I don't know, every couple of

大概有六七年吧。我们还开玩笑，呃，你知道，我也不知道，每隔几

47:15

months we're out for beers like, man, we really shouldn't figure out like how bad this problem is.

个月我们出去喝啤酒的时候就会说，哥们儿，我们真的不该去搞清楚这个问题到底有多糟。

47:19

Why is that a problem? Okay. So it's a problem where it's very easy to imagine. You're going to

为什么那是个问题？好。

47:24

try to train models to understand chemistry. So you're going to get the data from lots of groups

所以这是个很容易想象的问题。

47:30

because it's very tempting to get data from lots of groups. And if the models can identify the chemist

你会尝试训练 models 去理解化学。

47:36

by looking at the molecule alone, then maybe the models can infer chemist intent, which means maybe

所以你会从很多组那里拿数据，因为从很多组拿数据这件事非常诱人。

47:42

they don't learn anything at all. Well, they learn something. They learn something. Yeah. They don't

而如果 models 光看分子就能认出是哪个化学家，那也许 models 能推断出化学家的意图，这意味着它们可能根本什么都没学到。

47:47

learn. I'm not learning the thing you're trying to get to learn. Yeah. And so we were really afraid of

嗯，它们学到了点东西。它们学到了点东西。对。它们没学到。

47:51

this. They learned that certain chemists are better than other ones. Yeah. We learned all kinds of

我没学到你想让它学会的那个东西。对。

47:54

stuff. But maybe to sit to set this up for the audience. Um, so we went out and like we can

所以我们真的很怕这一点。

47:58

not republish this, you know, over Christmas, you can look it up. We have a leaderboard. We have a

别转载这个，你知道的，圣诞期间你可以去查一下。我们有一个 leaderboard。我们有一个

48:02

goofy online game you can play. Um, but we downloaded all chemicals from all chemists. It

很搞怪的在线小游戏可以玩。呃，但我们下载了所有化学家的所有 chemicals。它

48:08

kumbled, which is the biggest repository of these things and their activities. Every chemist

kumbled，是这些东西和它们活动的最大 repository。每个化学家

48:13

who published more than, I think it was 30 papers and 600 molecules. So get on molecules and all

发表超过——我觉得是 30 篇论文和 600 个 molecules。所以搞 molecules 和所有

48:17

the stuff. It's 1815 chemists. And then we tried to class fire to predict a chemist from looking

这些东西。一共 1815 个化学家。然后我们试着用 classifier 来预测化学家，只靠看

48:24

at the molecule alone. And we got a top five accuracy of 60%. It's a very, very easy problem.

molecule 本身。我们得到了 60% 的 top 5 accuracy。这是个非常非常容易的问题。

48:30

Yeah. Yeah. It's super easy for the models to do. Where they get confused is where you have two

是啊。是啊。对 models 来说这超级容易。它们会搞混的地方是，当你有两个

48:35

chemists who run the same publications and work the same institution. And so we think it's very

化学家发表在同一批出版物上，并且在同一个机构工作。所以我们觉得这非常

48:40

likely is that there is a distinctive chemical signature style that individuals may have when

很可能是，存在一种独特的 chemical signature style，每个人可能会有，当他们

48:46

they make molecules. And then when those individuals have students or trainees, then they adopt

制造 molecules 时。然后当这些人有学生或学员时，他们会吸收

48:51

some fraction of that. And so that was, that seemed pretty, pretty clear to us. And so that

其中的一部分。所以那对我们来说，似乎非常、非常清楚。所以那

48:56

was horrifying enough of in its own. And this is a data set that many, many academic projects,

本身就够吓人了。而且这是一个 data set，很多、很多学术项目、

49:02

many PhDs have trained on, tried to predict on fine patterns. And this is really like a

很多 PhD 都在上面训练过，试图去 predict fine patterns。而这真的就像是

49:09

major data set for the field. Yeah. Like, if you are not generating your own data, you're using

这个领域一个很重要的 data set。对。就像，如果你不是自己生成数据，那你就是在用

49:15

this. And then if you are generating your own data, then you are probably victim to this

这个。而如果你是在自己生成数据，那你很可能也会中招，

49:20

because your organization probably has multiple chemists who have their own distinctive chemical

因为你的组织里很可能有多个化学家，他们各自都有自己独特的 chemical

49:24

styles and that they are using their styles to go after different types of projects.

风格，而且他们是在用自己的风格去追不同类型的项目。

49:28

But the key here, I think, is that those chemists also tend to solve similar problems. Like,

但我觉得，这里的关键是，那些化学家也往往会解决类似的问题。就像，

49:35

I might have a career here. And I have a style, but I also am working on a class of problems.

我可能在这里有一份职业。我有自己的风格，但我也在做一类问题。

49:42

Right. And so that your style is also you're learning the style instead of the problem.

对。所以你的风格也在于，你学的是风格，而不是问题。

49:48

You're learning the style, you're learning the sociology of the problem. And the other thing

你在学风格，你在学这个问题的社会学。还有另一件事

49:51

that you're learning is that when you're a chemist at these organizations, you know, ordering

你学到的是，当你在这些组织里做化学家时，你知道，下单让

49:56

a molecule to be manufactured and tested is kind of a picking. Right. So what will happen is

一个分子被制造和测试，其实有点像挑选。对。所以接下来会发生的是

50:01

you will make many, many choices of things you will not pursue and you do not write those down.

你会做出很多很多选择，决定哪些东西你不去追，而且你不会把那些写下来。

50:07

Oh, yeah, it will be dumb to do that. And it's like, yeah, that's their motivation.

哦，对，那样做太蠢了。而且就像，对，那就是他们的动机。

50:11

Like, it's like our team couldn't just just wouldn't make it. Like I could say, like, we could do

就像，就像我们团队根本没法，就是做不出来。就像我可以说，像，我们可以做

50:15

this, but then what they would cost too much or the, the lab team will say, the chemistry team

这个，但那样的话，它们成本会太高，或者，那个，实验室团队会说，化学团队

50:19

will say, now we can't make it now. Or, you know, we don't have the building blocks on.

会说，现在我们现在做不出来。或者，你知道，我们连 building blocks 都没有。

50:23

I don't think it's going to work. Like there's all these reasons why you might not make a choice.

我觉得这行不通。就像，有各种各样的理由让你可能不做出某个选择。

50:28

And then not making that choice is a fingerprint that is going to land on what does get made.

然后，不做出那个选择，会像一个指纹一样，落在最终真正做出来的东西上。

50:34

And then, you know, when you talk to some of this little chemists, some of them are,

然后，你知道，当你跟这些小化学家中的一些人聊的时候，他们有些人，

50:37

you know, blessed and they're superstitious people. And some of them will say like, oh,

你知道，是有福气的，而且他们是迷信的人。然后他们有些人会说，哦，

50:42

this is red. Yeah, I bring this up like every single time. This one chemists are some of the most

这是红的。对，我基本上每次都提这个。这个嘛，化学家们是我见过最

50:46

superstitious people I've ever met. Right. And so, and so it's like, oh, well, this building block

迷信的人之一。对。然后，然后就像，噢，好吧，这个 building block

50:51

worked in this, in this program that made it to the clinic, I'm really fond of it.

在这个，在这个进了临床的 program 里管用，我特别喜欢它。

50:54

Maybe it helps us side building. So I'm going to put it in things that I like. And so then

也许它能帮我们做 side building。所以我就把它放进我喜欢的东西里。然后

50:58

your model will learn like, oh, the chemists likes this because they put so that means it's

你的 model 就会学到，噢，化学家喜欢这个，因为他们把它放进去了，所以这意味着它

51:02

probably working. And so it learns all this weird stuff. And so like, we were very afraid of this.

可能管用。然后它就学到所有这些奇奇怪怪的东西。所以，我们特别怕这个。

51:08

And so we took, you know, we ran this, this model to identify who the chemist was.

然后我们就，你知道，我们跑了这个，这个 model 来识别到底是哪个化学家。

51:14

And then we chose to represent the molecules, not with any chemical structures or formulas at all,

然后我们选择表示 molecules 的时候，完全不用任何 chemical structures 或 formulas，

51:20

but just a vector of the probability of creator. So now it's this 1800 element, you know,

而只是创建者 probability 的一个 vector。所以现在它就是这 1800 个 element 的，你知道，

51:25

vector and it's got some probability of who made it. And you can use that representation and then

vector，而且它带有一些关于是谁做的 probability。然后你可以用那个 representation，接着

51:30

train a model to predict activity from Campbell. And it does as well as if you use chemical formulas.

train 一个 model 来 predict 来自 Campbell 的 activity。而且它的效果跟你用 chemical formulas 一样好。

51:37

And if you combine them according to sociology only, yes, if you combine them, it doesn't,

而如果你只按照社会学把它们结合起来，是的，如果你把它们结合起来，它并不会，

51:42

we're meaningfully better. And so yeah, we are, you know, we really think that it's learning a lot

我们明显更好。所以是的，我们，你知道，我们真的觉得它学到了很多

51:47

of sociology. It's not learning ground truth. It's just learning like the choices and the

社会学。它不是在学 ground truth。它只是在学，像是那些选择和

51:50

fingerprints of the people who were trying to develop these molecules. And it's not a knock on them,

那些试图开发这些 molecules 的人的指纹。而且这不是在贬低他们，

51:54

right? I mean, that's that's not what they're trying to do. They're trying to, you know,

对吧？我的意思是，那不是，那不是他们想做的事。他们想，你知道，

51:57

get two drugs by making the fewest number of chemical series possible. Of course, they're going

用尽可能少的 chemical series 搞出两种药物。当然了，他们会

52:01

to do it like that. But when you train models on the exhaust, you might be in trouble.

就这么干。但当你拿 exhaust 来 train models 的时候，你可能就会遇到麻烦。

52:05

So this means that this is not a great machine learning training set.

所以这意味着，这不是一个很好的 machine learning training set。

52:09

We do not believe so. It's funny in my own, in my own life, you know, I've had many of these

我们不这么认为。有意思的是，在我自己，在我自己的人生里，你知道，我经历过很多这种

52:15

similar lessons. It's well known that things like molecular weight will just, you know,

类似的教训。众所周知，像 molecular weight 这种东西，你知道，就是

52:20

all you have to do is sort by molecular weight. That's like one of the single, single biggest

你只需要按 molecular weight 排序就行了。那差不多是你能做出的单个、单个最大的

52:24

predictions you can possibly have. Like wasn't it cast 16 or it was tied for first place?

predictions 之一。就像它不是 cast 16 吗，还是并列第一？

52:29

Yeah, yeah. It's like it was really close. Yeah. Yeah. So all these people pulling out their

对，对。就像它真的非常接近。对。对。所以所有这些人都掏出他们的

52:34

fancy machine learning models, you just sort the molecules by heavier molecules, better,

花哨的 machine learning models，你只要把 molecules 按更重来排序，越重越好，

52:39

better binders. Yeah. Yeah. Yeah. Yeah. So. Yeah. But this is not a intrinsic property to chemistry.

更好的 binders。对。对。对。对。所以。对。但是这不是 chemistry 的 intrinsic property。

52:45

This is a property of the data set. Yes. Yes. This is a property of the sociology of what gets

这是 data set 的一个属性。对。对。这是关于什么被做出来、什么被测试的社会学的一个属性。嗯。而且就像 medicinal chemists 的工作方式，更重的 molecules 很多时候更有可能 bind。这是因为他们看到改进会影响其他属性。这是其中的一部分。而且你可能还有些部分要做。基本上就像你也有一个 chemist。他们找到一个 hit，然后开始优化它，然后他们会说，哦，我有个假设，这就是 binding site。所以让我加这个 functional group 在这里 bind。好。然后在一个给定的 pocket 里，那些恶性的 amino acids 在驱动 binding 行为，也许你试着把它们都独立地破坏，然后你每次重做 screens，并且对什么有个好的感觉。

52:49

made, what gets tested. Yeah. And like the way medicinal chemists work heavier molecules more

做出来了，什么会被测试。

52:55

likely to bind a lot of times. It's because they're seeing improvements to affect other properties.

是啊。

53:00

That's part of it. And you probably got some parts to do. It's like basically you have a chemist

而且就像 medicinal chemists 的工作方式，更重的 molecules 更

53:04

too. They find a hit and then they start optimizing it and they're like, oh, I have a hypothesis

可能在很多时候 bind。那是因为他们看到一些改进会影响其他 properties。

53:08

that this is the binding site. So let me add this functional group to bind here. Okay. And then

这是其中的一部分。而且你可能也有一些部分要做。这基本上就像你也有一个 chemist

53:11

they're like, okay, well, that was a stronger binder. And then so they're like, oh, what if we add

他们会说，好吧，嗯，那个 binder 更强。然后他们就想，哦，如果我们加上

53:15

this other functional group to bind to this thing? And what's that happening? They just keep adding

另一个 functional group 来 bind 到这个东西上？那会发生什么？他们就一直加

53:20

things into it. It's really good. And then yeah, until it's sufficiently good. And then,

东西进去。效果真的很好。然后，对，直到它足够好。然后，

53:25

you know, then instead of using a combinatorial approach,

你知道，然后不是用 combinatorial approach，

53:28

or to say is the inside also, then you have these other properties. You're like, okay,

或者说是内部也，然后你还有这些其他的 properties。你就像，好吧，

53:31

well, now we have this really good binder, but it's not soluble. So let me throw on, you know,

嗯，现在我们有了这个非常好的 binder，但它不 soluble。所以让我加上，你知道，

53:35

some, you know, hydrophilic functional or some functional groups on the other side of the molecule,

一些，你知道，hydrophilic functional 或者一些 functional groups 到分子的另一侧，

53:42

which is not in the binding pocket to like make it, you know, and you just sort of you just keep

不在 binding pocket 里的那一侧，来让它，你知道，然后你就只是，你就一直

53:47

tacking these things on. And next thing, you know, you get this big good binder. And so you just

把这些东西都加上去。然后不知不觉，你就得到了这个又大又好的 binder。于是你只要

53:52

sort by molecular weight. And it's like the single biggest victor that you can have in most public

按 molecular weight 排序。这就像你在大多数 public

53:56

data sets. But that has nothing to do with chemistry. No, like I mean, I think a good way of

data sets 里能拿到的最大胜利。但这跟 chemistry 一点关系都没有。不，我是说，我觉得一个很好的

54:01

intuition is like, if a thing's working, they're going to do more stuff to it. Because I like it.

直觉就是，如果某个东西管用，他们就会对它做更多东西。因为我喜欢它。

54:07

Yeah. Yeah. It's like you just do more stuff to it. It gets more baroque or you have more like

对。对。就像你不断对它做更多东西。它就会变得更繁复，或者你会有更多那种

54:12

complicated expensive reactions. Or, you know, you get like rear building blocks, they're hard to

复杂又昂贵的 reactions。或者，你知道，你会用到像 rear building blocks 这种东西，它们很难

54:16

get, you know, things that you wouldn't do if you didn't think it was working. And so it's all

搞到，你知道，如果觉得它没效果，你根本不会去做的事情。所以这些全都是

54:20

these signatures that the models would just kind of pick up. And so yeah, it's quite terrifying.

这些 signatures，models 就会多少捕捉到。所以是啊，这相当吓人。

54:24

So this is like the horse, something.

所以这就像那匹马，什么的。

54:27

Clever huns. Clever huns.

Clever huns。Clever huns。

54:28

Clever huns, the horse.

Clever huns，那匹马。

54:29

He can you tell the story about Clever huns?

嘿，你能讲讲 Clever huns 的故事吗？

54:31

Yes, yes, so for a little history lesson here is there was a, there was a horse trainer

对，对，所以这里简单讲点历史，有一个，有一个驯马师

54:39

in Germany. I think it was like the late 1800s. I mean, I read about this. I should remember it.

在 Germany。我想大概是 1800 年代末。我是说，我读到过这个。我应该记得的。

54:43

And there's a horse named Clever Huns. And what Clever Huns seemed to be able to do was

然后有一匹叫 Clever Huns 的马。而 Clever Huns 似乎能做到的是

54:47

complicated math. So you would show the horse, you know, like a like a chalkboard that's

复杂的数学。所以你会给那匹马看，你知道，就像一块，一块黑板，那是

54:52

got a math problem. And then the horse would stamp the horse's foot on the ground and then stop

就给它出一道数学题。然后马会用蹄子在地上跺，接着停下来

54:56

when the correct answer, you know, was arrived at. And everyone was amazed that the horses were

等正确答案，你知道，得出来的时候。然后大家都惊叹这些马竟然

55:01

brilliant. And then someone had the good idea to go get a questioner who did not know the answer

这么聪明。然后有人出了个好主意，去找一个提问者，他不知道

55:07

to the math problem.

这道数学题的答案。

55:08

Yeah. And then the horse completely fails. And then it turns out that the horse is just looking at

对。然后马就彻底失败了。然后发现，马其实只是在看

55:13

like the posture and the micro expressions mostly of the owners. So that when it got to, you know,

比如主人的姿态和 micro expressions，而且主要就是主人的。所以当数到，你知道，

55:18

let's say the answer seven, then the older gets like real giddy when it's at six, right?

比如说答案是七，那数到六的时候主人就特别兴奋，对吧？

55:21

Yeah, the horses stop stamping. And there's this like wonderful paper put out by a German psychologist

对，马就停止跺蹄了。还有一篇很棒的论文，是一位德国心理学家发表的

55:28

who actually evaluates the actual micro expressions and postural changes that he'd observed on

到底是谁在真正评估他观察到的那些实际的 micro expressions 和 postural changes，这些

55:33

the horse's owner is quite extensive. Okay. Yeah. Like if you're dying for a silly reading, go look

在马主身上其实是相当广泛的。好。对。就像如果你特别想要个傻乎乎的解读，那就去看

55:38

out. So then the analogy is that you're reading the face and not the mask. That's right. Yes,

吧。所以这个类比就是，你读的是脸，不是面具。没错。是的，

55:43

you do not understand math. The models do not understand chemistry at all. They're just, you know,

你并不理解数学。这些 models 根本不理解化学。它们只是，你知道，

55:47

remembering the sociology of the creators. Got it. Going back to Hermes and your epistemic rigor

在记住创造者的社会学。明白了。回到 Hermes 和你的 epistemic rigor

55:55

and you know, you're in trying to build something useful out of all the data you've trained. So

而且你知道，你在试图从所有你训练过的 data 中构建出有用的东西。所以

56:00

you find, you know, public data sets are untrustworthy. And you have a bunch of data which you

你发现，你知道，public data sets 不可信。而你有一堆 data，你

56:06

trust internally, but you don't really have a good way of or but now you need to understand how do we

在内部信任它们，但你又没有真正好的办法，或者说，但现在你需要理解我们该怎么

56:12

accurately and carefully use this to understand to do something useful in the real world. Yeah.

准确、仔细地用它来理解，从而在现实世界里做点有用的事。嗯。

56:17

So how do you, how did you bridge that gap for us to think about this? Like we did a bunch of

所以你是怎么、你当时是怎么弥合这个差距，让我们能这样去思考的？就像我们做了一堆

56:22

these splits, like I kind of described it earlier, but then the next thing that seemed very obvious

这些 splits，就像我之前大概描述过的那样，但接下来有一件事看起来非常明显

56:26

for us to do is like, well, you know, we've trained models on internal data and we kind of don't

对我们来说要做的事就是，嗯，你知道，我们已经在 internal data 上 train 过 models，而我们有点不

56:30

like to train models on external data. So maybe we should try to predict on external data instead.

喜欢在 external data 上 train models。所以也许我们应该试着改成在 external data 上做 predict。

56:35

Which without ever training on the entire day without ever seeing it. And so that does put us in

也就是从没在整个一天上 train 过，也从没见过它。所以这确实让我们处于

56:39

a nice position because I haven't seen any other group train models for this task exclusively

一个不错的位置，因为我没见过其他任何团队专门为这个 task train models

56:47

on internally generated data. I think it's very tempting to go get these other sources, especially

在 internally generated data 上。我觉得去拿这些其他 sources 非常诱人，尤其是

56:51

because there's so many possible protein targets out there because people don't screen as many as we

因为外面有太多可能的 protein targets，因为大家通常不会像我们这样筛那么多。

56:54

do typically. And so we took large collections of external molecules and then we predicted them

所以我们拿来了大量外部的 molecules，然后对它们做预测，看看我们在很多不同的 proteins 上表现如何，还有很多不同的 positive negatives，都是我们非常仔细挑出来的。

57:01

to see how well we would do across, you know, many different proteins, lots of different positive

我想说，这差不多是 Hermes paper 里总的测试数量。

57:06

negatives that we chose very carefully. I want to say it was like the total number of tests in

我觉得大概是 1300 万之类的，横跨很多，大概 500 个 protein targets。我觉得那是个巨大的数字。

57:10

the Hermes paper. I think it's like 13 million or something across like many, it's like 500

然后我们把它和当时现有的、算是 state of the art 的 structural model 做了 benchmark，那个模型是 Boltz team 做的。

57:13

protein targets. I think it's some huge number. And then we benchmarked it against the existing

那是个非常好的模型。

57:18

what was at the time like the state of the art structural model, which is done by the Boltz team.

而我们的模型没法做到那个，很多那些……

57:22

And so which is a very good model. And like ours cannot do that a lot of the things that that

所以这是个非常好的 model。而像我们的就做不到，很多事情那个那个

57:26

model can do. But what we were able to do was do about as well as those guys on those public data sets

model 能做到的。但我们能做到的，是在那些公开 data sets 上，表现得和那些人差不多好

57:31

in terms of like this binding problem. When you compare the bolts, they required some some

就这个 binding problem 来说。当你比较 bolts 时，它们需要一些一些

57:36

sort of tricks to kind of have a various comparison. I think bolts also trained on this data.

某种 tricks 才能进行各种比较。我觉得 bolts 也在这个 data 上训练过。

57:41

Was that right? Sort of. So there are a couple large collections that we weren't totally clear

那是正确的吗？算是吧。所以有几个大型 collections，我们并不是完全清楚

57:48

on like the different subsets that they had trained on. And so they're I could think of one example

比如它们训练时用到的不同 subsets。所以，我能想到一个例子

57:52

in particular where it was very, very clear. And I'll talk about that in a minute. But a lot of the

特别是这一点非常非常清楚。我过一会儿会讲到那个。但很多

57:56

evals that we used were ones that were also surfaced by the Boltz team only we put a lot more

我们用的 evals 也是 Boltz team 提出来的，只不过我们放了更多

58:01

molecules in there to go see how we go. So yeah, like I think I think there. And also if you look

molecules 进去，看看我们做得怎么样。所以，嗯，我觉得我觉得是这样。还有，如果你看

58:06

at the different performances between herries and bolts, you know, like we're better on some

在 herries 和 bolts 的不同表现上，你知道，就像我们在一些

58:10

proteins and protein families that that bolts is worse at and vice versa. And we suspect this

proteins 和 protein families 上更好，而 bolts 在这些上更差，反过来也一样。而且我们怀疑这

58:15

is because of like different distributions of those in the training sets so that they understand

是因为它们在 training sets 里的分布不同，所以它们理解得

58:18

a little bit more or you know, honestly, maybe like memorize a little bit more and they're able to

稍微多一点，或者你知道，说实话，可能就记得稍微多一点，然后它们能够

58:21

extend that a little bit. One thing that was very validating for us is one of our test sets was

把这一点稍微扩展一些。有一件事对我们来说特别有验证意义，就是我们的一个 test sets 是

58:27

like another orthogonal chemical library that our models did not see that we had used internally.

就像另一个 orthogonal chemical library，我们的 models 没见过，但我们在内部用过。

58:32

But it was a copy of a chemical library that had been out in the literature. And so there was one

但它是一个已经在文献里公开过的 chemical library 的副本。所以有一个

58:37

target in particular. It was called as a protein. It's called BRD4 where there's an identical molecule

特别的 target。它被称为一种 protein。它叫 BRD4，其中有一个一模一样的 molecule

58:43

from our screen that was also in the PDB, the protein data bank training set that bolted almost

来自我们的 screen，它也在 PDB，也就是 protein data bank training set 里，bolted 几乎

58:49

certainly seen. Our models did kind of bad on that molecule. There's a far away. It was just a

肯定见过。我们的 models 在那个 molecule 上表现有点差。它离得很远。它只是个

58:53

chemical library. They weren't very good at it. And bolts was perfect at it. And it's because

chemical library。它们对它不太擅长。而 bolts 却做得完美。这是因为

58:58

we suspect that bolts had seen this protein and this individual molecule in a crystal structure

我们怀疑 bolts 已经在 crystal structure 里见过这个 protein 和这个单独的 molecule

59:03

already as part of the training set. So they're all semantics. Oh yeah, that checks out. It feels

已经作为 training set 的一部分。所以这些都是 semantics。哦对，那说得通。感觉

59:06

pretty good. So does this mean that I mean in your opinion, this validates your scaling hypothesis

挺不错。所以这是不是意味着，我是说，在你看来，这验证了你的 scaling hypothesis

59:12

add more proteins, more more diverse chemical libraries. And now you can basically fill in in

加入更多 proteins，更多、更多样化的 chemical libraries。而现在你基本上可以在这个

59:18

this phase more broadly or we hope so. Yes, we hope so. Yeah, like we think that going after

阶段更广泛地填补，或者说我们希望如此。是的，我们希望如此。对，就像我们认为接下来去攻克

59:24

large amounts of diversity and in particular going after intentionally chosen what you might

大量的 diversity，尤其是去追那些有意选出来的、你可能会

59:30

call activity cliffs is the way to do this and kind of unpack that a little bit. It's either on

称之为 activity cliffs 的东西，这就是做这件事的方式，而且稍微展开讲一点。它要么是在

59:35

the protein side or the small molecule side. Do you have two things that look very, very similar

protein 这边，要么是在 small molecule 这边。你有没有两个东西，看起来非常、非常相似

59:40

except for a tiny change in that tiny change drives like outsized changes in behavior.

除了一个微小的变化，而那个微小的变化会带来大得不成比例的表现变化。

59:46

So like a real classic example of this that medicinal chemical talk about is this idea. And

所以，一个真正的经典例子，就是 medicinal chemistry 里会聊到的这个想法。然后

59:51

it's kind of mythical. It's not exactly real, but this thing called a magic methyl and the idea is

它有点像是神话。它并不完全真实，但有这么个叫 magic methyl 的东西，而这个想法是

59:55

that you've got a molecule. And then maybe you've got like this one like little thorn that is or

就是你有一个 molecule。然后也许你有这么一个像小刺一样的东西，它要么

60:01

is not there. And that presence or absence of that thorn means like either works great or it

在，要么不在。而那个小刺的有或无，就意味着它要么效果特别好，要么它

60:06

doesn't work at all. When you have the situations, then maybe your model will start to learn like man,

根本不起作用。当你遇到这些情况时，也许你的 model 才会开始学习，就像，老兄，

60:11

like having a thorn could be really meaningful. And so maybe I need to think about like these

就像有个刺可能会真的很有意义。所以也许我需要想想这些

60:15

my very small changes. And so one thing that we do a lot and it's baked into delts because you've

我非常小的改变。所以我们经常做的一件事，它已经融入到 delts 里了，因为你有

60:21

got you know two out of three building blocks might be similar across the large numbers of examples

你知道，三个 building blocks 里可能有两个在大量的 examples 中是相似的

60:24

and have variations of behaviors. And then on the protein side, we very intentionally engineer

并且有行为上的变化。然后在 protein 这边，我们非常刻意地设计

60:30

mutations in those proteins to do the same kind of thing. A little confusion because if you have

那些 proteins 里的 mutations 来做同样的事。有点困惑，因为如果你有

60:36

these building blocks and the size of your variation is multiple atoms right many at how big is

这些 building blocks，而你的变化大小是多个 atoms，对吧，很多，那有多大

60:43

like a building block in terms of atoms right right it might be it might be like six to like 30 or

像是一个 building block，按 atoms 来说，对吧对吧，可能是，可能是大概 6 到 30 或

60:47

something you're for not ignoring like if we're going like the hydrogens but you can you can imagine

有些你不该忽略的东西，比如说如果我们在讲 hydrogens 的时候，但你可以，你可以想象

60:51

that their building blocks that themselves are very similar with like a single atom substitute.

它们的 building blocks 本身非常相似，只是像换了一个单独的 atom。

60:54

Oh I see okay so maybe there's two building blocks that have one one thorn and not a thorn and

哦，我懂了，好吧，那也许有两个 building blocks，一个有一个 thorn，另一个没有 thorn，而且

61:00

they're you and then so you can see that in the in the sequencing that's right that's right all right

它们就是你，然后你就能在 sequencing 里看到这一点，对，对，好吧

61:05

and maybe that thorn just changes position okay and then as you're and as you're thinking about

也许那个 thorn 只是换了位置，好吧，然后当你在，当你在想

61:09

this you're like okay well I might have tension between my goals I might want lots of like

这个的时候，你会想，好吧，我可能在我的目标之间会有拉扯，我可能想要很多，像

61:15

rougher more coarse diversity because I'm trying to find binders across large numbers of protein

更粗糙、更 coarse 的 diversity，因为我想在大量 protein 上找到 binders

61:20

targets that make me very different or I may want like super fine grained changes that are very

targets，让我变得非常不同，或者我可能想要那种 super fine grained 的变化，这些变化非常

61:25

very subtle and you know try to like really teach the models like these like very very subtle

非常细微，而且你知道，得试着真的把这些教给 models，就是这种非常非常细微的

61:30

differences and so as we think about designing your own chemical material like we have to

差异。所以当我们考虑设计自己的 chemical material 时，我们得

61:33

you know span this kind of tension so yeah I would think about it a lot. And both of those matter

你知道，去跨越这种张力，所以对，我会经常思考这件事。而且这两者都重要

61:38

both of those matter both of those matter and it also matters in the protein side too right it

两者都重要，两者都重要，而且它在 protein 这一侧也很重要，对吧，它

61:43

matters very much the protein side and so as you look at current like sort of co-folding models

在 protein 这一侧非常重要。所以当你看现在的，像是某种 co-folding models 时

61:49

a couple different groups have shown this you can take protein and it's got this amino acid sequence

有几个不同的团队已经展示了这一点：你可以拿一个 protein，它有这么一段 amino acid sequence

61:53

and it codes some protein that's got a big pocket in it and maybe that protein is represented

然后它会编码出某种 protein，里面有一个很大的 pocket，也许那个 protein 已经被表示在

61:56

in the training sets so the model has a pretty good idea about what it looks like when there's a molecule

training sets 里，所以 model 对它长什么样有一个相当好的概念，当有一个 molecule 的时候

62:00

in there or a small molecule chemical in there so you got this pocket I think it goes in the pocket

在那里面，或者那里面有个 small molecule chemical，所以你就有了这个 pocket，我觉得它会进到 pocket 里

62:04

you can just change the amino acid sequence that you give to the model to consider and you can

你只要改变你给 model 去考虑的 amino acid sequence，然后你就可以

62:10

use that to break the pocket you could fill the pocket in like you're filling it with cement or you

用它来破坏 pocket，你可以把 pocket 填起来，就像你在用水泥把它填进去一样，或者你

62:14

could remove it entirely so that there's nothing there or it's gone or it's much bigger you do

可以把它完全移除，让那里什么都没有，或者它没了，或者它变得大很多。你做

62:18

any of these things and you get at least what we've seen so far if you use current state of the

其中任何一件事，然后至少就我们目前看到的来说，如果你用当前

62:22

our models to go see what the model will do it will show the small molecule in where the pocket used

我们的 models 去看看 model 会做什么，它就会显示 small molecule 在 pocket 原来所在

62:28

to be because it just remembers where it is yeah and that's best looking at the large scale structural

的位置，因为它就记得它在哪里。对，而且那最好是在看 large scale structural

62:34

things at night some or like the hypothesis anyway is that it's doing this yeah the hypothesis is

的东西，在晚上，有些，或者像 hypothesis 那样，反正就是它在做这个。对，hypothesis 是

62:39

that it doesn't quite understand you know when you when you have a pocket and a protein there's going

就是它不太理解，你知道，当你——当你有一个 pocket 和一个 protein 的时候，会有少量 amino acids，它们的 hydrogen bond positions 会伸出来，在这些别的东西里还会有一点点 high stack，你知道，某种程度上驱动着这种 chemistry，而那些东西真的应该非常非常重要，而且当你在现实世界里测量它们的时候，它们也确实非常非常重要，所以如果你把它们破坏掉，而 model 又懂 physics，那它大概就应该知道它不能把 molecule 显示在那里，而这就是问题所在，就像很多这些 structural training sets，它们没有很好的 negatives，因为能进到那里的东西就是——对，我就在想，这就是关键，对吧，你知道

62:43

to be a small number of amino acids that will have hydrogen bond positions that are going they're

是少量的 amino acids，它们会有 hydrogen bond 的位置，那些会，它们

62:48

sticking out there's going to be a little bit of high stack in these other things that you know

伸出来，会有一些 high stack 在这些别的东西里，你知道

62:52

sort of drive this chemistry and those things really ought to be very very important and they are

某种程度上驱动这个 chemistry，而那些东西真的应该非常非常重要，而且它们确实是

62:56

very important in the physical world when you're measuring them and so if you break them and the

在物理世界里非常重要，当你测量它们的时候，所以如果你破坏它们，而

63:00

model understands the physics then it probably should know that it can't show the molecule in there

model 理解 physics，那么它可能应该知道它不能把 molecule 显示在那里

63:05

and that that's the thing is like a lot of these structural training sets like there aren't great

而且那就是问题所在，就像很多这些 structural training sets 一样，就像没有很好的

63:10

negatives because what makes it into there is yeah I was thinking that's the thing right you know

negatives，因为能进到那里的东西是，是的，我在想那就是问题所在，对吧，你知道

63:14

they don't have a lot of these variations that really you know disrupt these behaviors and no one's

它们没有很多这种 variations，你知道，真的会 disrupt 这些 behaviors，而且没人会

63:18

going to make crystal structures with a bunch of these newtons like we you know like we're currently

像我们现在这样，用一堆这些 newtons 去做 crystal structures，你知道，就像我们目前

63:22

doing first about how do you get negatives the way that we get negatives is that we have a really

正在做的。首先，negatives 怎么来？我们拿到 negatives 的方式是，我们对

63:26

good understanding of the distribution of the number of those molecules in a baseline condition

那些 molecules 在 baseline condition 下数量的 distribution 有非常清楚的了解。

63:32

okay so like when we do the screen we know oh okay these bound a lot because now we're enriching

好吧，所以就像我们做 screen 的时候，我们知道，哦，这些 bound 得很多，因为现在我们正在对它们做 enriching，

63:39

for them when we pull the protein out and now let's break the pocket and they are no longer enriched

等我们把 protein 拉出来，然后我们 break 这个 pocket，它们就不再 enriched 了。

63:44

okay they are no longer going that one amino acid substitution that we made is making a bunch of

好吧，它们不再继续了。我们做的那个 amino acid substitution 让一堆

63:49

these chemical series dropout but this is a one-off experiment this is not a del no this is a del

这些 chemical series dropout，但这是一次 one-off experiment，这不是一个 del，不，这是一个 del。

63:53

oh no so you're just you're not saying you broke it you're saying I find the case where it was

哦不，所以你就是——你不是说你把它弄坏了，你是说，我找到它本来

63:58

broken no we break it so we break the so we take the so let's say we screen the protein and the

就是坏掉的那种情况。不，是我们把它弄坏。所以我们把那个——我们把那个——这么说吧，我们 screen 这个 protein，然后那个

64:02

and the pocket is intact yeah sounds great okay we get a bunch of hits yeah um let's let's go

那个 pocket 是完整的，对，听起来不错，好，我们拿到一堆 hits，对，嗯，我们——我们继续

64:08

even further here let's say that we know someone else has already identified a chemical that goes

再往深一点，这么说吧，我们知道别人已经找出了一个 chemical，它会进到

64:13

in that pocket let's buy it let's stick it in there run the screen again yeah that means all the

那个 pocket 里。我们买它，把它塞进去，再跑一次 screen。对，这意味着所有

64:18

chemicals that normally go in the pocket will be displaced by a high cost oh you have a

本来会进那个 pocket 的 chemicals 都会被一个 high cost 挤掉。哦，你就有一个

64:24

competitive asset oh we do all the time okay okay do this all the time yeah okay and then you say

competitive asset。哦，我们一直这么干。好吧好吧，一直这么做。对，好，然后你说

64:29

okay well that looks pretty good I believe that and then you say all right now we know that this

好吧，嗯，这看起来挺不错，我相信。然后你说，好吧，现在我们知道这个

64:33

chemical that someone else's characterized goes in this pocket because maybe they made a crystal or

别人已经 characterized 过的 chemical 会进到这个 pocket 里，因为可能他们做出了 crystal，或者

64:36

some other thing yeah great let's look at the amino acids in that pocket and then we will change

别的东西。对，太好了，我们来看看那个 pocket 里的 amino acids，然后我们会改

64:42

the DNA construct that makes the protein for us we will now mutate physically this one residue

那个为我们制造 protein 的 DNA construct，我们现在会实际地 mutate 这一个 residue

64:48

so now like this hydrodist you know this this physical amino acid now we put a big bulky one in

所以现在就像这个 hydrodist，你知道，就是这个这个实际的 amino acid，现在我们放一个很大、很 bulky 的进去

64:53

another pocket full or like the different side chain or something yeah all these things you can do

另一个 pocket 里塞满，或者像不同的 side chain 之类的。对，所有这些事你都能做

64:56

yeah yeah and this is really easy now because we have alpha-fold and you know bolts or all these

对呀对呀，现在这真的很简单，因为我们有 alpha-fold，还有你知道的 bolts，或者所有这些

65:01

structure models you can you can use this to design your proteins as you you can use this you can

structure models，你可以你可以用这个来 design 你的 proteins，就像你可以用这个，你可以

65:06

use and for things that are just what characterized you know like we use conservation you know like we

用，而且对于那些只是 what characterized 的东西，你知道，就像我们用 conservation，你知道，就像我们

65:10

stare at we have a structure why I just have a kind of look at these things like oh yeah you know

盯着看，我们有一个 structure，为什么我就是会有点，看这些东西，就像，哦对，你知道吧

65:13

this guy we're craft is amazing and so all right so you know now we're going to make this one mutation

我们正在 craft 的这个家伙太厉害了，然后，行吧，所以你知道，现在我们要做这一个 mutation

65:17

let's do get another screen and oh look all the chemicals that got booted out when we hit it with the

我们再做一次 screen 吧，然后，哦看，所有那些 chemicals，当我们用那个——打它的时候，都被踢出去了

65:24

we with that one thing that's blocking the hole they're also gone we broke the pocket so now we have

我们用了那个堵住洞的东西，它们也都没了，我们把 pocket 弄坏了，所以现在我们有

65:29

really good evidence that they're all going into that pocket right right and then here's the point

非常好的证据，证明它们都在进入那个 pocket，对，对，然后接下来就是关键点

65:33

where you know like the protein chemists come in and they say well if you made a mutation how do you

就是，你知道，像 protein chemists 进来说，嗯，如果你做了一个 mutation，你怎么

65:37

know that you didn't just break the whole thing right reasonable question you say okay well we only

知道你不仅仅是把整个东西弄坏了呢？对吧，这是个合理的问题。你说，好吧，我们只是

65:42

lose the chemical series they got kicked out by that competitor all the other ones that are binding

失去了那个 chemical series，它们被那个 competitor 踢出去了，所有其他还在 binding 的

65:47

everywhere else in the protein are intact in the attack yeah so you didn't just break the protein

protein 里其他地方在 attack 里都是 intact 的，yeah，所以你不是只是把 protein 弄坏了

65:51

you didn't just break it right and so and then you think okay well if you're going to be

你不是只是把它弄坏了，对吧，然后你就想，好吧，如果你要

65:54

out screening a hundred protein constructs a week which is what we do on big weeks like they

一周 screening 一百个 protein constructs，这就是我们在忙的那几周会做的事，就像他们

65:59

sorry so you've broken the pocket oh yeah oh yeah then how do you also find that there are other

抱歉，所以你把 pocket 弄坏了，哦 yeah，哦 yeah，那你怎么还能发现还有别的

66:06

things that do that start to connect to bind to the protein so when you change things around you

东西会那样，开始连接、bind 到 protein 上，所以当你改变周围的东西时，你

66:11

might change the shape in a way that doesn't encourage other things to bind we have definitely seen

可能会以一种不促进其他东西 bind 的方式改变形状，我们肯定见过

66:15

and and is that an information source for you it is it is and so another thing that's really nice

而且，而且这对你来说是一个信息来源吗？是的，是的，所以另一个很棒的地方

66:20

about this and the nice part about delts is you can you can imagine if you're not thinking in large

关于这个，还有 delts 好的地方是，你可以，你可以想象，如果你不是在大规模地思考

66:25

numbers that oh I've got this one chemical that was spining and now I broke the pocket and now it's

数字，那个，哦，我有这么一个 chemical，它之前在 spinning，现在我把 pocket 弄坏了，现在它

66:30

gone yeah but because of the combinatorial nature of the delts maybe I have five different

不见了。对，但因为 delts 的 combinatorial 性质，也许我有五个不同的

66:36

chemical series it's like it's like hundreds of molecules and there's another group that's

chemical series。就像，就像几百个 molecules，而且还有另外一组

66:41

the same hundreds of molecules with a single change in the building block so like you have these

一样的几百个 molecules，只是 building block 上有一个改动。所以就像你会有这些

66:45

repeated examples where it's like oh this one tiny thing shifted the population behavior

重复出现的例子，就像哦，这一个小小的东西把 population behavior

66:50

enormously and it's a big population and so it's like oh I am now really convinced I can show

巨大地改变了。而且这是一个很大的 population，所以就像，哦，我现在真的相信了，我可以给你们看

66:54

you guys a plot that we I don't know if you you know put pictures on on a block structure but

你们一个 plot，我们，我不知道你们是不是，呃，把图片放在 block structure 上，但

66:59

it's very obvious like oh this one side group on this building block is clearly interacting with

很明显，就像哦，这个 building block 上的这个 side group 明显在和

67:04

this one amino acid residue because we have like five independent examples of population behavior

就这一个 amino acid residue，因为我们大概有五个独立的 population behavior 例子

67:09

shifting because of it is beautiful it's very very clear yeah go ahead tell us about this plot

因为这一点而发生偏移，这很美，非常非常清楚。对，继续，给我们讲讲这个 plot

67:14

we pulled this thing up and the reason that we pulled it up is we think it's really important for

我们把这东西调出来了，而把它调出来的原因是，我们觉得非常重要的是，

67:19

there to be a lot of variation in the proteins as well as the small molecules and a reason that we

在 proteins 以及 small molecules 里要有大量的 variation，而我们之所以

67:23

think this is you want to be able to get the you know and a model trying to learn medicinal chemistry

认为这一点是，你想能够拿到，你知道，而且一个试图学 medicinal chemistry 的 model

67:29

to really start to put together you're having these hydrogen bond interactions between like this

真正开始把它们拼起来，你有这些 hydrogen bond interactions，在像这个

67:34

part of this molecule in this part of the protein we've seen other work that I talked about earlier

molecule 的这一部分和 protein 的这一部分之间。我们见过其他研究，我之前讲过

67:38

about how you can if you digitally break a pocket and you use the current state there are models

关于你可以怎样，如果你 digitally break 一个 pocket，并且你使用 current state，就会有一些 models

67:43

that they have a really hard time understanding the physics and they'll just show the molecule

就是他们真的很难理解 physics，然后他们就只是把 molecule

67:46

back in that thing anyways so it makes this suggest that they're memorizing and so we thought one

放回那个东西里，反正，所以这暗示他们是在死记硬背，然后我们就想，一个

67:50

way to go around this is to experimentally engineer mutations in binding regions of these proteins

绕过这个的办法，就是实验性地在 these proteins 的 binding regions 里 engineer mutations

67:56

and then re-screening and so what that looks like physically you take the DNA of the wild

然后再 re-screening，那么具体操作上是什么样呢，就是你拿 wild

68:04

type protein that you've already screened and then you use microbiology techniques to change one

type protein 的 DNA，你已经 screened 过了，然后你用 microbiology techniques 去改变一个

68:08

or more than one amino acids in that protein so that the functional part where the small molecules

或者不止一个 amino acids，在那个 protein 里，这样那个功能部分，也就是 small molecules

68:12

are binding is no longer the same it's different in some way we have a different methods to try to

binding 的地方，就不再一样了，它以某种方式不同了，我们有不同的方法来尝试

68:18

identify which amino acids are the smartest to break sometimes you break ones that you're confident

识别哪些

68:22

will do things sometimes you don't know and so you but we can do this at a pretty big scale so

有时候会做一些你不知道的事情，所以，你，但我们可以在相当大的 scale 上做这个，所以

68:26

it's not a big problem to us to take a couple you know wingers at each one and so then you run

对我们来说不是什么大问题，对每一个都来几个，你知道，wingers，然后你就运行

68:30

the screen and now when the pocket is broken where the molecules go what do the molecules do

这个 screen，现在当 pocket 被破坏时，molecules 会去哪儿，molecules 会做什么

68:36

do they still go in there do some of them go in there is there something common among the ones that

它们还会进去吗，有些会进去吗，那些

68:41

stick around or does something come in about the ones that get dislodged or or or pitched out

留下来的之间有没有什么共同点，或者那些被 dislodged 的，或者或者或者被踢出去的，会不会有什么共同点

68:45

and then you have lots of different examples of these where you might have three to five

然后

68:49

vicious amino acids in a given pocket that are driving the binding behaviors maybe you try to

一样。他们找到一个 hit，然后开始优化它，他们会说，哦，我有个 hypothesis

68:54

break all them independently and then you do redo the screens each time and get a good feel for what

这里就是 binding site。所以让我加这个 functional group，让它 bind 到这里。好。然后

68:58

part of the small molecule is interacting with the part that you have changed and then when

small molecule 的一部分正在与你改过的那个部分发生相互作用，然后当你

69:03

you do this with a large dowel then you'll have lots and lots of repeated examples of similar

用一个大 dowel 来做这件事时，你就会有很多很多重复的相似

69:08

molecules they're all following the same behavioral trajectories and that way you get a really good

molecules，它们都沿着同样的 behavioral trajectories，这样你就能有一个非常好的

69:13

feel for oh I've got lots of examples they're all pretty similar but the one thing is super common

感觉：哦，我有很多例子，它们都挺相似的，但有一件事超级常见

69:19

is this one little bit on the small molecule so I have a suspicion that they's interacting with

就是 small molecule 上的这一小块，所以我怀疑它们是在跟

69:23

the pocket that I just broke what are we looking at on this diagram so what we're looking at on this

我刚破坏的那个 pocket 相互作用。这个 diagram 上我们看的是什么？我们在这个

69:28

plot is all those little dots are small molecules from a chemical library that we have

plot 上看到的是，所有那些小点都是来自一个 chemical library 的 small molecules，这个 chemical library 是我们

69:33

designed and made along the axes are how well the small molecules bound so if it's like really

设计和做出来的，坐标轴上表示的是这些 small molecules 结合得有多好，所以如果它真的

69:39

down the axis it's binding well if it's more towards the origin it's not really binding at all

沿着这个 axis，它 binding 得很好；如果它更靠近 origin，那就根本不太 binding。

69:44

and then we've colored a couple different blobs and those blobs are different chemical

然后我们给几个不同的 blob 上了色，而这些 blob 是不同的 chemical

69:48

series that chemicals like bracing more to each other but the difference is where we've pointed

series，那些 chemical 更喜欢彼此之间互相 bracing，但区别在于我们指向的位置

69:52

that kind of chemical looking structure and so if you're squinting and you look at these

那种看起来像 chemical 的结构；所以如果你眯着眼看这些

69:57

they're all pretty similar but there's a couple subtle differences and that whole population

它们其实都挺像的，但有几个细微差别，而整个 population

70:01

we believe is driven by interactions of those differences to the one view that we made so in this

我们认为，是由那些差异之间的 interactions 驱动的，对于我们做的那个 view 来说，所以在这个

70:06

plot there's on the bottom the wild type target and then on the on the horizontal or the vertical

plot 里，底部是 wild type target，然后在横向或纵向的

70:13

axis you have the mutant type target and so there's this red one and the red one is pointing

axis 上，你有 mutant type target，所以有这一个红色的，而这个红色的指向

70:19

basically up it is binding better it is binding better so that particular one is like the mutation

基本上，向上就是 binding 更好，就是 binding 更好，所以那个特定的那个，就像是 mutation

70:27

made that particular one bind better that's which is speaking to the question that I asked a few

让那个特定的那个 bind 得更好，这正好说明了我几分钟前问的

70:32

minutes ago and then maybe some of the most of these seem to be above the axis so most of this

那个问题，然后可能，这些里面大多数似乎都在 axis 上方，所以大多数这个

70:39

mutation caused most of these to bind better that's correct yeah there's a second mutation that

mutation 让这些大多数都 bind 得更好了，对，没错，对，还有第二个 mutation，那个

70:44

pocket that I can also share with you guys where all of the all of those population all those

pocket，我也可以跟你们分享，在那个里面，所有的，所有那些 population，所有那些

70:49

blobs get kicked out entirely and they do not bind so then the pocket was broken into meaningful

blobs 都被完全踢出去了，它们不 bind，所以那个 pocket 被以有意义的方式

70:53

way and then here the binding is increased and so they basically everything will be along the bottom

破坏了，然后这里 binding 增加了，所以它们基本上所有东西都会沿着底部

70:58

that's correct okay that's correct yeah so you can give this yeah yeah yeah awesome so doing this you

没错，好，没错，对，所以你可以给这个，对对对，太棒了，所以这样做你

71:04

can map out both well you can you you can treat the activity cliff issue both in compound space

可以把两者都映射出来，嗯，你可以，你你可以在 compound space 里处理 activity cliff 问题

71:10

and in protein space that's correct and this is particularly useful if you're let's say working at

也可以在 protein space 里，没错，而且如果你比如说是在做

71:15

kinases where you there's like a hundred kinases which differ only by like and the important

kinases 的话，那里有大概一百种 kinases，它们之间的区别只在于，比如，那个重要的

71:20

binding site you might differ by one residue at a time or something so this can be incredibly

binding site 可能一次只差一个 residue 之类的，所以这会非常

71:25

powerful for certain classes like that in addition to your own sanity yeah that's right yeah and we

强大，对某些像那样的类别来说，另外也能让你自己保持理智，对，没错，对，而且我们

71:31

do this you know as I mentioned we do this we do this a lot yeah many many hundreds of these types

会做这个，你知道，就像我提到的，我们做这个，我们经常做这个，对，很多很多这类

71:35

of screens one of the things you can do with this is Hermes sees this cross attention so you can

screen，你能用这个做的一件事是，Hermes 会看到这个 cross attention，所以你可以

71:40

actually interpret where Hermes thinks the binding pocket is so it's learning the binding pocket

真正解读出 Hermes 认为 binding pocket 在哪里，所以它在学习 binding pocket

71:46

through these types of maps yes so one thing that we've sort of written about is using these attention

通过这类 maps，是的，所以我们大概写过的一件事，就是用这些 attention

71:52

mapping approaches to have a look at where Hermes is you know looking at an amino acid space and when

mapping approaches 来看看 Hermes 在哪里，你知道，看看 amino acid space，然后当

71:59

we first did this work we were expecting to for it to either look at the binding pockets that was

我们第一次做这个工作时，我们原本预期它要么去看 binding pockets，这算是

72:03

like the first intuition and then maybe like a more measured intuition was like well maybe

最初的直觉；然后也许一个更谨慎的直觉是，嗯，也许

72:07

what we'd hope it for it to do is maybe look at all those residues because it probably doesn't

我们希望它做的是，也许去看所有那些 residues，因为它可能并不

72:10

really know what's going on yet and then when we looked it was not looking at either of those

真的知道现在发生了什么；然后当我们去看时，它并没有在看这两者中的任何一个

72:14

things it was looking at really short sequences of amino acids that were away from the functional

东西，它看的是非常短的 amino acid 序列，这些序列远离 protein 的功能

72:21

part of the protein and also away from part of the protein that looks similar across many proteins

部分，也远离在很多 protein 中看起来相似的那部分 protein

72:26

and so for maybe the non biology people in the room the way a lot of molecular evolution has worked

所以，可能对在座不是做 biology 的人来说，很多 molecular evolution 的运作方式是，

72:33

is that there's a stretch of a protein that does a pretty good job and then often gets shuffled

就是有一段 protein 做得挺不错，然后经常会被打乱重排，

72:38

among other proteins as well so that you have this conserved regions called domain or you know

和其他 proteins 混在一起，所以就有了这些叫 domain 的 conserved regions，或者说，

72:42

conservative protein domain and so that part looks pretty similar across many proteins it's something

conservative protein domain，所以那部分在很多 proteins 里看起来都挺像，它是某种

72:46

that biologists will use to infer protein families and relationships and functions so they all

做 biology 的人会用来推断 protein families、relationships 和 functions 的东西，所以它们都是

72:52

the same kind of piece they're probably doing something kind of similar and so we expected the models

同一种片段的话，很可能在做差不多类似的事，所以我们本来以为 models

72:55

look at that but they don't they look outside that and they look at these little stretches that

会看那个，但它们没有，它们看那之外，它们看这些小小的片段，

73:00

people don't think typically do a whole lot they're like a little bit different they're

人们通常不觉得这些片段能起多大作用，它们有点不一样，它们

73:03

non conserved there's a lot more variation in those and we realized that Hermes was probably coming up

non conserved 的那些里面，variation 多得多，然后我们意识到 Hermes 很可能是在

73:08

with a nickname for the protein it's not thinking about oh this is the functional part it's like

给这个 protein 起个外号。它不是在想，哦，这是 functional 的部分，而是像

73:14

I don't need to learn that I can remember you know 10 or 12 amino acids that are different

我不需要学那个，我可以记住，你知道，10 个或 12 个 amino acids 是不同的

73:19

here than everyone else and I can just memorize it and we thought that was bad

在这里跟其他所有人都不同，我可以直接记住它，而我们还觉得那很糟糕

73:24

because maybe many of the like if you have variations in the pocket or whatever

因为可能很多，就像，如果你在 pocket 里有 variations，或者什么的

73:30

but they all have these other parts that are similar then maybe in the variation the pocket

但它们都有这些相似的其他部分，那么也许在 variation 里，pocket

73:37

mostly don't matter but some of them do then you might the model might learn oh I can just

大多都不重要，但有些确实重要，那你可能，model 可能会学到，哦，我只要

73:41

pay attention to these yeah the concern was that the model again was not really learning any physics

关注这些就行了。对，当时的担忧是，model 又一次并没有真正学到任何 physics

73:46

it's like oh I can just memorize this stuff it's very much in keeping with when you try to break

就像是，哦，我可以直接把这些东西背下来，这跟当你试图打破

73:50

the pocket with the co-folding models and they shove it in there anyway it's like oh it doesn't

那个 pocket 时一样，用 co-folding models，而它们还是硬把它塞进去，就像，哦，它根本不

73:53

know what's going on it's just made this nickname yeah so yeah man so how do you fix that and then

知道发生了什么，它只是起了这个绰号。对，所以，哥们，那你怎么解决这个问题，然后

73:58

that's when we hit upon trying to do these mutations at scale and so what happens is if you have

就在那时我们想到，试着做这些 mutations at scale，而接下来发生的是，如果你有一个

74:03

a protein that looks almost exactly the same but there's one tiny difference and then the behaviors

protein，看起来几乎一模一样，但有一个很小的差异，然后那些行为

74:08

change a lot as a consequence of that one tiny difference you're kind of encouraging the models

会因为那一个小差异发生很大变化，你其实是在促使这些 models

74:13

to look at all the amino acids again I can't do that for the section that's non conserved because if

重新去看所有的 amino acids。对于 non conserved 的那一段，我不能那么做，因为如果

74:17

it does that with these two things in different behavior it's gonna do a bad job yeah and so we

它面对这两个行为不同的东西还那么做，它就会做得很差。对，所以我们就

74:21

went back and did well you know trade models with these additional mutations like the one on this

在某个 pocket 里驱动 binding behaviors 的 vicious amino acids，也许你会试着

74:25

pot that you guys are seeing and then you know look back at the attention maps and now Hermes will look

你们看到的那个 pot，然后你知道，回头看 attention maps，现在 Hermes 会看

74:29

at all of these things much more evenly so you're basically using the attention map to find these

所有这些事情时都均匀得多，所以你基本上是在用 attention map 来找这些

74:34

cases where it's memorizing nicknames yeah and then enrich the mutations in the pocket to kind

它是在记住昵称的那些情况，yeah，然后丰富 pocket 里的 mutations，来稍微

74:41

of encourage it to pay attention to the pocket instead yeah yeah and it's not necessarily just to

鼓励它转而关注 pocket，yeah yeah，而且这不一定只是要

74:46

be across any one gene it's more like the general problem of like oh I can't use this nicknames

横跨某一个 gene，更像是那个普遍问题，就像，哦，我没法用这个昵称

74:52

strategy effectively anymore you want to teach the model that now don't do that you don't do that no

策略实际上不再有效了，你想教 model，现在别那么做，你别那么做，不行

74:58

no that's that's lame don't do that so I love this because we've seen where I think that

不，那、那太逊了，别那么做。所以我特别喜欢这个，因为我们已经看到，我觉得

75:03

the most pragmatic use of mecanterp has been in biology and probably others related sciences where

mecanterp 最务实的用法一直是在 biology，可能还有其它相关科学，其中

75:14

in the reason why I think it's qualitatively different is because you go backwards from

而我觉得它之所以在性质上不同，是因为你是从反方向来的：

75:19

you're learning about biology from the model rather than the just trying to interpret what

你是从 model 那里学习 biology，而不是只是试图解读什么

75:26

like what is like what is it thinking in this case maybe it is actually that for a case but I've

比如它到底在想什么，在这种情况下，也许对这个案例来说确实就是那样，但我

75:31

just seen so much mecanterp being applied to biology because you're really trying to understand

只是看到这么多 mecanterp 被应用到 biology，因为你真的是想理解

75:37

you don't have the language in these machine learning models there is no language to lean on so

在这些 machine learning models 里，你并没有语言，没有什么语言可以依靠，所以

75:43

you really have no clue what is thinking unless you're inspecting the model well it's really

你真的完全不知道它在想什么，除非你去检查 model。嗯，这真的

75:48

interesting so yes and one thing that we have found very fruitful in the very recent past is

很有趣。所以是的，而且我们在最近这段时间发现非常有成效的一件事是

75:53

actually transitioning to language models okay and what we are currently doing is we're using

实际上是在转向 language models，OK，而我们现在做的是，我们在用

76:00

readable interpretable English to represent the small molecules and the proteins and their behaviors

可读、interpretable 的英文来表示 small molecules 和 proteins 以及它们的行为

76:04

wow okay and that turns out to work pretty good yeah I mean this is yeah yeah I want to just

哇，OK，而且结果效果还挺不错的。对，我是说，这……对，对，我只想先

76:11

frame this a little bit before you tell us about this because yeah there's actually a really

在你跟我们讲这个之前，我先稍微铺垫一下，因为对，其实真的有一篇很

76:16

interesting paper in the single cell domain called gene PT in which people so I count over the

有意思的论文，在 single cell 领域，叫 gene PT，里面的人……所以我数了下

76:25

authors are from Stanford and it was somebody who was a statistician not even really in bio

作者们来自 Stanford，而且其中有个是统计学家，甚至根本不是真正搞生物的

76:31

maybe they are now but that person they took the textual embeddings of the descriptions of

也许他们现在是了，但那个人，他们拿了那些描述的 textual embeddings，关于

76:40

genes yep and then they use those and so like you're using the SM embeddings so for her museum

基因，对，然后他们用这些，所以就像你在用 SM embeddings，所以对她那个 museum 来说

76:46

so you are using the SM they were just like let's just use the textual embeddings and the beauty

所以你们是在用 SM，他们当时就说，干脆就用 textual embeddings 吧，而美妙之处

76:50

of this is because now you have like a gazillion years of human time put into understanding each of

在于，现在你有如天文数字般的人类时间投入到了解每一个

76:58

these genes and that's all captured to some extent by these textual embeddings that knowledge of

这些 genes，而这些 textual embeddings 在一定程度上捕捉到了所有这些——那种对

77:03

the embedding so like yeah so please tell us about what you're doing yeah so we started this project

embedding 的知识。所以，嗯，那请跟我们讲讲你们在做什么。嗯，我们开始这个项目

77:09

a little while ago because one thing that we realized is that we had this like growing collection

是在不久前，因为我们意识到的一件事是，我们有这么一批不断增多的

77:16

of activity cliffs and profiles of the same chemical material across large numbers of proteins

activity cliffs 和 profiles，是同一种 chemical material 在大量 proteins 上的

77:20

and we wanted to know if we could actually use reasoning agents to go explore it and use it in

我们想知道，我们是不是真的能用 reasoning agents 去探索它，并把它用在

77:25

a lab and a loop kind of situation and so that demanded that we started working with reasoning models

一种 lab 和 loop 式的情形里。所以这要求我们开始使用 reasoning models

77:31

and language models in order to kind of let them go play and so what that looks like is we took some

还有 language models，某种程度上是为了让它们能去玩一玩，而具体做法就是，我们拿了一些 open source language models，比如我们做了一些 quins 和 GLM，还有其他人在用的那些东西，然后我们教它们 medicinal chemistry，方式是你说：这里有一个结构，用 natural English 描述它；用 natural English 给我两个分子之间的 diff；然后我们把它推进到这一步：我们说，OK，这里有一些分子是 binders，这里有一些不是 binders，你能告诉我们你觉得这些 binders 为什么会 binding 吗；然后你也可以给它 hard negatives。所以是的，有这两组 collection，它们会说，OK，我觉得这个 chemical series 有这个，你懂的，然后它会把这些东西用 rebel English 写下来，我们觉得这是

77:36

open source language models like we did some quins and the GLM and the stuff that other people are

open source language models，就像我们做了一些 quins 和 GLM，还有其他人在用的那些东西，

77:40

using and we taught them medicinal chemistry in ways where you say here is a structure describe it in

然后我们教它们 medicinal chemistry，方式是你说这里有一个 structure，用

77:45

natural English give me the diff between two very between two molecules in natural English and then

natural English 描述它，给我两个分子之间的 diff，用 natural English，然后

77:51

we got it to the point of where we say okay here are some molecules that are binders here are some

我们把它做到这个程度：我们说 OK，这里有一些 molecules 是 binders，这里有一些

77:55

that are not binders can you tell us why you think the binders are binding and then you can also give

不是 binders，你能告诉我们为什么你觉得这些 binders 在 binding 吗？然后你也可以给

78:00

it hard negatives so yeah have these two collections that they go okay well I think this chemical

它 hard negatives，所以对，有这两个 collections，它们就说 OK，我觉得这个 chemical

78:04

series has this you know and it will write these things down in and rebel English which we think is

系列有这个，你知道，然后它会把这些东西写下来，用 rebel English，我们觉得这是

78:08

great and then we said okay now that you sort of know these basic concepts we're going to give you 10

很好，然后我们说：行，既然你现在差不多懂这些基本概念了，我们会给你 10 个

78:14

binders and 10 non binders that are close ask you to decide why you think the binders are binding

binders 和 10 个很接近的 non binders，让你判断为什么你觉得这些 binders 会 binding

78:19

and now we're going to give you the opportunity to run your own experiments here is a catalog of

现在我们会给你机会自己跑实验，这里有一份目录，里面有

78:23

millions of molecules why don't you pick five or 10 and test your hypothesis about why they're binding

数百万个 molecules，你为什么不挑五个或 10 个，来验证你关于它们为什么会 binding 的假设

78:29

and so then the model will go into the data that we have already collected it will pull out what it

然后 model 会去看我们已经收集好的 data，它会挑出它

78:33

thinks are the most informative choices and we'll get back the answer instantly like it's a

认为信息量最大的选择，然后我们会立刻拿到答案，就像它是一个

78:37

coding sandbox and then it will update its hypothesis about the binding and then the ultimate goal is

coding sandbox，然后它会更新自己关于 binding 的假设，然后终极目标是

78:42

to get the best binder in as few steps as possible very much like a medicinal chemist would do and

用尽可能少的步骤拿到最好的 binder，非常像药物化学家会做的那样，然后

78:46

then we're just doing this based on your a priori run experiments but it's looking out for

然后我们只是根据你 a priori 运行的实验来做这个，但它是在寻找

78:50

a database that's right right so now so you're basically created reinforcement roll outs it is a

一个数据库，对，对，所以现在，所以你基本上创建了 reinforcement roll outs，它是一个

78:56

reinforcement rollout of a large scale that's growing all the time that's full of these activity cliffs

大规模的、一直在增长的 reinforcement rollout，充满了这些 activity cliffs

79:01

in full different variations and so now we have and so you can use this and since we know that when

有各种不同的变体，所以现在我们有，所以你可以用这个，而且既然我们知道当

79:06

you change the molecules in these small activity cliff ways it is a causal change like we know that

你以这些小的 activity cliff 方式改变分子时，这是一个 causal change，就像我们知道

79:11

when you put this thing on there the behavior changes in this reproducible measurable way and so you

当你把这个东西放上去时，行为会以这种可重复、可测量的方式改变，所以你可以

79:15

can also use these trajectories to come up with reasoning traces kind of at a pretty good scale

也可以使用这些 trajectories 来得出 reasoning traces，规模相当不错

79:20

and so right now we have these agents running through the data set we just give them 10 binders and

所以现在我们有这些 agents 在跑 data set，我们只给它们 10 个 binders 然后

79:25

10 almighters you give them a budget of 100 molecules and they'll go through you know probably like 102

10 个 almighters，你给它们 100 molecules 的预算，它们会走过，你知道，大概像 102

79:30

holes where they're just going through this long many step reasoning problem and then come up with

个 holes，就是一路穿过这个很长、很多步的 reasoning 问题，然后得出

79:35

what they think of the answers to the best at the end and so why do you think this is superior to

它们最后觉得最好的答案，所以为什么你觉得这比

79:40

something like Hermes or maybe like something you know sort of scaling up their Hermes approach where

像 Hermes 这样的东西更优，或者也许像某种，你知道，把它们的 Hermes 方法 scaling up 的东西，在那里

79:47

you're you know sort of building some you know attention. Yeah we think so we think it's much more

你，你知道，在构建某种，你知道，attention。对，我们这么觉得，我们觉得它更

79:52

resistant to the memorization problem and so like while we are very happy with where the attention

能抵抗 memorization problem，所以就像，虽然我们对 attention

79:57

goes with Hermes when you throw in all these mutants like as you add more parameters to a model like

在 Hermes 里怎么走非常满意，当你把所有这些 mutants 都扔进去，就像你给一个 model 加更多 parameters 时，就像

80:02

they're going to memorize the mutants too right like and we want this thing to really understand

它们也会 memorize 这些 mutants，对吧，就像，我们希望这个东西真正理解

80:06

like these causal changes and it seems like using natural language is like a pretty good regularizer

就像这些 causal changes，而且看起来用 natural language 就是个挺不错的 regularizer

80:11

for the problem and like we are unaware of other problems where you have a reinforcement learning

对于这个问题来说，而且就好像我们并不知道还有别的问题，是那种你有一个 reinforcement learning

80:16

loop that is physically grounded that is this scalable this fast like you can imagine for you know when

loop，它是 physically grounded 的，能这么 scalable、这么快，就像你能想象，你知道当你

80:22

you're doing lab and loop stuff where you're going off and making robots and going off to these

你在做 lab 和 loop 这些东西，你要出去造机器人，跑去处理这些

80:27

problems they will also necessarily require like limited resources they will still make some choices

问题的时候，它们也必然需要像有限的资源，它们仍然会做一些选择

80:32

about what to make in the first place what not to make in the first place and then doing those loops

关于一开始要造什么、一开始不造什么，然后跑那些 loops

80:36

will take a little bit longer and if there's a scalable way to measure that landscape and then

会花稍微更长一点时间，而如果有一种 scalable 的方式去衡量那个 landscape，然后

80:41

continually make that bigger and bigger and bigger and to allow the agents to play in it we think it's

不断把它搞得越来越大、越来越大、越来越大，让 agents 在里面玩，我们觉得它

80:45

probably going to go a lot faster for this particular problem that's interesting yeah I mean you

对这个特定问题来说，可能会快得多，这挺有意思的，对，我是说你

80:48

I would say it's it's Excel from Zera is getting there but they're not actually language there but

我会说，它是，Excel from Zera 快成了，但它们其实并不是真正的 language，但

80:55

they are encoding as a sequence so it's yeah in some ways the perturbation is similar but this is

它们是在编码成一个 sequence，所以，对，在某些方面 perturbation 是相似的，但这是

81:03

yeah this is different and then Lila also rhymes with this but it does like all these guys are really

对，这不一样，然后 Lila 也跟这个有相似之处，但确实，像这些人都真的

81:09

smart and pushing like these automation meetings yeah but but they're but they're they're not

很聪明，在推这些 automation meetings，对，但，但他们在，但他们在，他们没

81:14

doing these pulled screens and then using that as they have it as far as they know having figured

做这些 pulled screens，然后把那个拿来用，就他们所知，已经搞清楚

81:22

out how to make that into just a database look up so you can do huge number of roll outs yes yes

怎么把它变成只是一个 database lookup，这样你就能做大量的 roll outs，对，对

81:28

so right now the way that we like to think about it is if you're a medicinal chemist and you're

所以现在，我们喜欢这样想：如果你是药物化学家，而你

81:32

working at a pharmaceutical company you do this thing called a structure activity relationship

在制药公司工作时，你会做一件叫 structure activity relationship 的事

81:36

and what that means is oh when I change the molecule in this way it changes the behavior and so I'm

意思就是，哦，当我这样改变 molecule 的时候，它的表现就会变，所以我

81:41

trying to optimize for these different properties and you go through what's called an S.A.R.

想针对这些不同的 properties 做优化，然后你会经历所谓的 S.A.R.

81:44

campaign that's early drug development it's making them all cool better you might see a couple

campaign，那就是 early drug development，就是把它们都变得更酷、更好，你可能见过

81:48

dozen of these maybe 50 of these in your career and so right now you know last weekend we pulled out

几十个这种，也许在你的职业生涯里见过 50 个，所以现在，你知道，上周末我们弄出了

81:54

over 10,000 of them right so yeah and so now is the opportunity now we'll be restricted to the

超过 10,000 个，对吧，所以是的，现在机会来了，现在我们会受限于

82:01

space of things that we have screened and then at the end you can ask the models to make additional

我们已经 screened 过的那些东西的范围，然后到最后你可以让 models 做额外的

82:06

choices from say catalogs and then you can go you know in the future go order those and

选择，比如说从目录里选，然后你可以，你知道，将来去订购那些，然后

82:10

retest them like you might for like a traditional lab in the loop approach but right now we give

重新测试它们，就像你可能对传统的 lab in the loop approach 做的那样，但现在我们给

82:16

these agents 10 binders give them a budget of a hundred and they'll get the best binders out of

这些 agents 10 个 binders，给它们一百的 budget，它们会从

82:20

six and a half million in that budget and so it seems like it's working yeah so like like tell

六百五十万里挑出最好的 binders，在那个 budget 内，所以看起来它起作用了。对，所以，呃，跟我们讲讲

82:25

us more about why you think this is I get the argument about it's a better regularizer it's

多说说你为什么觉得这是这样。我理解这个论点，说它是更好的 regularizer，它是

82:31

but also why do you believe that this is effective right so like well you can imagine

但还有，你为什么相信这是有效的呢？对吧，所以，呃，你可以想象

82:38

well it doesn't get things wrong so badly but it also doesn't get things that right but

嗯，它不会把事情搞错得那么离谱，但它也不会把事情搞对到哪里去，但

82:44

it's I think it's because you give it the opportunity to make all those choices I mean it's like

我觉得这是因为你给了它机会去做所有这些选择。我的意思是，这就像

82:49

running these coding sandboxes you just have like a really rapid reinforcement reward signal

跑这些 coding sandboxes，你只是有一个非常快速的 reinforcement reward signal

82:53

and but the thing is you also have to include the choices that a human being would not have made

但问题是，你还得把那些人类不会做出的选择也包含进去

82:58

because we're measuring all the molecules across all the proteins yeah so it has to be able to

因为我们是在测量所有 proteins 里的所有 molecules，对，所以它必须能够

83:02

make the counterintuitive decision and get a feedback on that it's got to have the counterfactuals

做出反直觉的决定，并得到关于这个的反馈，它必须得有 counterfactuals

83:07

it's got to make mistakes yeah and you need to make sure that the mistakes are you know present in

它必须犯错，对，而且你得确保这些错误，你知道，是存在于

83:10

the data or future potential mistakes for present in the data in ways that are not really biased

数据里，或者未来可能犯的错误也存在于数据里，而且方式不会真的带有偏见

83:15

for it to you know learn chemists sociology fingerprints yeah that's really fascinating so through

让它，你知道，了解化学家、社会学、指纹，对，这真的很有意思，所以通过

83:20

all of the you know language pre-training corpus it kind of has an intuition for some basic chemistry

所有那些，你知道，language pre-training corpus，它多少对某些基础化学有一种直觉

83:25

that I think most participants might have an idea of and now you're giving it sort of like an

我觉得大多数参与者可能都有一点概念，而现在你又给了它一种有点像

83:30

infinitely fast love in the loop that's correct so maybe reasoning to always say that reasoning

无限快的 love in the loop，这没错，所以也许 reasoning 就是总要说那个 reasoning

83:35

skills are might even be more important right the ability to say oh in my reinforcement campaign

技能甚至可能更重要，对吧，就是那种能在我的 reinforcement campaign 里说“哦”的能力

83:41

I've learning campaign I've learned things and now now that allows me to reason about the next

我在 learning campaign 里学到了东西，现在，现在这让我能推理下一个

83:46

yep harder question and then just get making choices yeah you know and and you know right now

对，更难的问题，然后就只是做选择，对，你懂的，而且，而且，你懂的，现在

83:51

we're like we're really excited about the really signal and also you know like we're at

我们就像，我们对那个真的很 signal 的东西特别兴奋，而且，你懂的，我们处在

83:54

tiny shop right there's only ten of us like I'm really looking forward to when you know our

一个小店，对吧，我们只有十个人，就像，我真的很期待，你懂的，当我们的

83:58

experimental corpus grows it you know like 2 x 10 x you know like a hundred x the rate that

experimental corpus 增长的时候，你懂的，像 2 倍、10 倍，你懂的，像 100 倍，那个速度

84:03

it currently does because then like the amount of space for these agents to explore like

它目前确实是这样，因为那样的话，就像这些 agents 可以探索的空间量，就……

84:07

expand dramatically let's talk about that space because so my understanding and this is from me

大幅扩展。我们来聊聊那个领域吧，因为，呃，我的理解是——而且这是我自己的理解——

84:12

like you know talking to cloud and googling and whatever but my understanding is that Del's are

比如，你知道，跟 cloud 聊、用 Google 搜之类的，但我的理解是，DELs 是

84:17

like limited in that they are they tend to be create flat molecules or flatter molecules

受限的地方在于，它们——它们往往会生成平面的分子，或者更扁平的分子

84:24

depends on the building blocks okay it depends on the building blocks yeah so there are you can

这取决于 building blocks，好吧，取决于 building blocks，对，所以，你可以

84:27

imagine that you will be somewhat constrained by designs in terms of like oh I'm going to get

想象一下，你会多少受到设计的限制，比如，哦，我要拿到

84:32

some number of building blocks usually two to three and then stitch them together in some way

一定数量的 building blocks，通常是两到三个，然后把它们以某种方式拼起来

84:37

and so like we'll maybe do chemical library design choices yeah typically it's like like one one

所以，比如说，我们可能会做 chemical library design 的选择，对，通常就像，就像，一个，一个

84:42

that we run a lot right now is a forked design it's like a y-shaped molecule and then there's

我们现在经常跑的一种，是 forked design，它就像一个 y-shaped molecule，然后还有

84:46

other ones that are linear shaped molecules there's yet other ones that are kind of like circular

其他的就是 linear shaped molecules，还有另外一些有点像 circular

84:50

kind of macro cycle shaped things and then you have to think about where is the DNA going to go

有点像 macro cycle 形状的东西，然后你得想 DNA 要往哪儿放

84:54

and like how is that going to cause me problems and then also the the commercial availability of

然后就像，这会怎么给我带来问题，然后还有那个那个……的商业可获得性

84:58

building blocks and the chemistry is required to stitch this together you can go for custom stuff

building blocks，以及把这些拼接在一起所需要的 chemistry，你可以选定制的东西

85:03

takes longer it's kind of more appealing to us to go faster yeah and there's certainly going to be

花的时间更长；对我们来说，能更快一点更吸引人。对，而且肯定会有

85:07

corners of chemical space that we're probably bad at and then you have to think about well future

chemical space 里我们可能不太擅长的那些角落，然后你得想，嗯，未来的

85:11

libraries should have this kind of geometry and then I'm going to think about the different proteins

libraries 应该有这种 geometry，然后我会去考虑不同的 proteins

85:14

are going to throw at it and there'll be like a pretty uneven distribution of like a whole chemical

会往它上面扔，而且会有一种相当不均匀的分布，像一整个化学

85:20

library made this way might be really good for these types of targets it might be really crappy

用这种方式做出来的 library，对这类 targets 可能真的很好，它可能真的很烂

85:24

for these other types of targets and so then you get to think about all right one of my what

而对另外那些类型的 targets 来说，然后你就会去想：好吧，我的……有一个……什么

85:28

what corners of protein space am I optimizing for and those kinds of tensions yeah so so it sounds

什么 protein space 的角落是我在 optimizing 的？还有那些矛盾，对啊，所以所以听起来

85:33

like the argument is well we know that we're not optimizing the whole space but it's a lot more than

像是这个论点是：好吧，我们知道我们并没有在 optimizing 整个 space，但它已经比

85:39

it was before so that's better and then we are as we get better at this and understand or we have

以前多很多了，所以这就更好；然后随着我们在这件事上变得更好、更理解，或者我们有

85:46

bigger budget or whatever it is and we'll expand to these corners that we know that we're lacking

更大的 budget 或者不管什么，我们就会扩展到那些我们知道我们缺的角落

85:50

in but or maybe we don't even know we're lacking but we will find that out right exactly do we

但或者也许我们甚至都不知道我们缺，但我们会发现这一点，对吧，没错，我们

85:56

want to move on to the last base yeah you can go for it let's see how it goes all right so for

要不要进入最后一个 base？好啊，你来吧，看看会怎么样，好，那么对于

86:01

the audience context um Ian requested when I confirmed that this was his idea not my day

听众背景，嗯，Ian 要求的，当时我确认了这是他的主意，不是我的日子

86:08

I was ready a don't ask me or come out me bro section where we we ask the hardest problems

我当时准备好了一个“别问我，或者来怼我啊兄弟”的环节，我们会在那儿问最难的问题

86:14

we can cut we can have a what which is we only have to probably tell is not very hard for me

我们可以剪掉，我们可以有个什么，就是，我们可能只需要说，这对我来说不是很难

86:19

yeah I think I've asked my hard questions already we only had about 24 hours to prepare so I

对，我觉得我已经问过我的难题了，我们只有大约 24 小时准备，所以我

86:25

think most of these will be Claude's hardest questions but anyway we'll we'll see how this goes yeah

觉得这些大多会是 Claude 最难的问题，不过不管怎样，我们、我们看看会怎么样吧，对

86:29

the context for this as I was at um I was at a cancer mucker therapeutics meeting last week and uh

这个的背景是，我当时在，嗯，我上周在一个 cancer mucker therapeutics 会议上，然后呃

86:35

and the the format of the beating is half of the time is for content presentation and the other

而且这个挨批的流程是，一半时间是用来做内容展示，另一半

86:40

half is for very aggressive grilling uh and it was uh it was it was horrifying and inspirational and

是用来做非常激烈的拷问，呃，而且那，呃，那真的既吓人又鼓舞人心，而且

86:46

so uh here I am willing to do it and public let's do this dals is uh RJ was just alluding to oftentimes

所以呃，我在这儿，愿意做，而且公开，咱们就干吧；dals 呢，呃，RJ 刚刚也提到，它往往有非常特定的 properties，嗯，你可以在优化 indel 上做得非常好，但最终它们经常在某些事情上失败，而那些很难转化到真正的药物上。对，所以某些 failure modes 就是，比如任何容易组装的东西也容易拆解，这对你知道的 ADMET properties 很糟糕；你能继续构建的空间就那么多，因为你已经有点用掉了一些像 combinatorial building blocks 之类的东西。总之，这在长期的 out-of-library generalization 方面是怎么限制的？我觉得很大程度上会有点取决于 building block availability，以及你愿不愿意进入 custom synthesis。所以如果我们……

86:52

had very specific properties um and you can get very good at optimizing indel but ultimately they

有非常特定的性质，嗯，而且你可以变得非常擅长优化 indel，但最终它们

86:58

frequently fail at certain things which are hard to translate to actual medicine yep so certain

经常在某些事情上失败，而那些事情很难转化到实际医学中，是的，所以某些

87:04

failure modes being such as anything which is easy to assemble is easy to disassemble bad for

失败模式比如，任何容易组装的东西都容易拆解，这对

87:09

you know ad met properties you only have so much room to kind of build on because you've already

你知道，ADMET properties 来说是不利的，你只有那么多空间可以往上构建，因为你已经

87:14

sort of use some like combinatorial building blocks and so on anyway how is this limiting in terms

算是用了一些类似 combinatorial building blocks 的东西，等等，总之，这怎么构成限制，就

87:19

of like long term out of library generalization I think a lot of it is going to be somewhat dependent

长期的 out-of-library generalization 而言，我觉得很大一部分会有点取决于

87:25

on like building block availability and like ones willingness to go into custom synthesis so if we're

像 building block 的可得性，以及人们有多愿意去做 custom synthesis，所以如果我们

87:30

especially for the fine grain mapping stuff you know being able to say move a nitrogen all the way

尤其是对于 fine grain mapping 这些东西，你知道，能说把一个 nitrogen 一路

87:35

around the ring at all one of those positions is something that's very very desirable but it's also

绕着 ring 移到其中任何一个位置，这件事非常非常理想，但它也

87:38

something that's pretty unlikely that you're going to get like commercially for all sorts of

是你不太可能像商业化那样搞到的，对于各种

87:42

building blocks you're currently generating so like is that going to be a failure mode for sure

你现在正在生成的 building blocks，所以这肯定会是一个 failure mode 吗？

87:45

yes yes I would I would love to be able to get more of those and like beef up our custom synthesis

是的是的，我我真的很想能弄到更多那些，然后像加强我们的 custom synthesis

87:50

that's something that we're looking at uh but right now trying to get those as challenging

那是我们正在看的事，呃，但现在想弄到那些还是挺有挑战的

87:54

one critique that clawed might be thinking about in terms of molecules easy to put together into

clawed 可能会思考的一个批评点是，关于 molecules 容易组装和

87:58

disassemble for many years it was uh the most common way to attach uh building blocks to new

拆解。很多年来，它是，呃，把 building blocks 接到新的上面最常见的方式

88:03

adult chemistries is a might balance or protein bonds and so those are quite digestible uh you

adult chemistries 是一种可能平衡或 protein bonds，所以那些相当容易消化，呃，你

88:07

know in in in the body but in the past few years uh they've really really passed like four-ish years

知道，在在在体内，但过去几年，呃，真的真的已经过了大概四年左右

88:13

there've been like a really nice um fluorescence of chemistries that are friendly to DNA that don't

出现了，就像，一种非常不错的，嗯，fluorescence of chemistries，它们对 DNA 友好，不会

88:19

destroy it during the untouchable process because that's very important that will give you carbon carbon

在 untouchable process 期间破坏它，因为那非常重要，那会给你 carbon carbon

88:23

bonds carbon nitrogen bonds these kinds of things so it's actually pretty good for work a day simple

bonds、carbon nitrogen bonds 这类东西，所以它其实相当适合日常简单的

88:28

chemistry yeah so like I'm actually not too worried about that part and more worried about building

chemistry，对，所以就像，我其实不太担心那部分，更担心 building

88:33

block diversity uh and their sources from that yeah what about admin properties in general like

block diversity，呃，以及它们从那里来的来源，对，那 admin properties 总体上怎么样，就像

88:38

you know you you've been focusing very much on binding we have yeah so uh projects that we've done

你知道，你你一直非常关注 binding，我们有，对，所以，呃，我们做过的项目

88:44

uh trying to look at reasonable admin and like PKPD properties for what it's worth we've gone

呃，试着看看合理的 admin 以及类似 PKPD 的性质，不管怎么说，我们已经

88:48

directly from uh from chemical library hits to good PKPD in amounts in six months with a very small

直接从呃，从 chemical library hits 到在量上有良好 PKPD，只用了六个月，而且只用了非常少的

88:54

number of webcam cycles so at least chemical material that we currently have or at least that

webcam cycles，所以至少我们目前有的 chemical material，或者至少

88:58

you know those few series that we did that for uh you know they turned out pretty good in our hands

你知道，我们为此做的那少数几个 series，呃，你知道，在我们手里结果还挺好的

89:02

I guess you know we will continue to see like how that modeling looks and also of course we want to do

我猜你知道，我们会继续看看那个 modeling 看起来怎么样，当然我们也想做

89:06

more of these like and that target models and then to get about how to incorporate other types of

更多这类，呃，target models，然后去弄清楚如何把其他类型的

89:10

you know more emergent models along with our designs was that speed driven by your ability to create

你知道，更 emergent models 和我们的 designs 结合起来——那个速度是因为你们能创建

89:15

custom like local combinatorial libraries or was that speed driven like what allowed you to get

custom 的、像 local combinatorial libraries 这样的东西，还是说那个速度是由，比如，什么让你能够做到

89:20

there in six months like um the drift so we when we design this chemical libraries at the beginning

六个月后，就像，嗯，那种 drift，所以我们在最开始设计这些 chemical libraries 的时候

89:25

we very much optimize that space uh for drug like this and so we think that's part of it is that

我们很大程度上是在优化那个空间，呃，针对像这样的 drug-like，所以我们觉得这是其中一部分，就是

89:30

was like a really good starting point you know when we do our chemical library designs uh we have

是一个非常好的起点，你知道，当我们做 chemical library 设计的时候，呃，我们有

89:34

this uh kind of like Thompson sampling level approach where we come up with an optimization function

这个，呃，有点像 Thompson sampling 层面的方法，我们会想出一个 optimization function

89:40

you say okay well I want the libraries to have these properties and so some of those properties are

你会说，好吧，我希望这些 libraries 具备这些 properties，而其中一些 properties 是

89:43

going to be I want to bind these things I want to avoid these things and then some of these properties

会是，我想 bind 这些东西，我想避开这些东西，然后其中一些 properties

89:47

are going to be oh there's other admin model approaches that you can use to make predictions

会是，哦，还有别的 admin model 方法，你可以用它们来做 predictions

89:52

about how your chemical material might turn out and so you put that as part of the optimization

关于你的 chemical material 最后可能会变成什么样，然后你把它作为 optimization 的一部分

89:56

and then what you do is you say okay have all the chemicals that I could make given my geometry

然后你要做的就是，你说，好，把所有我根据我的 geometry 能造出来的 chemicals 都列出来

90:01

I don't want to make that many I want to make a smaller number uh and so you you get you basically

我不想造那么多，我想造少一点，呃，然后你就，你就，基本上

90:06

set up and generate a bunch of synthetic libraries and then you score them on the optimization

设置好，然后生成一堆 synthetic libraries，接着你用 optimization

90:10

function and so we do that with this library uh and the starting material look pretty good so I'm

function 给它们打分，所以我们用这个 library 这么做，呃，starting material 看起来相当不错，所以我

90:13

sure that we'll continue to do that. Along those same lines has anything that you've done like made it

确定我们会继续这么做。顺着同样的思路，你做过的任何东西有没有，比如，推进

90:20

to so it made it to animals. Yeah so pkpd and a mouse is that's yeah so what that what that means

到——所以它已经做到动物实验了。对，所以 pkpd 和小鼠，那是，对，所以那，那是什么意思

90:25

is uh can you dose a mouse with the thing uh and then see how it gets into the mouse's body

是，呃，你能用这个东西给小鼠 dose 吗，呃，然后看看它怎么进入小鼠的身体

90:30

and tissues. Yeah so the pkpd is pharmacogenetics, pharmacodynamics, dynamics,

和组织。对，所以 pkpd 是 pharmacogenetics、pharmacodynamics、dynamics，

90:35

training. Yes. How does the body move the the drug through the or how does a metabolism move the

training。对。身体是怎么把 drug 送过……或者说 metabolism 是怎么把

90:41

body the drug through the body? Yeah and then there are biomarkers of the drugs intended activity

drug 送过身体？对，然后还有 biomarkers，用来反映 drug 预期要产生的 activity

90:46

and so what that means here is this is like a drug this is a compound we're working off really the

所以这里的意思是，这就像一种 drug，这是一个 compound，我们真正基于的其实是

90:50

inflammation and so we were following uh cytokine levels in mouse blood and they and they did what we

inflammation，所以我们在追踪呃小鼠血液里的 cytokine levels，而且它们，它们做了我们

90:55

wanted them to do and expected them to do based on the targets of the drug. So Sean

希望它们做的事，也预期它们会根据 drug 的 targets 做的事。所以 Sean

91:00

likes to ask when money I want to ask when people. Yeah I mean for those we talked to we have

喜欢问钱是什么时候，我想问人是什么时候。对，我是说，对那些我们聊过的人，我们有

91:08

a couple of these things uh that we're talking to folks about partnering out to spin up new entities

几个这样的东西，呃，我们正在跟一些人谈，合作出去 spin up 新的 entities

91:11

to go chase them down. Okay and so yeah and of course like all these things take like sometimes you've

去追它们。好，所以对，当然，所有这些事情都需要，就像有时候你已经

91:15

got to do like there are these traditional things forward what you declare what's called like a

你得做，就像有这些传统的事情要往前推进，你要声明所谓的像是一个

91:20

development candidate then you've got to do a bunch of other work about manufacturing and additional

development candidate，然后你还得做一堆关于 manufacturing 和额外的

91:24

safety work before I can get into people so all these things take a lot longer. Yeah and so you're

safety work，然后我才能进入人体，所以所有这些事情都要花长得多的时间。对，所以你要

91:28

going to spin out the idea is to spin out a so people maybe even internal they will say like I

去 spin out，这个想法就是 spin out 一个，所以人们，可能甚至内部的人会说，我

91:34

really want to take this this thing to market or try to get it like take it to the next step.

真的很想把这个东西推向市场，或者试着让它进入下一步。

91:38

Yeah I mean one way that we kind of think about it is um you know we built this really great engine

对，我的意思是，我们有点像这样想的一种方式是，嗯，你知道，我们建了一个特别棒的 engine

91:43

for like data generation and and model training and and now as we've talked about like these

用来做 data generation 和 model training，而现在就像我们聊过的这些

91:48

reinforcement learning reasoning approaches um and most other domains when you build an engine to go

reinforcement learning reasoning approaches，嗯，而在大多数其他领域，当你建一个 engine 去

91:53

do that stuff to go like help on those problems um the substrate that you're generating along the

去做那些事，像是去帮忙解决这些问题，嗯，你在这个过程中生成的 substrate 本身并没有内在价值。

91:58

way is not itself intrinsically valuable no one wants to sell reasoning traces about like political

没人想卖关于政治理论之类的 reasoning traces，但是，就像，因为做了所有这些 screening，我们会生成一些小小的金子。

92:04

theory but like there are going to be a little bits of gold that we had generated as a consequence

因为你实际上不只是在推理这个 model 好——或者抱歉，是这个 molecule 好——而是在推理这里就是它为什么好，这里就是它为什么好。

92:10

of doing all this screening because you're actually reasoning about not only that this model is good

而且这里有我们专有的化学材料，它有，你知道，针对某些 protein target classes 的 activities，而那些 target classes 别人可能觉得很难 drug，对吧。

92:15

or sorry this molecule is good but this here is why that's good here is why that's good and here is

所以你就有了 starting material，你知道，现在我们有 starting material，看起来对我们挺 selective 的，针对，呃，我想。

92:21

chemical material that's proprietary to us that has you know activities that are you know protein

chemical material是proprietary to us的，它有你知道的activities，是你知道的protein target classes，别人可能觉得很难drug的，所以你有了starting material

92:26

target classes that others might find hard to to drug right and so you've got starting material

for你知道现在我们有starting material，看起来对我们pretty selective，对呃我觉得

92:30

for you know right now we have starting material that looks pretty selective to us for uh I think

这是对的，kind of turns out那些proteins非常hard to drug，因为各种原因

92:34

like 700 targets like it's a norms number uh you don't want to chase all those down uh but you know

比如 700 个 targets，就像它是个 norms number，呃，你不想把所有那些都追到底，呃，但你知道

92:39

as as opportunities they are all opportunities with different levels of you know kind of economic

作为机会来说，它们都是机会，只是有不同程度的，你知道，那种 economic

92:44

value capture and desirability and so you think all right well we've got all this stuff like let's see

value capture 和吸引力，所以你会想，好吧，我们手上有这么多东西，那就看看

92:49

what we can get up let's see what we can get out of it discovery is like a small percentage

我们能搞出什么，看看我们能从中得到什么，discovery 大概只占很小一部分

92:55

of the total cost of developing a drug and you can argue that well yeah well any given discovery

在开发一款药的总成本里，你可以争论说，嗯，是啊，任何一次 discovery

93:03

is cheap maybe but that you know because of the failures you have a lot of them so but even then you

也许很便宜，但你知道，因为失败，你会有很多失败，所以，但即便这样，你

93:09

know to get one drug one drug like all the steps of a single drug candidate that is successful is

知道，要得到一款药，一款药，就像单个成功的 drug candidate 的所有步骤，是

93:18

maybe you know on the low side 200 million on the high side of billion um depending on who you

也许，你知道，往低算两亿，往高算十亿，嗯，取决于你

93:25

have to test on and lots of other factors but and then you know the discovery is a couple million

还得在上面测试，还有很多其他因素，但然后你知道，这个 discovery 就是几百万

93:32

so so are you tackling the right problem yeah um and so I think there are a number of

所以所以，你是在解决正确的问题吗？是的，嗯，所以我觉得有不少

93:39

early discovery problems that are worth that much you know like like reasonable amounts and so

早期的 discovery 问题，值那么多钱，你知道，就，就还算合理的金额，然后所以

93:45

what does that mean there's a bunch of targets right now with like really good genetic evidence

那意味着什么呢？现在有一堆 targets，有那种非常好的 genetic evidence

93:48

that are very challenging to drug with small molecules um you know a pretty straightforward example

这些非常难用 small molecules 来 drug，嗯，你知道，一个挺直白的例子

93:53

is uh for a lot of these um uh kind of cytokine inflammatory uh indications folks are using antibodies

就是，呃，对于很多这些，嗯，呃，算是 cytokine inflammatory 的，呃，indications，大家在用 antibodies

93:59

to drug them and going after a small molecule therapeutic that you could take a sipill instead of an

来 drug 它们，然后去做一种 small molecule therapeutic，你可以服用 sipill，而不是一个

94:05

infusion one that doesn't require uh like cold chain you know transport uh to get to different patient

infusion 的那种，不需要，呃，像是 cold chain，你知道，运输，呃，送到不同患者那里

94:10

populations like having those things is enormously desire meeting you have to have refrigerated

对于人群来说，拥有这些东西是非常令人渴望的，意味着你得有冷藏卡车，是啊，对，是啊。

94:15

trucks yeah right yeah like and like if you want to get these if you want to get these antibody-based

就像，然后就像，如果你想把这类东西，如果你想把这类 antibody-based drugs 给到其中一些人，这就像是一个非常难的问题，是啊，呃，这很贵，而且非常贵。

94:19

drugs to some of these folks like this is like a very hard problem yeah uh it's expensive and very

所以，所以大家就真的很兴奋：你能不能找到一种 small molecule therapeutic，能做到同样的事，呃，然后针对这些的 antibody，对吧。

94:23

expensive and so and so folks are like really quite excited about can you find a small molecule

所以有些 antibody，它有效，是的，然后你就参与其中，然后你就想，为什么我们不能直接针对同一个 target 呢。

94:27

therapeutic that does the same thing uh then antibody against these right so there's some antibody it's

没错，结果发现那些 proteins 因为各种原因非常难 drug。

94:31

it works yep and so you're in and so you're like well why can't we just go after the same target

但我们愿意认为，你知道，我们正在开发的 technology 能，能拿到好的 material。

94:37

that's right kind of turns out that those proteins are very hard to drug for variety of reasons

但我们like to think that你知道我们正在developing的technology能够能拿到good material

94:41

but we like to think that you know the technology that we're developing can can get good material

internal data only，并且我们能够extend并且做出generalized predictions，进入到

94:44

and so we do in fact have selective material for a lot of these kinds and so that's one example

所以实际上，对于很多这类东西，我们确实有 selective material，这就是一个例子

94:48

of something that's like a very desirable target where of course like there's going to be much more

是那种非常理想的 target，当然，要做到那一步，会有多得多

94:53

a capital requirements to like get it across the line to actually get it to market but the interest

的资本要求才能把它推过线，真正推向市场，但兴趣

94:57

is sufficiently high enough so you're going to ruin the antibody market basically like if you can

已经足够高，所以你基本上会把 antibody 市场搞垮，如果你能

95:02

like that it's good for some stuff like it's great for some stuff and like there's a lot of energy

做到的话，它对某些东西挺好，对某些东西很棒，而且有很多精力

95:06

being thrown you know at those groups because we're good at protein fully right and so like they're

被投入到那些团队里，你懂的，因为我们擅长 protein fully，对吧，所以他们

95:09

making a progress and maybe more broadly I would say is like going after the right problem

正在取得进展，也许更广泛地说，我会说这就是在追正确的问题

95:14

it is a little bit of a lamppost problem where the things where we can make enough scalable

这有点像路灯问题，就是那些我们能做出足够 scalable 的东西

95:19

measurements in the life sciences are these really problems we can't really run a million mice

生命科学里的测量，这些真的是那种问题吗？我们没法真的跑一百万只小鼠

95:25

experiment right like we can't run many millions of phase one trials and so we're going to get better

实验，对吧，就像我们没法跑几百万个 phase one trials，所以我们会变得更好

95:31

at that stuff we're going to start to like go up these different layers of biological abstraction

在这些事情上，我们会开始像这样往上走，走过这些不同的 biological abstraction 层级

95:35

with like better data and better reasoning of physics models but like right now like the

用更好的数据，以及更好的 physics models 推理，但就像现在，就像

95:38

the technologies that we have to really make the number of measurements that we need are going to

我们真正必须用来做出我们所需数量的测量的那些技术，会

95:42

be like these early approaches I think yeah yeah yeah and I would argue also that you can if you

像这些早期方法，我觉得。对呀对呀对呀，而且我还会说，你可以，如果

95:48

have a oracle some AI oracle that says oh this like all the admin properties will be good and this

有一个 oracle，某个 AI oracle 说，哦，这个，就像所有 admin properties 都会不错，而且这个

95:54

generalizes to exact this population so you have this oracle filter then all of your failures

generalizes 到正好这个 population，所以你有了这个 oracle filter，那你所有的失败

96:00

in clinical trials go away yeah you can imagine a world where you can come up with an org you have

在 clinical trials 里都消失，对，你可以想象这样一个世界：你能搞出一个 org，你有

96:04

some new target that you know some new evidence of covers uh and then the org could give you small

某个新 target，你知道，有一些新证据能覆盖，呃，然后这个 org 能立刻给你 small

96:08

molecule material instantly that you would feel comfortable running straight to like a mouse study

molecule material，让你觉得可以直接跑去做，比如说，一个 mouse study

96:11

right right so that requires training on data from mouse studies clinical trials whatever

对对，所以这需要在 mouse studies、clinical trials 之类的数据上训练

96:18

I mean maybe I mean I could say definitely like admet properties if you're confident in the target

我是说，也许，我是说，我肯定可以说，比如说 ADMET properties，如果你对 target 有信心

96:23

and you need to make sure that it's not hitting all these other off targets uh but even then like

而且你需要确保它不会命中所有其他这些 off targets，呃，但就算那样，也

96:27

that's you probably don't need clinical results to feel the confidence to run off and run a

那就是，你大概不需要 clinical results 就能有信心，直接跑去做一个

96:31

mouse study right away all right well we're on you everyone's favorite pet question my favorite pet

mouse study，马上。好吧，我们现在进入大家最喜欢的那个问题，也是我最喜欢问的

96:35

question is what I call the cusserie paradox of machine learning and uh

问题就是我所谓的 machine learning 的 cusserie paradox，以及呃

96:38

discovery yes yes um okay so uh it's re-cusserie lesson great tweet I love this

discovery，对对嗯好吧所以呃这是 re-cusserie lesson，很棒的推文，我喜欢这个

96:45

thing it was uh yeah yeah I mean he was honest a few weeks ago we hopefully will release his

东西，它是呃是啊是啊我是说几周前他很坦诚我们希望能发布他的

96:49

episode in the next few weeks um sadly I actually didn't even bring this up on the

这期节目在接下来几周里嗯遗憾的是我其实都没在

96:54

episode I'm so disappointed anyway uh but yeah the tweet is basically what is the business model

那期节目里提到这个我太失望了反正呃但是啊这条推文基本上就是商业模式到底是什么

97:00

in machine learning for drug discovery because in order to build a good machine learning model

在 machine learning 里用于 drug discovery，因为要构建一个好的 machine learning model

97:05

you need lots of data in good high quality data at scale you clearly have that but then if you

你需要大量 data，需要高质量、大规模的 data，你显然有这些，但接下来如果你

97:11

need that to train the model like what do you need the model for because you just go collect the stuff

需要那些来 train 这个 model，那你到底要这个 model 干什么呢，因为你直接去收集这些东西就行了

97:15

and you clearly can collect this stuff quite well so yeah we talked a lot about data collection we

而且你显然能把这些东西收集得相当好，所以，对，我们聊了很多关于 data collection 的事，我们

97:20

talked a lot about machine learning and I I'm actually kind of curious in some ways like where does

聊了很多关于 machine learning 的事，而且我，我其实在某些方面有点好奇，比如说，到底在哪里

97:26

the actual machine learning is it necessary for your business model so I mean of course we depend

实际的 machine learning 呢，它对你们的商业模式是必要的吗？所以我是说，当然我们依赖

97:30

on the business model I think if your only rationale is to go off and get good chemical material

商业模式。我觉得如果你的唯一理由就是去搞到好的 chemical material

97:36

for a particular target then of course we is right right and you you you would have it you could get

针对某个特定的 target，那当然，我们是对的，对吧，对吧，而且你、你、你就会有的，你就能

97:40

started if you had enough chemical material broadly cast across all these other targets in a way

起步了，如果你有足够的 chemical material，广泛覆盖所有这些其他 target，以某种方式

97:47

that can uh uh you know kind of drive a deeper machine understanding of the chemical behaviors

这种方式，呃，呃，你知道，能某种程度上推动对 chemical behaviors 更深的 machine understanding

97:53

than the way by which you optimize those early materials have the potential to go much faster

比起你 optimize 那些早期材料的方式，它有潜力快得多

97:58

especially if you can do it at the scale that we've done it especially if you can do it at like the

尤其是如果你能在我们做到的那种规模上做到，尤其是如果你能在像

98:01

low capitol infusions that we've been able to do this we have not raised a lot of money at least

我们能够做到的那种低资本注入下做到这一点，至少我们没有融到很多钱

98:04

and we screen it in an gross amount and then I think getting even you know kind of even stepping back

而且我们以很大的总量去 screen 它，然后我觉得，甚至，你知道，甚至可以说退一步看

98:09

you know one of the reasons that I brought up this language model stuff is because we are starting

你知道，我提到这些 language model 东西的原因之一，是因为我们开始

98:13

to wonder if it would enable uh you know a much larger sort of like scope of capabilities

想，它能不能带来，呃，你知道，一种大得多的、那种能力范围

98:20

than just identifying novel chemical material like if we get really good at chemical reasoning

而不只是识别 novel chemical material，比如如果我们真的擅长 chemical reasoning

98:25

then we can get really good at lots of different industries in their applications without

那么我们就能非常擅长很多不同行业及其应用，而不用

98:30

having to physically go and screen in all those different domains like what's an example of that uh

必须亲自去所有那些不同的 domains 里 screen，比如，那有什么例子吗，呃

98:34

so let's say we get really good at making these predictions then we can run off and go work in agriculture

那假设我们真的很擅长做这些 predictions，那我们就可以跑去搞农业了

98:39

like immediately right you could go work in veterinary science you go work off material science

就马上，对吧，你可以去搞 veterinary science，你可以去搞 material science

98:44

and you don't have to if you're really good at it the you don't have to screen all those proteins too

而且如果你真的很擅长这个，你也不一定非得……你也不用 screen 所有那些 proteins

98:48

because you have a good fund amount of understanding of this right and the same thing could be argued

因为你对此已经有相当好的理解，对吧，而且同样的说法也适用于

98:53

for like all the protein folding companies right it's like now that these models are good

就像所有 protein folding 公司，对吧，现在这些 models 已经很强了

98:56

there are teams that are going after protein targets that have not been screened yet or a very

有些团队在追那些还没被 screen 过的 protein targets，或者非常

99:00

challenging two screen and getting like really good starting material like the novel guys did it

难 screen 的，然后拿到特别好的 starting material，就像 novel guys 做的那样

99:04

with this uh basically like a receptor on cells uh that has like a little tiny piece of a mutated

用这个，呃，基本上就像 cells 上的一个 receptor，呃，它有一小片 mutated 的……

99:11

cancer gene that's like really hard to do uh and they managed to pull it off and so you can

cancer gene，这就像真的很难搞，呃，他们居然成功做到了，所以你可以

99:15

imagine know that could be a world where we do this with small molecule material as well

想象一下，你知道，那可能是一个我们也能用 small molecule material 来做这件事的世界

99:18

and so like I don't think three is wrong uh in terms of finding like starting you know if you're

所以就像，我不觉得 three 是错的，呃，就找起点这件事来说，你知道，如果你

99:22

just trying to drug this one thing you screen a lot and then you already had the thing you

只是想 drug 这一个东西，你 screen 很多，然后你已经有了这个东西，你

99:25

don't need a model for the thing uh but that's not exactly the same as generalizing across all

就不需要为这个东西搞一个 model，呃，但那并不完全等同于 generalize 到所有

99:30

things and you know we are hopeful that if you have enough of this variation engineered activity

东西，而且你知道，我们很希望，如果你有足够多的这种 variation、engineered activity

99:34

cliffs that you will actually be at a point you could do all things in which case at which point

cliffs，你实际上就会达到一个点，你可以做所有事情，那样的话，到那个时候

99:39

you don't need to screen it all ever again yeah well and arguably you guys it's I don't I don't know

你就不需要再把所有这些都 screen 一遍了，是啊，嗯，而且可以说，你们，这个……我不，我不知道

99:43

exactly how much you raised so I mean you're for a campaign you're talking that's like a couple

到底融了多少钱，我的意思是，你是在为一个 campaign 募资，你说的这差不多就是几个

99:49

campaigns you that would cost the same amount so uh yeah yeah I mean maybe to put some real

campaigns 得花同样的钱，所以，呃，对，对，我的意思是，也许为了提供一些真实的

99:55

perspective on it um so we have uh developed and screened about two thousand different protein

参照，嗯，所以我们呃开发并筛选了大约两千种不同的 protein

100:01

constructs um a total of probably about five thousand screens in the history of the company across

constructs，嗯，公司历史上总共大概做了五千次 screen，覆盖

100:05

at least six and a half million molecules each um which I don't want to get into like a like a

每次至少六百五十万个 molecules，嗯，我不想把它搞成一种，一种

100:11

huge numbers game but it's certainly almost as much as anyone else on earth maybe more than

巨大的数字游戏，但这肯定几乎和地球上任何人一样多，也许比

100:16

everyone else probably not there's probably a couple groups that like if you know but like the

其他所有人都多，可能也不是，可能有一两个组，如果你知道的话，但就像

100:20

number of mutations I think it's probably more than everyone else that seems likely to me uh and

mutations 的数量，我觉得可能比所有人都多，这在我看来很可能，呃，而且

100:24

so yeah like we just put a huge amount of work into making this like very cheap and very scalable

所以对啊，就像我们刚投入了大量工作，把它做得非常便宜、非常 scalable

100:29

and if I was doing one at a time yeah if you went to CRO and you did one at a time it probably cost

而且如果我一次做一个的话，对，如果你去找 CRO，一次做一个，那大概要花

100:34

you a hundred K each depending on you know how much hand holding you got right yeah you just said

你每个 10 万，取决于，你知道的，他们手把手带了你多少，对吧，对，你刚说

100:39

five million six was six and a half million molecules yep that's not a lot of molecules especially

五百六十万，是六百五十万个 molecules，对，这不算很多 molecules，尤其是

100:43

since it's combinatorial space you know I'm gonna say 10 to the 40 year or something 10 to 60

因为它是 combinatorial space，你知道的，我会说 10 的 40 次方年，或者 10 的 60 次方

100:48

big numbers yeah I mean obviously you don't need to understand all the chemical space

大数字，对，我是说，显然你不需要理解整个 chemical space

100:52

no you just need to make drugs consistently repeatedly that's that's your goal but is that actually

不，你只需要持续、反复地做出 drugs，那才是你的目标，但这是否真的

100:58

enough chemical space to meaningfully do anything outside of your chemical space great question um

足够的 chemical space 来有意义地做你 chemical space 之外的事情吗？好问题，嗯

101:03

so the the the pleasant answer is um you know I I said earlier that we have trained models on

所以那个那个那个比较让人满意的回答是，嗯，你知道，我之前说过，我们只在 internal data 上训练过 models，而且我们能够扩展到 chemo，也就是 chemical space 中非常不同的角落，并做出 generalized predictions。所以，是的，我们觉得这够了吗？远远不够。你知道，这其实是我们未来 roadmap 的一部分：在不同角落的 chemical space 里做出更大量的 chemical libraries，并且大致覆盖我之前讲到的范围，其中一些是这些非常细粒度的变化，也就是我更欣赏的 activity cliffs，而另一侧则是这种大幅度的变化，试图捕捉新的 proteins 以及各种 protein shapes。抱歉，这不是一个难问题，我只是真的很好奇。

101:09

internal data only and that we're able to extend and make generalized predictions into

chemo，也就是chemical space非常不同的corners，所以yes，我们觉得这够吗，not

101:13

chemo which is very different corners of chemical space so yes do we think it's enough not

remotely，你知道所以这是我们roadmap往前走的part，就是making你知道larger numbers of

101:17

remotely you know it's so part of our roadmap of the forward is making you know larger numbers of

远程地，你懂的，所以这也是我们 forward roadmap 的一部分，就是让，你懂的，更大数量的

101:22

chemical libraries in different corners of chemical space and kind of spanning kind of what I

chemical libraries 分布在 chemical space 的不同角落，也有点覆盖了我

101:27

talked about earlier where some are like these very fine grain changes that I more appreciate

之前聊到的那些，其中有些就像是这些非常细粒度的变化，我更欣赏

101:30

activity cliffs and then the other side it's like this big course changes to try to capture new proteins

activity cliffs，然后另一边就像是这种很大的粗粒度变化，想试着捕捉新的 proteins

101:35

and sort of protein shapes this is sorry not not a hard question just I'm genuinely curious

以及 protein shapes 这类东西。抱歉，这不是，不是什么难的问题，我只是真心好奇

101:39

are you able to predict like potency like relative potency meaningfully like can you do KD curves like

你能有意义地预测 potency 吗，像 relative potency 那样，你能做 KD curves 吗，类似这样？

101:45

where you you know titrate like more more larger libraries or so people have wrestled with this

就是你可以，你知道，titrate 更多、更大的 libraries 之类的，大家在这个问题上也挣扎过

101:51

a little bit in terms of trying to get the KD directly out of Dell's doing things like maybe

也就是在尝试直接从 Dell's 里拿到 KD 这方面稍微试过一点点，比如可能

101:54

varying protein concentration or something that folks do one thing that we see is uh when we

改变 protein concentration，或者人们会做的类似事情。我们看到的一件事是，呃，当我们

101:59

when we do replicates of screens we get back the same hits and roughly the same order which makes

当我们做 screens 的 replicates 时，我们拿回来的 hits 是一样的，顺序也大致相同，这让我们

102:04

us you know kind of optimistic that you could get these kinds of things we started to move more

你知道，有点乐观，觉得这类东西是能拿到的。我们开始更多地转向

102:08

into a SPR biophysics related measurements try to figure out if we can maybe come up with like a

SPR biophysics 相关的测量，想搞清楚我们能不能也许搞出类似一个

102:14

like a potency ladder that we could then be as part of the chemical material that we're screening

比如一个 potency ladder，然后我们就能把它作为正在 screening 的 chemical material 的一部分

102:19

uh to try to capture those things but I won't say that like we're good at it yet like you can't

呃，是想试着捕捉这些东西，但我不会说，好像我们已经很擅长了，就像你没法

102:23

use light you can't just use reads as a proxy for KD or something um I think the relative ranking yes

用光，你不能只拿 reads 当 KD 的 proxy 之类的，嗯，我觉得 relative ranking，是的

102:28

but I think we can read right rates reads for KD not really and then you start running it a

但我觉得我们可以读，对吧，rates、reads 来算 KD，其实并不行，然后你就开始跑它，一个

102:32

weird stuff so like there's lots of proteins that we've encountered where it turns out love like

奇怪的东西，所以就像，有很多我们遇到过的 proteins，结果发现，呃，就像

102:36

the DNA barcode itself there there are proteins that are not described as DNA binders that we have

DNA barcode 本身，呃，有些 proteins 并没有被描述成 DNA binders，而我们

102:41

discovered are and so you can imagine that that might affect the KDs of things and you have to

发现它们其实是，所以你可以想象，那可能会影响这些东西的 KDs，而你必须

102:45

figure out and mitigate those problems and that kind of stuff so like yeah it's I think the relative

弄清楚并缓解这些问题，诸如此类，所以就像，是啊，我觉得 relative

102:49

ranking is something that's quite doable I think translating it into KDs is something that we're

ranking 是相当可行的，我觉得把它转化成 KDs 是我们

102:52

actively working on and we think we might really get like some reasonable you know kind of like

正在积极地做，而且我们觉得我们可能真的能达到某种还说得过去的，你知道，有点像

102:55

sloppy translation but I wouldn't say that we're there yet okay this one this one was pretty brutal

草率的翻译，但我不会说我们已经到那一步了。好吧，这个，这个真的挺残酷的

103:01

oh the bitter the bitter lesson that yeah I'm just gonna read this this one is pure clot I had no

哦，the bitter，the bitter lesson，那个，对，我就直接读这个吧，这个纯粹是 clot，我完全没

103:06

thought in this okay the bitter lesson is really the law what stops a better funded lab from

想过这个。好吧，the bitter lesson 真的就是那个法则：什么能阻止一个资金更充足的实验室

103:10

outscaling your Dell engine oh yeah of course people can ask us this um so and the reason like

outscaling 你的 Dell engine。哦对，当然人们会问我们这个，嗯，然后原因就像

103:15

you know huge what I would really say is that uh both at least end of recursion uh uh my team has

你知道，很重大，我真正想说的是，呃，both at least end of recursion，呃呃，我的团队已经

103:20

made so many mistakes in accidentally bumping into the cheating modes that just kind of come out

犯了太多错误，不小心撞进了那些 cheating modes，它们就有点像是从

103:28

of the consequences of doing laboratory work there are so many of these and and so like you have to

做实验室工作的后果里冒出来。这种东西太多了，而且，而且所以就像，你必须

103:33

chase these down and a lot of it has to do with tracking batch effects with metadata for example

把这些追查到底，而其中很大一部分都跟用 metadata 追踪 batch effects 有关，比如说

103:38

and so then you have to build a culture of your organization where it's like all right we have to

所以接下来，你必须在你的组织里建立一种文化，就是那种，好吧，我们必须

103:42

track all these things we've got to build tools track all these things we have to make those tools

追踪所有这些东西，我们得构建工具来追踪所有这些东西，我们还得让这些工具

103:45

frictionless that the people doing the work actually do will record these things uh and then you

用起来毫无阻力，让真正在做这些工作的人确实会把这些东西记录下来，呃，然后你

103:49

have to go back and look at all the data back into this founding history of the company to make

还得回头去看所有数据，一直回溯到公司创立之初的历史，来确保

103:54

sure you're getting really consistent results and then when you see a big changes you have to identify

你得到的结果真的非常一致，然后当你看到大的变化时，你必须找出

103:57

what drove them then you have to mitigate those and it's just like it's like a huge huge amount of

是什么驱动了它们，然后你必须去缓解这些问题，而这就像，就像，是巨大巨大量的

104:01

work and we also rebuilt kind of from scratch the selection process we're rebuilt from scratch a lot

工作，而且我们还差不多从零重建了筛选流程，我们从零重建了很多东西

104:06

of the protein expression process and then we have all these other proxies that we use to evaluate

protein expression 过程的一部分，然后我们还有所有这些我们用来评估的其他 proxies

104:10

how well the protein production went and if it's good or it's bad and these other filters and it's

protein production 进行得怎么样，是好还是坏，还有这些其他的 filters，然后它

104:14

just um it takes kind of a unique culture and patience in going to make these mistakes and then

只是，嗯，这需要一种独特的文化和耐心，去犯这些错误，然后

104:21

had the team like really think very deeply about how to address them and I think unless you've got

让团队就像真的非常深入地去思考如何解决它们，而且我觉得除非你有

104:27

folks who you know really have a nose for these kind of experimental mistakes it's going to be pretty

你知道的，真正对这些实验错误有敏锐嗅觉的人，否则这将会挺

104:31

it's going to take some time to replicate the process mode yeah I really think so there's just so

要花些时间才能复现 process mode，对，我真的这么认为，只是有太

104:36

many things where it can go sideways and you have to be paying enough attention all the time to

多地方会出岔子，你必须一直足够留意，才能

104:41

identify them when they're going sideways and there might be things that you can write automated

在它们出岔子的时候识别出来，而且可能有些东西你可以写成 automated 的

104:44

filters for and maybe not early on um and so like you know like one pretty good example I'm happy to

filters for，而且也许不是早期，嗯，所以，就像，你知道，就像，有一个挺不错的例子，我很愿意

104:51

talk about is um people ask us a lot how do we know our proteins are really good we have lots of ways

聊的是，嗯，人们经常问我们，我们怎么知道我们的 proteins 真的很好，我们有很多办法

104:56

to figure this out I think one of my favorite ones right now is um we have noticed that when proteins

来搞清楚这一点。我觉得目前我最喜欢的一个方法是，嗯，我们注意到，当 proteins

105:01

misfold and we're able to identify misfolding in very crude cases like if you put the protein you know

misfold 的时候，我们能识别出 misfolding，在一些非常粗略的情况下，比如如果你把 protein，你知道

105:07

run out of gel which is this protein trick where you try to figure out how big it is um and it

跑 gel，这是那种 protein 的小技巧，你想弄清楚它有多大，嗯，然后它

105:11

misfolds then you can kind of detect that but if a misfold is a little bit it could be hard

misfold 了，你就能大概检测到，但如果 misfold 只是一点点，可能就很难

105:15

there are drugs whose job it is to refold misfolded proteins correct like O-shaperones okay

有些药物的职责就是 refold 那些 misfolded 的 proteins，对吧，比如 O-shaperones，好吧

105:22

you can call creatures too that's what you would call them yeah uh there's a very famous one that

你也可以叫它们 creatures，你就是会这么叫它们，对，呃，有一个非常著名的，那个

105:26

treats uh cystic fibrosis patients called lumacaftor we have a bunch of chemicals in our chemical

治疗呃 cystic fibrosis 病人的 lumacaftor，我们有一堆 chemicals 在我们的 chemical

105:31

library that looks like lumacaftor this thing whose job it is to refold misfolded proteins guess

library 里，看起来像 lumacaftor，这东西的工作是重新折叠 misfolded proteins，你猜

105:36

what kind of proteins it hits when we screen them sick to sick fibrosis uh

它会命中什么样的 proteins，当我们 screen 它们 sick to sick fibrosis 呃

105:41

misfolded proteins oh yes folded proteins that's what it's for it's like and it's in our libraries

misfolded proteins，哦是的，folded proteins，它就是干这个的，就像，而且它就在我们的 libraries 里

105:48

and so you know like and we discovered this by accident like we saw this you know these

所以你知道，就，我们是偶然发现这个的，就像我们看到了这个，你知道，这些

105:52

chemical series keep popping up across all these different protein types like man like what is

chemical series 不断出现在所有这些不同的 protein 类型中，就像天哪，就像什么

105:55

that lumacaftor and then you start looking at the protein expression and then like how it went

那个 lumacaftor，然后你开始看 protein expression，然后就像它怎么变化的

106:00

it's like oh it goes when when things aren't going well and so now we can use it as one filter

就像，哦，它会在事情不太顺利的时候变化，所以现在我们可以把它当作一个 filter 来用

106:06

for if we think if if we get a bunch of lumacaftor ads that protein probably didn't fall

因为如果我们想，如果，如果我们拿到一堆 lumacaftor ads，那个 protein 可能没 fold

106:11

fold right we probably don't want to use that as training data we probably want to rewrite it with

对，我们可能不想把它当作 training data，我们可能想把它改写成

106:15

slightly different conditions and make it more friendly to the protein then we'll see what those

稍微不同的条件，让它对 protein 更友好，然后我们看看那些

106:18

hits look like and so like unless you're paying that kind of attention like it's really I mean

hits 长什么样，然后就像，除非你付出那种注意力，真的，我是说

106:23

it's really hard yeah and you have to build tools so that you can just look at these things and

真的很难，对，而且你得做工具，这样你就能直接看这些东西，然后

106:28

have people who know to look at these things and recognize them it's uh yeah it's it's a hard problem

让懂的人来看这些东西并认出它们，这，呃，对，这，这真的是个难题

106:32

yeah going back to uh everyone who touches delts comes out traumatized somehow you've uh you

对，回到，呃，每个碰过 delts 的人不知怎么最后都受到创伤，你，呃，你

106:38

figured it out you're also traumatized oh that good way yeah oh it's really hard and you know

搞明白了，你也受到创伤，哦，那也挺好，对，哦，真的很难，你知道

106:45

and that's why it's and and you know when you had the motivation to just finding like one or two

这就是为什么，而且你知道当你有了动力，只想找到一两个真正厉害的好东西来启动一个药物项目时，你就不会去追查assay的所有这些不同方面，所以我理解，但如果你有动力让它成为machine learning的素材，那就得关注所有这些东西。我们最后总是会问的一个问题是，如果你能凭一道命令消除一个瓶颈，那会是什么？不，当我想到凭一道命令的时候，这其实是目前正在发生的事情之一，就是看着sequencing成本像石头一样往下掉，我在想，好，如果我们想要更多chemicals，想要更多measurements，我们真正想做的是什么？

106:49

really awesome hits to start off a drug program like then you're not gonna chase down like all

一开始就有特别棒的 hits 来启动一个 drug program，那你就不会去追着深挖所有

106:54

of these different aspects of the assay and so I get it but like if you're motivated for it to be

assay 的这些不同方面，所以我懂，但就是，如果你是为了让它成为

106:58

machine learning fodder then like yeah you have to pay attention to all this stuff one of the

machine learning 的素材，那，是啊，你就得注意所有这些事情。我们

107:01

questions that we always ask at the end is if you could by fiat remove a bottleneck

最后总是会问的其中一个问题是：如果你能一声令下就移除一个 bottleneck

107:10

what would that be no when I think about by fear and it's one of these things that is actually

那会是什么呢，不，当我出于恐惧去想的时候，而它其实是那些实际上

107:15

currently happening which is watching the sequencing costs drop like a rock like I'm thinking

正在发生的事情之一，就是眼看着 sequencing costs 一落千丈，就像我在想着

107:21

about okay what do we really want to do if we want more chemicals and we want more measurements

好吧，如果我们想要更多化学品、想要更多测量，我们真正想做的是什么

107:25

those are the things that I want to make my life a little bit easier like at the problems that

那些就是我想让我的生活稍微轻松一点的事情，比如那些

107:29

are immediately on the ground so you want sorry I don't understand cheaper sequencing more molecules

就摆在眼前的问题，所以你想要——抱歉，我没听懂，更便宜的 sequencing、更多 molecules？

107:34

okay yes and sequencing like can you give some context for how much it's dropped

好，是的，还有 sequencing，比如你能给点背景，说说它降了多少吗？

107:39

oh I got like many many many orders of magnitude it is uh I think the first genome cost a

哦，我印象里好像是很多很多很多个 orders of magnitude，呃，我想第一个 genome 花了

107:45

billion dollars and now we can do a human genome for a thousand three I think it might be a hundred

十亿美元，而现在我们做一个 human genome 只要一千三，我觉得可能是一百

107:49

a hundred yeah a hundred sorry yeah yeah I there are stretches where was dropping faster than

一百，对，一百，抱歉，对对，我……有些阶段它的下降速度快于

107:53

more law yeah was sorry yes and what so wait what why then why do you want to even cheaper sequencing

Moore's law，对，是，抱歉，是的，然后什么？等等，什么？为什么？那你为什么还想要更便宜的 sequencing？

107:59

oh I just I'm just happy with the trend okay yeah what would you remove it by fiat if you like

哦，我只是……我只是对这个趋势很满意，好吧，对，如果你愿意的话，你会用强制手段把它取消掉吗？

108:04

that hasn't been removed yeah here's a good thing what I would want is there are different gizmos

那个还没被移除，是的，有件好事，我想要的是有各种不同的小玩意儿

108:10

that you can use to kind of purify small molecules okay so when you uh right now when you make them

你可以用来纯化一下 small molecules 的，好吧，所以当你，呃，现在你制造它们的时候

108:16

you you put some stuff on a pot and then you heat it up or do some other conditions and you get

你你往一个罐子里放些东西，然后加热，或者做其他一些条件处理，然后你得到

108:20

that output and and that will work at some level of efficiency might be great might be crappy

那个产物，而且，而且那会在某个效率水平上起作用，可能很棒，可能很烂

108:26

and so in either case especially if it's crappy there are these other gizmos that you can use to

所以无论哪种情况，尤其是如果它很烂，还有这些其他的小玩意儿你可以用来

108:30

kind of clean it up so people use like hplc or some of these other tricks none of them are super

把它清理一下，所以人们会用像 HPLC 或者一些其他的技巧，它们都不是超级

108:34

paralyzable that are really really good and so being able to like clean up stuff at a massive scale

parallelizable，真的真的很好，所以能够像大规模地清理东西

108:40

will be really really nice because like we you know well what do we want to do we run all these

会真的真的很好，因为就像我们，你知道，我们想做什么呢，我们跑所有这些

108:45

molecules in this funky assay that we've already described as given peaceful lots of trauma

在这个我们已经描述过的古怪 assay 里，molecules 被给了平静的大量创伤

108:48

it will be great to then get you know large numbers of molecules in a different assay that

然后能在另一个 assay 里，你知道，弄到大量 molecules，那就太好了，那个 assay

108:54

be really clean and purified and easy to work with uh and then go do that next because then you

会非常干净、purified、好操作，呃，然后接下来就去做那个，因为那样你

108:59

could really orthogonally validate all those measurements that you're making so yeah that'd be really

就能真正 orthogonally validate 你正在做的所有那些 measurements，所以对，那会真的

109:03

really nice and of course everyone complains about like GPUs and stuff like yeah GPUs that's the

非常棒，当然大家都会抱怨像 GPUs 之类的东西，对，GPUs，那就是

109:09

yeah yeah yeah yeah yeah yeah yeah yeah yeah yeah yeah yeah yeah yeah yeah yeah yeah yeah yeah yeah yeah

对对对对对对对对对对对对对对对对对对对对对

109:14

all right so and then the other question is um what would you like people to take away from that

好，那么，然后另一个问题是，嗯，你希望人们从

109:18

episode I would like people to take away from the episode that like you know at least for us we

那一集里带走什么？我希望人们从那一集里带走的是，你知道，至少对我们来说，我们

109:23

we believe that we're building soon and that's really really exciting and you know like what feels

我们相信我们很快就能把它做出来，这真的真的超级令人兴奋，而且你知道，就像是一种感觉上 physically grounded 的 substrate，能做各种有意思的 reasoning 和 biomedical 应用。

109:28

like physically grounded substrate for all sorts of interesting reasoning and biomedical

所以如果他们对这个感兴趣，请一定联系我们，你知道，我们在招 machine learning 的人，我们想和合作伙伴聊聊。

109:32

applications and so if that's of interest to them like please reach out to us like we're you know

所以你知道，我们正在做这个东西，我们觉得它会有变革性，所以联系我。

109:36

we're hiring machine learning people we want to talk to partners and so you know we're building

太棒了，酷，那太好了，非常感谢，谢谢你来，谢谢，嗯嗯。

109:40

this thing that we think will be transformative so contact me amazing cool that was great so much

这个我们认为会有变革性影响的东西，所以联系我。太棒了，太酷了，那真是太好了，非常

109:45

thank you for coming thank you yeah yeah

感谢你来，谢谢，耶耶

Lab-in-the-Loop at Compute Speed — Ian Quigley, Leash Bio

Play Queue