Controlling Biology - Fabian Theis, Helmholtz Munich ·

00:00

We can really now start guiding the experimentation.

我们现在真的可以开始引导实验了。

00:02

So, where before you, we build these human-cell

所以，以前呢，我们构建这些 human-cell athleticities 的时候，我们多少有这种 mapping 的层面。

00:04

athleticities, we had sort of this mapping aspect.

然后我们谈到了 world model 的 prediction 问题。

00:06

Then we talked about the world model prediction problem.

从 mapping，我们走到了 prediction。

00:08

From mapping, we went to prediction.

但现在，我们可以实际去引导下一个实验了。

00:10

But now, we can go to actually steering the next experiment.

但说实话，很多时候论文最后都停在某个 model 上，对吧？

00:13

But quite frankly, very often the papers

但说实话，很多时候这些论文

00:15

end with some model, right?

最终都会落到某个模型上，对吧？

00:16

Maybe some sort of concept figure.

也许是某种 concept figure。

00:18

Maybe at a little bit validation.

也许做一点点 validation。

00:19

But entropy, you start a next project, right?

但 entropy，你会开始下一个项目，对吧？

00:20

And of course, you learn a bit, but it's not like

当然，你会学到一点，但它不像是

00:22

a continual learning type of thing.

那种 continual learning 的东西。

00:24

A new thing, which we call lab compass,

一个新东西，我们叫它 lab compass，

00:26

which is an inverse generative model that actually allows you,

它是一个 inverse generative model，实际上能让你，

00:28

once you have a forward world model, whatever,

一旦你有了一个 forward world model 之类的，

00:30

to tell you what you need to do in order to reach a certain state.

来告诉你为了达到某个状态需要做什么。

00:32

Obviously, there are the ultimate vision

显然，最终愿景是

00:34

being that we want to RL everything.

我们想把所有东西都用 RL 来做。

00:35

We want to do RLs in this loop.

我们想在这个 loop 里做 RL。

00:37

It's a bit out there, but I think we're slowly getting that.

这有点超前，但我觉得我们正在慢慢实现它。

00:41

Welcome to the latent space AI for science podcast.

欢迎收听 latent space AI for science podcast。

00:44

I'm RJ Hanuki, CTO of mereomics.

我是 RJ Hanuki，mereomics 的 CTO。

00:47

I'm joined by my co-host, Brandon Anderson,

和我一起的是我的共同主持人 Brandon Anderson，

00:50

who develops RNA therapeutics at atomic AI.

谁在 atomic AI 开发 RNA therapeutics？

00:53

I can't emphasize enough how much Fabian Tess

我怎么强调都不为过，Fabian Tess

00:57

and his labs and companies have contributed

以及他的实验室和公司

00:59

to the development of systems biology

对 systems biology 的发展

01:02

and various iterations of the virtual cell concept.

以及 virtual cell concept 的各种迭代做出了多大的贡献。

01:05

Papers and models come out of his lab faster than I can read them.

他实验室产出的论文和 models 快得我都读不过来。

01:09

And the lab is consistently moving the ball forward

而且这个实验室一直在把这件事向前推进

01:11

on our understanding of how to observe and model

推动我们对如何观察和建模的理解

01:14

the behaviors and interactions of cells.

细胞的行为和相互作用。

01:17

Just as importantly, his work often focuses on creating

同样重要的是，他的工作常常聚焦于打造研究人员和科学家在日常工作中真正会用到的工具。

01:20

the tools, the researchers, and scientists actually use

大多数计算生物学家可能都在某个时候用过 Kampai 和 NDATA。

01:24

in their day-to-day work.

而这项工作促成了 SCverse 的诞生，它是一套非常棒的开放工具集合。

01:25

Most computational biologists have probably used

大多数计算生物学家可能都用过

01:28

Kampai and NDATA at some point.

Kampai和NDATA。

01:30

And this work has led to the creation of SCverse,

这项工作促成了SCverse的诞生，

01:33

which is an amazing collection of open tools

它是一个很棒的开放工具集合，

01:36

that has attracted many, many contributors.

这吸引了非常非常多的贡献者。

01:38

Of course, Fabian is one of those academics

当然，Fabian 就是那种学者

01:41

who has way more honors than you can possibly list.

他的荣誉多到你怎么列都列不完。

01:44

But suffice it to say that he leads a bunch of the AI

但简单来说，他领导着许多 AI

01:47

and computational biology efforts at Helmholtz Munich

以及 Helmholtz Munich 的 computational biology 方面的工作，

01:50

and related institutions and participates also

还有相关机构的工作，同时也参与

01:53

in many other industry efforts.

很多其他行业项目。

01:56

Fabian, it's really great to meet you.

Fabian，真的很高兴见到你。

01:58

Thanks a lot for joining us.

非常感谢你来参加我们的节目。

02:00

Thanks for having me.

谢谢邀请我。

02:01

I'm really looking forward to the discussion then.

那我很期待接下来的讨论。

02:02

Yeah, I've been tracking your work for a long time.

是啊，我关注你的工作很久了。

02:05

And so it's a great honor to finally get you on the show

所以终于能邀请你上节目，

02:09

and talk to you.

和你聊聊，真是莫大的荣幸。

02:10

Let's kick things off with some of the hype happening right now.

我们先从眼下的一些热潮聊起吧。

02:12

There's a ton of interest right now in structural biology

现在大家对 structural biology 的兴趣特别大。

02:16

andthropic and open AI bioforma initiatives

andthropic 和 open AI 的 bioforma initiatives

02:19

focus heavily on designing molecules right now.

现在特别聚焦在设计分子上。

02:22

On the other hand, virtual cell and systems biology approaches

另一方面，virtual cell 和 systems biology 这些方法

02:25

really haven't gotten as much attention recently.

最近确实没怎么受到关注。

02:29

Maybe Zera and Noetic are good counter examples there.

也许 Zera 和 Noetic 算是这方面不错的反例。

02:32

But bio hub just announced a $1.8 billion collaboration

但 bio hub 刚宣布了一项 18 亿美元的合作

02:36

with the DOE and NIH, Meda and Google

合作方是 DOE、NIH、Meda 和 Google

02:39

to collect and open source multi-scale biological data

来收集并 open source multi-scale biological data

02:42

that can be used to develop a new generation of virtual cells.

这可以用来开发新一代的 virtual cells。

02:46

So my question, do you think that we're about to see renewed

所以我的问题是，你觉得我们是不是即将看到

02:50

interest from the farm industry in virtual cells

农业产业对 virtual cells 重新燃起兴趣

02:53

based on the activity that we're seeing in academia right now?

考虑到我们目前在学术界看到的活跃情况？

02:56

I should disclose before I sort of go to the answer

在我大概要给出答案之前，我应该先说明一下

02:58

that obviously my opinion is quite academic

那就是，显然我的观点相当学术

03:02

in the sense that whenever I talk to my friends and colleagues

意思是，每当我跟朋友和同事聊天时

03:05

that really sort of day by day sit in the startups,

他们真的是那种每天泡在 startups 里的人，

03:09

they have like what I sort of have as an overview

他们有的东西，有点像我所拥有的那种总体概览

03:11

and sort of a bit of intuition what the different research labs do.

以及大概对不同研究实验室在做什么的一点直觉

03:14

They have the same thing for the startups.

他们对 startups 也有同样的东西

03:15

And this is sort of the hot big thing

而这差不多就是那个热门的大事

03:18

that you know, this and this investments are happening.

也就是，你知道的，这个、这个投资正在发生

03:19

I don't really have that on that scale.

我其实没有到那个规模

03:21

But having said that, for example, you mentioned a bio hub initiative

但话虽如此，比如你提到了一个 bio hub initiative

03:25

which by the way is also heavily coupled to human cell Atlas

顺便说一句，它也跟 human cell Atlas 紧密耦合

03:28

which is sort of this internationally really worldwide consortium

这算是一个国际性的、真正覆盖全球的联盟

03:30

that sort of pushes this next level of data set generation.

它在推动 data set generation 达到下一个层级。

03:34

I do think that there's a whole bunch of reasons

我确实认为，有一大堆原因

03:38

why we're going to see that sort of scale up

会让我们看到那种 scale up

03:41

from all these models working well in terms of being generative

从所有这些模型在 generative 方面表现良好这一点出发

03:44

but also sort of lapping the loop on the molecule level

同时也在 molecule level 上把这个循环走完

03:47

now being exported and sort of really working one level up

现在被向外输出，并且真的往上升了一个层级

03:50

on the cell level.

也就是在 cell level 上。

03:51

And I think there's a bunch of reasons for that.

而且我觉得这背后有很多原因。

03:54

And the first one I think and maybe the most important one

而第一个，我觉得，也许也是最重要的一个

03:55

which you really hear is the data generation.

你确实经常听到的，就是 data generation。

03:58

I think we had that concept of a virtual self for a while

我觉得我们已经有 virtual self 这个概念有一段时间了

04:02

like my lab we started this perturbation prediction modeling

比如我的实验室，我们开始做这个 perturbation prediction modeling

04:06

and using machine learning in 2019

并且在 2019 年用上了 machine learning

04:08

with a paper we called SEGEN at the time

当时有一篇我们叫 SEGEN 的论文

04:10

which was like, it was a non-linear model

它大概就是，一个 non-linear model

04:12

but essentially sort of in a latent space

但本质上，差不多是在一个 latent space 里吧

04:14

did some linear transformation.

就是做了一些 linear transformation。

04:15

And it was actually quite okay and even nowadays

而且其实效果还挺不错的，甚至到现在

04:18

many of these sort of pseudo linear type of things

很多这种 pseudo linear 类型的东西

04:20

are still doing good job

依然做得挺好

04:21

but the reason for that being that I think the data says

但原因在于，我觉得 data 所说明的是

04:25

and the questions that we asked on top

以及我们在此基础上问的那些问题

04:27

were just sort of still on this simplistic side

也都还是有点偏简单化那一侧

04:29

that's why also sort of it didn't yet take up.

这也是为什么它某种程度上还没真正兴起。

04:31

So whenever people talk about virtual cells

所以每当人们谈到 virtual cells

04:33

maybe also your question I think what you actually ask is

也许还有你的问题，我觉得你实际想问的是

04:37

how does can you predict how a cell would behave

你怎么能预测一个细胞会如何表现

04:40

on the all kind of perturbations.

在各种 perturbations 下。

04:41

And that's sort of I think a quite deep question

而且我觉得这某种程度上是个挺深的问题

04:43

and if you really think about it

如果你真的仔细想想

04:44

that's maybe even a bit of a holy grail

这甚至可能有点像个圣杯

04:46

and computation barge of systems biology

以及 systems biology 的 computation barge

04:48

sort of this input output system that is a cell

就像是这个 input output system，也就是一个 cell

04:52

with memory and like with all these specifications

有 memory，还有所有这些 specifications

04:54

and with all the different sort of levels of molecules

以及各种不同层面的 molecules

04:57

and so on, I'm not just our name.

等等，我不只是我们的名字。

04:58

I think being able to map this out

我觉得能把这一切梳理出来

05:01

what does even understanding mean

到底什么叫理解

05:02

and I've been the sort of bit of a meta problem

而且我有点像是个 meta problem

05:04

but I think you understand at least your modular system

但我认为，至少你对自己的 modular system

05:08

very well when you sort of predict its output behavior

在某种程度上预测它的 output behavior 时，理解得还是很透彻的

05:10

given different inputs.

也就是给定不同输入的时候。

05:12

So I think it's a very good question

所以我觉得这是个很好的问题

05:13

and this is why often people nowadays equate

这也是为什么现在人们经常会把

05:16

a virtual cell with perturbation prediction.

virtual cell 和 perturbation prediction 等同起来。

05:20

I think it's a bit too simplistic

我觉得这有点太简单化了

05:21

but I think this perturbation prediction type of models

但我觉得这类 perturbation prediction 模型

05:23

and that's sort of maybe what you mean by virtual cell

这大概就是你说的 virtual cell 的意思吧

05:25

is taking off a lot and the startup landscape

现在发展得特别快，而 startup 格局

05:28

I think is a bit exploding there.

我觉得在那边有点爆发式增长。

05:29

We talk a lot about US Europe

我们经常谈到 US、Europe

05:32

but also in the Chinese ecosystem

但在中国生态里也是

05:34

there's a lot of these additional efforts now happening

现在有很多这类额外的努力正在发生

05:36

and I hope that these are not just sort of now

我也希望这些不只是现在

05:39

kindling off of a flame

借着一团火苗点起来的

05:40

because everyone sort of adds AI to their things

因为大家多多少少都在把 AI 加到自己的东西里

05:42

but because farmer is really desperately

但因为 farmer 真的非常迫切需要做到 cell type context specific

05:45

needing to be cell type context specific

以及他们的 screens——目前他们还没法在那种规模上做到

05:48

and their screens which at the moment they can't do yet

也许你可以稍微展开讲讲这到底指什么

05:50

on that scale.

所以现在的 perturbations

05:51

Maybe you could expand a little bit on what does this

比如我们最近和 Zera 聊过

05:55

so perturbations right now

所以现在做扰动

05:57

like we talked to Zera recently

就像我们最近跟Zera聊的那样

05:59

they're doing these perturbation screens

他们在做这些 perturbation screens

06:01

what does that unlock

这能解锁什么？

06:03

and then what do you think the next generation of data

然后你觉得下一代 data

06:06

is going to unlock in addition to that?

除此之外还会解锁什么？

06:08

I mean for me

我是说，对我来说

06:09

if I may stop when I'm sort of getting to put too much

如果我可以停一下——当我有点开始把太多东西

06:12

into these old stories

塞进这些老故事里

06:14

like you know I'm sort of I've been trained

就像你知道的，我有点像是，我一直是被训练成这样的

06:17

machine learning scientists for the longest time

machine learning 科学家们长期以来

06:19

doing all kinds of things

一直在做各种各样的事情

06:20

but now I started with single cell early on QPCR

但现在，我早期是从 single cell、QPCR 开始的

06:24

maybe initially if I'm really honest

也许一开始，如果我真说实话

06:26

not so much because it was like the biology

倒不是因为那个 biology

06:28

that was interesting but because we had high sample numbers

本身有多有趣，而是因为我们的样本数量很多

06:30

purely for that right

纯粹就是为了这个，对吧

06:31

because even early on it was sort of clear

因为即便在早期，这多少也已经很清楚了

06:34

that the number of QPCRs at scale

也就是规模化的 QPCRs 数量

06:37

of course transcriptomes was just getting really big

当然，transcriptomes 的规模当时正变得特别大

06:40

and you know when I started my career early on

你知道，我早年刚入行的时候

06:42

we had like a few bug RNA

我们大概只有几个 bug RNA

06:44

now we sort of had thousands of them or whatever

现在呢，我们差不多有成千上万个了，差不多吧

06:46

so this is like early times

所以这还像是早期的时候

06:47

but now you know we'll be going to to a regime

但现在你知道，我们要进入一种 regime 了

06:50

and that was sort of one of these eye opening moments

那算是那种让人眼前一亮的时刻之一

06:52

when parts by I would have

当 parts by 我本来会有的时候

06:53

I think they call it the gigalab or something like that

我觉得他们叫它 gigalab 还是什么的

06:55

and you know there's sort of similar ones

你知道现在也有点类似的东西

06:57

from competing companies now

来自现在的竞争公司

06:58

but when they had like in a single experiment

但他们有一次在单个实验里

07:00

10 million transcriptomes measured

测了 1000 万个 transcriptomes

07:02

right

对吧

07:03

Father was like next level stuff

Father 那简直就是另一个级别的东西

07:05

you know this step from just getting more PBMCs

你知道这一步，从只是获取更多 PBMCs

07:08

as a control tool like let's actually do something

作为一个 control tool，就像我们真的要做点事

07:11

at scale to it sort of having here the cell space

在大规模上，对它来说，差不多就是这里有了 cell space

07:13

but then having here the perturbation space

但接着这里有了 perturbation space

07:15

which typically was very scarce

而这一般来说非常稀缺

07:17

it was like control versus disease or so

它就像 control 对 disease 之类的

07:20

so this sort of number of points

所以这种 points 的数量

07:22

what you had here in terms of sampling

你这里在 sampling 方面所拥有的

07:23

on the perturbation space was typically much

在 perturbation space 上通常要

07:25

much more scarce than the number of observations

比 observations 的数量稀缺得多得多

07:27

in terms of cells that you had

就你手头的 cells 来说

07:28

but now this is exploding

但现在这块正在爆炸式增长

07:29

now you sort of have I know 100 different cell lines

现在你大概会有，我知道，100 种不同的 cell lines

07:31

where you do sort of full perturbacy

你可以做那种 full perturbacy

07:33

like full crisp out

比如 full crisp out

07:34

or you have like

或者你有那种

07:36

batteries of

一批批的

07:37

thousand drugs or so added

大概上千种药物被加进来

07:39

and suddenly this space gets really rich

然后这个领域突然就变得非常丰富

07:40

right and then you can sort of start doing

对吧，然后你就能开始做

07:42

all these fancy perturbation prediction type of things

各种花里胡哨的 perturbation prediction 之类的东西

07:44

and this is currently happening

而这是目前正在发生的

07:45

it's crazy what what scale is happening

现在这个 scale 的发展有多疯狂，简直难以想象

07:47

we're ourselves involved

我们自己也在参与其中

07:49

in a whole bunch of things

在一大堆事情上

07:50

where we actually go beyond cell lines

我们实际上会超越 cell lines

07:52

and do

然后做

07:53

all the noise perturbations

所有的 noise perturbations

07:54

and maybe if if time we can talk a bit more about that

也许如果有时间的话，我们可以再多聊一点这个

07:56

because I'm quite convinced that we need to go to multi-cell

因为我相当确信，我们需要走向 multi-cell

07:58

the systems to have actually really interesting readouts

的系统，才能真正得到有意思的 readouts

08:01

but I think everyone is scaling

但我觉得大家都在 scaling

08:02

these levels up and at some point you know

这些 levels 不断往上，然后到某个时候，你知道

08:04

we're going to go maybe even through human genetics

我们可能甚至会一路走到 human genetics

08:06

as a perturbation

作为一种 perturbation

08:07

can you just like explain what is a perturbation

你能不能就解释一下，什么是 perturbation

08:09

what do you mean by that just for the people

你这么说是什么意思，就当是给那些

08:11

who are not familiar

不太熟悉的人讲讲

08:13

yeah yeah actually would you mind

对对，其实你介不介意

08:15

maybe backing up even one step more than that

也许再往回倒一步，甚至比那再多退一步

08:17

like what do we even mean by like measuring

比如说，我们说的 measuring 到底是什么意思啊

08:20

transcriptomics

08:21

and you were talking about like PCR-based

然后你提到的像是 PCR-based

08:23

and then also perturbacy

然后还有 perturbacy

08:25

when I keep sort of public talks

我经常做那种公开演讲的时候

08:26

that I typically say you know human genome project

我通常会说，你知道，human genome project

08:29

we sort of have the genome for the body

我们算是有了人体的 genome

08:31

but it's pretty much the same everywhere right

但它在哪儿基本上都一样，对吧

08:32

so like why do we have all these different cell types

所以，就像，为什么会有这么多不同的 cell types 呢

08:35

right what sort of happening specifically

对吧，具体到底是在发生什么

08:36

and so first level what we typically have

然后第一层呢，通常我们会看到的是

08:39

one of the first levels is that you know

最早的一层之一就是，你知道

08:40

per cell different stuff is being read off

每个 cell 里被读取的东西都不一样

08:42

of the genome right

从 genome 上读，对吧

08:43

and these messenger RNAs

然后这些 messenger RNAs

08:44

which then sort of encode

它们随后有点像是在 encode

08:46

in the various downstream proteins

在各种 downstream proteins 里

08:48

where these are sort of differently expressed

它们在这些地方算是表达得不太一样

08:50

as one says in different cells

就像人们说的，在不同细胞里

08:52

and to say in a human maybe there's just something

而且这么说吧，在人类身上可能也就有那么

08:54

like 20,000 different types of these

大概 20,000 种不同的这类东西

08:56

we're splicing many many more

我们通过 splicing 弄出来的要多得多得多

08:58

but there's sort of a rough level

但大致上有一个大概的量级

08:59

so what you de facto have

所以你实际上得到的是

09:01

is per cell a 20,000-dimensional vector

每个细胞都是一个 20,000-dimensional vector

09:04

that has a whole bunch of zeroes and stuff

里面有一大堆零啊之类的东西

09:06

is not there but then a whole bunch of sort of numbers

它不在那儿，但接着又是一大堆那种数字

09:07

how many of these counts of these

这些东西的 counts 到底有多少

09:10

messenger RNAs you have

你有多少 messenger RNAs

09:11

because these things are called transcripts right

因为这些东西叫 transcripts，对吧

09:13

this is called the transcriptome

这叫做 transcriptome

09:16

yeah and and de facto then you have this sort of 20,000-dimensional space

是啊，然后实际上你就有了这么一个 20,000-dimensional space

09:19

that you very often encode in a sort of nicer latent space

你经常会把它们编码到一个更好用的 latent space 里

09:22

that sort of has like a sub manifold

它差不多有一个 sub manifold

09:24

like sort of this the sub space where these actual cell cloud lips

就像这个 sub space，实际的 cell cloud 所在的地方

09:28

and if the cell cloud then sort of like this

然后这个 cell cloud 如果差不多像这样

09:30

sort of 20,000-dimensional vectors

差不多是 20,000-dimensional vectors

09:31

each point then being one cell right

每个点就代表一个 cell，对吧

09:33

if they sort of cluster up then we call these things cell types

如果它们差不多 cluster 起来，我们就把这些东西叫 cell types

09:36

or something like that

或者类似这样的东西

09:37

and if they sort of go into a sausage or whatever

然后如果它们有点像变成一根香肠还是什么的

09:40

we call them interjectories

我们叫它们 interjectories

09:41

like a cell differentiation

就像 cell differentiation 一样

09:42

like the fertilized Excel

就像受精的 Excel 一样

09:44

then making all the cell taps in the body

然后形成身体里所有的 cell taps

09:45

or something like

或者类似这样的

09:46

so like DNA is sort of like a recipe book

所以就像 DNA 有点像一本食谱

09:49

or maybe it's a bunch of genes

或者也许它是一堆 genes

09:51

or a bunch of ingredients

或者一堆食材

09:53

and then as you each cell is its own sort of

然后每个细胞本身就像是一种

09:56

meal or dish or something which is a combination of different ingredients

一顿饭或一道菜之类的东西，是不同食材的组合

10:00

and now the transcriptomics is sort of figuring out

而现在 transcriptomics 有点像在弄清楚

10:04

taking different meals and then sort of decomposing them

把不同的饭菜拿来，然后有点像把它们拆解开来

10:07

like oh what ingredients went in to make this meal

就像，哦，做这顿饭用了哪些食材

10:09

so it used to set a dumb and healthy

所以它以前会设成一个又哑又健康的东西

10:12

I have a bunch bunch of foodies as kids

我有一大堆吃货孩子

10:14

I will use this metaphor a bit more often I like it

我会更经常用这个比喻，我喜欢它

10:17

thanks Brandon

谢谢 Brandon

10:18

I love it and then perturbations is when you somehow perturb

我很喜欢，然后 perturbations 就是你以某种方式去 perturb

10:24

the genome or the transcriptome

genome 或 transcriptome

10:26

so that it has different transcripts in it

这样它里面就会有不一样的 transcripts

10:29

and then typically that's done with a crisper knockout

然后通常是用 crisper knockout 来做的

10:32

where you remove a gene from the genome

也就是从 genome 里移除一个 gene

10:35

and that causes it to not express that gene obviously

那显然会导致它不表达那个 gene

10:38

and then from there then you see what

然后从那里开始，你就会看到

10:40

are the phenotypic or the actual changes in the physical biology

是 phenotypic 的变化，还是 physical biology 里实际发生的变化

10:45

that occur because of a perturbation

这些变化是因为某个 perturbation 而产生的

10:47

that more or less right

这差不多是对的吧

10:49

maybe I can take a step back there

也许我可以在这里先退一步

10:50

because I think RJ you're fully right

因为我觉得 RJ，你完全说得对

10:52

but this is one of the hard core type of perturbations

但这是最硬核的那类 perturbation 之一

10:54

like fully knocking out a gene

比如把某个 gene 完全 knocking out 掉

10:56

sort of in evolution happens a bit

在进化中，这种事多少会有点发生

10:58

but mostly these things are sort of a tad more subtle

但大多数情况下，这些东西其实要更微妙一点

11:00

but of course in order to understand the system

但当然，为了理解这个系统

11:02

it's a good idea to take out bits and pieces

把一些零碎的片段拆出来是个好主意

11:04

I think what happens on a normal development

我觉得在正常发育过程中会发生的是

11:07

or normal sort of us being in homeostasis

或者说我们正常处于 homeostasis 时

11:09

is of course that these cells sort of talk to each other all the time

当然就是这些细胞一直在互相交流

11:12

you know they're sort of living these complex

你知道，它们有点像在过着这种复杂的

11:14

multicellular states right

multicellular states，对吧

11:15

so there's all kind of communications happening between cells

所以 cell 之间发生着各种各样的 communication

11:17

so they always have stuff coming from the outside

所以它们总是有东西从外面过来

11:20

and then sort of doing something to them

然后有点对它们做些什么

11:22

so for me a perturbation space in fact is just an input space

所以对我来说，perturbation space 实际上就是一个 input space

11:25

to stuff that happens to a cell

对发生在 cell 上的东西

11:27

just maybe because it just sits to another one

也许只是因为它就挨着另一个

11:29

and there's like some signaling happening right

然后就像有一些 signaling 在发生，对吧

11:31

or some grade and some homeones coming in whatever right

或者某个等级、一些激素进来，随便吧，对吧

11:33

and all this stuff of course you cannot sort of tune up or tune down

当然，所有这些玩意儿你没法随便往上调或往下调

11:37

maybe even something as basic as

甚至可能是一些最基础的事情，比如

11:39

when you go these in vitro systems

当你用这些 in vitro systems 的时候

11:40

where farmers are struck screens

在那里 farmers 会撞上 screens

11:42

when you sort of put them in a dish

当你把它们放进培养皿里

11:44

you can do different what's called media conditions

你可以做不同的、所谓的 media conditions

11:46

and they so have so much impact right

而且它们的影响特别大，对吧

11:48

and it's like a bit of for me

而且对我来说，这有点像是

11:50

still like a bio-driven maybe even like an alchemy

还是一种 bio-driven 的东西，甚至有点像 alchemy

11:54

what type of cocktails people come up with

人们会想出什么样的 cocktails

11:56

to then sort of maybe take you know skin cells

然后也许就，你知道，拿 skin cells

11:59

and turn them into into more or less

把它们变成，差不多是

12:00

embryo like state cells right

embryo-like state cells，对吧

12:02

where you can then sort of make new cells

然后你就可以借此造出新的 cells

12:05

to heal different stuff in the body

去修复身体里各种不同的东西

12:06

or these what are called Yamanaka factors

或者这些所谓的 Yamanaka factors

12:08

for example these were noble laureate type of stuff

比如这些就是 Nobel laureate 级别的东西

12:12

right so I think having a space of perturbations

对吧，所以我觉得，有一个 perturbations 的空间

12:16

essentially of things that you can do to a cell

本质上就是你能对一个 cell 做的各种事情

12:18

this can be a very rich thing

这可以是非常丰富的一件事

12:20

it could can go on as much as actually changes in

它可以一直发展到，实际上是在

12:22

in the gene which would be like a full knockout

在 gene 里做改变，那就像是 full knockout

12:23

or something or subtle right which is sort of human genetics

或者很细微，对吧，这就有点像 human genetics

12:26

and I think all of that

而且我觉得所有这些

12:27

we need to put in a big space of stuff that we can do

我们都需要放进一个很大的、涵盖我们能做的各种事情的空间里

12:30

that has an influence of cell state

那些会对 cell state 产生影响的东西

12:32

and like a virtual cell model should take all of it

而像 virtual cell model 应该把这一切都纳入进来

12:34

right and also combinations there of of course

对吧，当然还有它们的组合

12:36

and maybe you have disease happening time

也许还有疾病发生的时间

12:37

whatever right that then impact that

不管是什么，对吧，然后都会影响那个

12:40

that's a great connection to the organoids idea

这和 organoids 的想法能很好地联系起来

12:45

so can you talk a little bit about okay so

所以你能不能稍微聊一下，好，那么

12:47

an organoid is what and how does that connect to this

organoid 是什么，它又怎么跟这个

12:52

concept of perturbation space and what do you do

perturbation space 的概念联系起来，以及除了这个 context 本身之外

12:55

to it in addition to just this context

你还会对它做什么

12:58

when I when I first heard about the learned about this concept

当我第一次听到、了解到这个概念的时候

13:00

I should say you know I'm an other trained biologist

我得说，你知道，我是另一名受过训练的生物学家

13:02

so I typically learn about things when I read a bit

所以我通常是在读一点东西的时候了解这些

13:04

or when I talk to colleagues and friends

或者是在跟同事和朋友聊天的时候

13:06

and when I first saw that people are able to take

然后我第一次看到，人们基本上能从 cancer 或随便什么的身体里取出 cells，对吧，

13:10

essentially cells out of a body of a cancer

再让它们长成小小的 multicellular systems，而且这些系统真的有点开始跟别的东西连接起来，我当时超级超级震撼。

13:12

or whatever right and let them then grow into small

所以你知道，现在人们比如可以取一些 cells，放进培养皿里，

13:15

multicellular systems that actually sort of start

然后养出小小的 mini brains，基本上就是一组 neurons。

13:17

connected to things I was super super impressed

连接到一些让我印象非常非常深刻的东西

13:20

so you know people can now for example

所以你知道人们现在可以比如说

13:22

take some cells put in a dish and then grow

取一些细胞放进培养皿里然后培养出

13:25

little mini brains essentially set of neurons

小小的迷你大脑本质上就是一组神经元

13:28

that essentially really start connecting

那基本上就真的开始连接起来了

13:30

talking to each other and people have shown

彼此交流，而且人们已经证明

13:32

that they can get those sort of mini brains

他们能弄到那种 mini brains

13:34

in a dish to compute similarly people can take heart cells

放在培养皿里进行计算；类似地，人们可以取来 heart cells

13:38

grow them into little heart organoids and we have

把它们培养成小小的 heart organoids，而我们有一堆

13:40

a bunch of local corporations with fantastic partners

本地公司，还有很棒的合作伙伴

13:43

that actually start beating like you have

它们真的开始跳动了，就像你有

13:44

beating little heart cell clumps in a dish right

在培养皿里跳动的一小团一小团 heart cell clumps，对吧

13:47

and then of course once you have these and your same

然后当然，一旦你有了这些，而且你的——同样，

13:50

thing is I spoke about brain there's even

我之前提到过 brain，甚至还有

13:52

simple systems like the colon where this works very well

像 colon 这样很简单的系统，这个方法在那里效果也很好

13:54

there's a whole bunch of cancer systems

还有一大堆 cancer systems

13:56

but you know also for let's say if you want to study

但你知道，另外，比如说如果你想研究

13:59

diseases with heart to acquire material

心脏相关的疾病来获取材料

14:01

let's say you want to look into the genetics

比如说你想研究某种

14:03

of some type of of neurodegenerative diseases or so

neurodegenerative diseases 的 genetics 之类的

14:06

actively it's very hard to get access to tissue right

主动去获取 tissue 其实非常难，对吧

14:08

so you take stem cells from the skin

所以你会从皮肤里取 stem cells

14:10

it takes out from the skin did you differentiate

从皮肤里取出来之后，你再做 differentiate

14:12

and make turn them into these middle brains

然后把它们变成这些 middle brains

14:14

and then you can potentially study some some things there

然后你就有可能在那里研究一些东西

14:16

so I think this concept of replicating little bits of organs

所以我觉得，复制一小块一小块的器官这个概念

14:20

it's by the called organoids in a dish

它被称为 organoids in a dish

14:22

it's very fascinating there's all kinds of limitations obviously

这非常有意思，当然也有各种各样的局限性

14:24

often immune cells this immune system is missing

经常是 immune cells，这个 immune system 缺失了

14:26

there's like no maybe feeding them or something like

就好像没有……也许是给它们喂东西之类的

14:29

so there's all kind of problems right

所以就会有各种各样的问题，对吧

14:31

but it's much much more than just a cell

但这远不止是一个 cell

14:33

and which is like one cell type

而这就像是某一种 cell type

14:34

moving to the modeling side how does how do you model that

转到 modeling 这边，那要怎么、怎么去建模这个呢

14:37

so that's I think a bit the research franchise

所以我觉得这有点像是 research franchise 那块

14:40

so I don't think people have agreed on what the

所以我觉得大家还没就 what the 达成一致

14:44

optimum way to model is particularly

最优的建模方式尤其

14:45

else because that's not as much data as one wants

别的方式，因为那并没有人们想要的那么多数据

14:47

Jay right you're you're involved if I understand

Jay，对，你——如果我理解

14:49

correctly in a whole bunch of spatial omics

没错，参与了一大堆 spatial omics

14:51

slash also sort of H&E sort of pathology type of

以及也算是 H&E 那种 pathology 类型的

14:57

average pathologist like so much samples are very

普通 pathologist，就像那么多样本都非常

15:01

right and now increasingly we sort of do a bit more

对，而且现在我们越来越会多做一点

15:03

spatial omics so I think how we even represent

spatial omics，所以我觉得我们到底该怎么 represent

15:07

spatially connected data sets and then you sort of break it

spatially connected data sets，然后你差不多把它拆开

15:10

down on an organoid level and so on people are not

到 organoid 层面等等，人们并不

15:12

as clear as it is with this associated cells so

像对 associated cells 那么清楚，所以

15:16

again taking taking this a bit simpler I think

再把这个稍微简化一点，我觉得

15:19

if we look at this space of this associated cells

如果我们看 associated cells 的这个空间

15:22

and we look at these groups we have this beautiful

再看这些分组，我们有一个很美的

15:24

metaphor of what the humans call a cell type

隐喻，也就是人类所说的 cell type

15:26

which you know nature doesn't care about

而你知道，自然并不在乎这个

15:27

but this is just us putting categories

但这只是我们在把 categories 加到

15:29

in top of stuff but it helps us understand

东西上面，不过它能帮我们理解

15:30

this complexity I don't think we have a similar

这种复杂性。我不觉得我们有一个类似的

15:33

metaphor for low connexious of cells maybe a little bit

隐喻来形容细胞的 low connexious，也许有一点

15:39

but also in H&E right whenever I talk and see

但同样在 H&E 里，对吧，每次我聊天、看到

15:42

what pathologist do it's often very qualitative

pathologist 做的事，往往都非常 qualitative

15:45

and it's maybe not and that level yet so I think

而且可能还没到那个水平，所以我觉得

15:47

we need to develop a bit of a categorization

我们需要建立一点 categorization

15:49

language for us maybe computation maybe we don't

语言对我们来说，也许是 computation，也许我们并不

15:52

need this so much but computation we also need to

那么需要它，但 computation 我们也得

15:54

sort of think a bit how we deal with this at the

稍微想一下，我们现在到底在

15:56

moment we essentially deal with this as

这个时刻怎么处理，其实我们基本上把它当作

15:58

distributions of of cells so this is sort of

cells 的 distributions，所以这算是

16:00

the solution at the moment but I think

目前的解决方案，但我觉得

16:02

you can be more much more subtle like graphs or something

你可以做得更细得多，比如 graphs 之类的

16:04

like that so on the one hand you have the bitter

类似这样，所以一方面你有那个 bitter

16:06

lesson and you know sort of you you're

这个经验教训呢，你知道，差不多就是，你你

16:08

a move inductive bias and you end up being able

去掉 inductive bias，然后你最终就能

16:11

to have a larger data set that's applicable

拥有一个更大的 data set，它能适用于

16:14

to your problem and then that helps you to get

你的问题，然后这能帮你得到

16:16

more accurate models because your your

更准确的 models，因为你的你的

16:19

distribution is wider and whatever so that's

distribution 更宽啊什么的，所以那就是

16:21

like the one and then but then there's this

像是那一面，然后但然后又有这种

16:24

like countervailing tendency which is like

像是 countervailing tendency，它就像是

16:28

maybe you could call it the anti-bitter lesson

也许你可以把它叫做 anti-bitter lesson

16:30

or like the flip side of the bitter lesson where

或者说像是 bitter lesson 的反面，在那里

16:32

when you're in data poor regimes for one thing

当你在某件事上处于 data poor regimes 时

16:37

then you have to introduce introduce a lot of

那你就必须引入、引入大量

16:40

inductive bias to get enough statistical power

inductive bias，来获得足够的 statistical power

16:42

to make good inferences but there's also a second

从而做出好的 inference，但还有第二个

16:45

side of that which I think John Platt in our

方面，我觉得 John Platt 在我们

16:48

discussion with him not too long ago brought up was

不久前和他的讨论中提到的，是

16:50

that you have these these like predictive

就是你有这些，这些像是 predictive

16:52

and then descriptive models and at what you

然后还有 descriptive models，而作为科学家你

16:55

really want as a scientist is to get to a

真正想要的是达到一种

16:58

simplification a descriptive model where you

简化，一种 descriptive model，在那里你

17:00

can say oh well I've tested all the apples

可以说，哦好吧，我测试了所有苹果

17:03

and the gravity works in such a way on apples

而重力对苹果的作用方式是这样的

17:06

and then but I'm modeling this as a universal

然后但我把这 modeling 成一个 universal

17:10

concept simply so that we know that it works on

concept，只是为了让我们知道它适用于

17:12

planets as well right and and so that I want

行星也一样，对吧，而且，而且所以我想

17:16

to understand from you what are the

从你这里了解，有哪些

17:18

simplifications or you know sort of maybe

简化，或者说，你知道，算是，也许

17:21

simplifications maybe insights maybe

简化，也许是洞察，也许

17:23

actual mathematical or otherwise

实际的数学的，或者其他方面的

17:26

scientific models that are becoming a parent

科学模型，它们正在逐渐显现出来

17:30

that are going to help to reduce all the

这些将有助于减少所有的

17:32

complexity so that we're able to actually make

复杂性，这样我们才能真正做出

17:35

meaningful gain meaningful understanding

有意义的提升，有意义的理解

17:37

from these really large data sets.

从这些真正很大的 data sets 里得到的。

17:39

Yeah that's that's a that's a very important

嗯，那、那是——那是一个非常重要的

17:42

very I think kind of deep point which I think

非常——我觉得算是挺深的一个点，我觉得

17:47

sort of breaks the field into maybe sort of a bit

算是把这个领域分成，也许算是有点

17:51

two different groups of people with answers right

两个不同的、有答案的群体，对吧

17:55

so how much do we need to put inductive biases

所以我们需要放多少 inductive biases

17:58

maybe priors into a model to maybe make it

也许还有 priors 到 model 里，才可能让它

18:04

better explainable right but potentially also

可解释性更好，对吧，但可能也

18:07

more robust and going out of distribution or how

更 robust，以及 out of distribution 的情况，或者说

18:10

much are we going to just rely on bigger data

我们在多大程度上会只是依赖更大的 data

18:12

to sort of have the model be more expressive

来让 model 更有表现力一点

18:14

and so on and I think and there I I need to

等等，而且我觉得，还有，我我我需要

18:17

be very clear before we may go into the

非常明确，在我们可能进入

18:18

philosophy I think literature has really

哲学之前，我觉得文献真的

18:20

clearly shown over I think now decades at

清楚地表明，我想现在几十年来

18:24

least for also very many data sets that if we

至少对很多 data sets 来说也是这样：如果我们

18:27

build in a priori inductive bias is too

把 a priori inductive bias 建得太强，

18:30

strongly that we might lose predictive power

就可能失去 predictive power。

18:33

right so it comes at a cost and if you now

对吧，所以这是有代价的。那如果你现在

18:35

let's say a physician and you just want to

比如说你是个医生，你只是想要

18:38

or maybe even if you're a patient and you know

或者哪怕你是个病人，你知道

18:39

you maybe you might have cancer not maybe you

你可能有癌症，不，也许你就是

18:42

want to have the best model as a predictor

想要最好的 model 来做 predictor

18:44

maybe she getting treated you want to have a

也许她在接受治疗，你想要有一个

18:46

model that is maybe a bit less good but

model，它也许稍微差一点，但

18:48

you understand it a bit more but for the

你对它的理解会多一点，但就

18:49

prediction at least I think you you want to

prediction 来说，至少我觉得你你想要

18:52

have the best and this is actually I think

最好的，而这实际上我觉得

18:53

not not entirely clear and I think in our

不、不完全清楚，而且我觉得在我们

18:56

field with this perturbation space scale the

领域里，用这个 perturbation space scale，这个

18:59

south guy I think is big but the patient

south guy 我觉得很大，但那个病人

19:00

very space scale like the stuff that we do

像我们做的那些东西那样的 very space scale，到它没那么大——我觉得我们还没进入 super data rich 的 regime。

19:02

to it not being as big I don't think we're

所以我觉得，在我们这种情况下，这些 hybrid markets 可能还是值得走的方向。

19:05

yet in the super data rich regime so I think

但当然，越来越多——这个我觉得，这，我不知道这算不算可悲，但这是从 big scale machine 得到的诚实 learning。

19:07

in our case these hybrid markets might be

在我们的情况下这些混合市场可能仍然

19:09

still the thing to go for but more and more

是值得追求的方向但当然越来越

19:12

of course and this is I think this I don't

而且我觉得这个我不知道

19:14

know if you can call it sad but the honest

你能不能称之为可悲但说实话

19:16

learning from from big scale machine

从大规模机器学习

19:17

learning is that at some point data beats

learning 就是，到某个时候，data 会胜过

19:20

human made additional concepts in these

人类在这些里面额外加入的概念

19:23

and I have this one simple example I often

我经常举一个很简单的例子

19:25

put is alpha fold where you know the first

就是 alpha fold，你知道，最早的

19:27

models sort of made this extraordinary

models 算是做出了这种非凡的

19:29

effort to to make you know the molecules

努力，去让，你知道，分子

19:30

rotation invariant is sort of building

rotation invariant，算是把它构建

19:32

into the learning process and only in

进 learning process 里，而直到

19:34

later versions you know sort of rotation

后来的版本，你知道，有点像 rotation

19:36

symmetry was automatically found by the

symmetry 是被

19:38

model just with more data and was a bit

model 自动找到的，只是靠更多 data，而且稍微

19:39

more flexible because whatever physical

更灵活了一点，因为不管你放进去的是什么 physical

19:41

constantly put in it's not always as rigid

constant，它并不总是那么 rigid

19:43

and then you sort of you gain something

然后你有点像是会得到一些东西

19:44

if you let the model learn from this so I

如果你让 model 从这里面学习，所以我觉得

19:48

think in the future the bigger part at least

未来更大的那部分，至少

19:51

of the sort of more used models will be

那种更常用的 models 会是

19:53

those more flexible very much database

那些更灵活的、非常偏 database 的

19:55

ones where you then apostuary put

也就是你随后 a posteriori 把

19:57

interpretation on top this also brings

interpretation 加在上面的，这也带出了

20:00

up the the question for me so that you

对我来说的那个问题，就是你

20:02

actually I think recently wrote a

其实我觉得最近写了一篇

20:05

or co-author paper on the statistical

或者 co-author 了一篇关于 statistical

20:08

power that the current generation of

power 的 paper，也就是当前这一代的

20:10

virtual cells have in terms of the data

virtual cells 在 data 方面有什么

20:13

that the the lift that you get from

也就是你从

20:16

additional data in the current

additional data 中获得的 lift，在当前

20:18

generation of virtual cell models and I

这一代 virtual cell models 里

20:20

think you said basically that the objective

而且我觉得你基本上是说，objective

20:23

function or the training methodology is

function 或 training methodology 才

20:25

potentially the problem there and not

可能是那里的问题，而不是

20:27

necessarily that the information is not

一定是因为信息不……

20:29

in the data so like what can you explain

在数据里，所以比如你能解释一下

20:33

that a little bit and why you think that maybe

那一点吗，以及你为什么觉得也许

20:34

there's plenty of untapped information in

信息里有很多尚未发掘的信息

20:37

the information and then like what where

然后比如那会带我们到哪里，我们还需要

20:40

does that get us and what what do we need to

做什么才能更进一步？对，我觉得可以公平地说

20:42

do to get further yeah I think it's fair to

你知道，早期我们有一些

20:45

say that you know early on we had some

那种非常小的、更有限的

20:48

sort of very small sort of more limited

有点小，有点更受限

20:50

prediction models of learning that say

那些关于学习的 prediction models 说

20:53

late in space is first and also sort of

在空间里靠后的反而排第一，而且还有某种

20:54

perturbations on top and then you know when

叠加在上面的 perturbations，然后你知道，当

20:58

like this really large-scale foundation

这种真正 large-scale 的 foundation

20:59

model representation came out everyone

model representation 出来的时候，大家

21:00

was sort of buying into the hype and

都有点买这波 hype 的账，而且

21:02

maybe potentially a bit bit over buying

可能还有点过度买单了

21:04

then you know obviously after this you know

然后你知道，显然在这之后，你知道

21:05

comes what's the called the the sobering

接下来，那个叫什么来着，那个 sobering

21:07

up value or something like that yeah and

up value，或者类似的东西，对，然后

21:09

then there were a whole bunch of papers

然后有一大堆论文

21:10

coming up where they were she saying you

陆续出来，里面他们在说，你

21:11

know hey maybe even more or less linear

知道，嘿，也许甚至差不多就是个 linear

21:14

model we're doing a good job and actually

model，我们做得不错，而且实际上

21:15

now I see Jen paper early on we had

现在我看到 Jen 的论文，早期我们已经有

21:17

already some of these benchmarks showing

其中一些 benchmarks 在显示

21:18

yet it's not so not so super easy in simple

不过也没那么、没那么容易，在简单

21:21

tasks actually to improve on that um I

任务上，其实要在这方面提升，嗯，我

21:25

think I think there's a truth to it that

觉得，我觉得这里面有一定的道理，尤其是

21:27

particularly in the current data state

在当前的数据状态下

21:29

for some things let's say you want to predict

对有些事情来说，比如说你想预测

21:31

the effect of a crisper in a cell and you

CRISPR 在细胞里的效果，那你

21:33

might not need the most fancy model whereas

可能并不需要最花哨的模型，而

21:35

if you ask a bit of if you want to predict

如果你稍微问一下，如果你想预测

21:37

cell type or something like that you know

cell type 之类的，你懂的

21:38

for that mostly interesting regression is

对那个来说，比较有意思的 regression 就

21:40

fine but if you want to predict something

还行，但如果你想预测一些

21:42

a bit more difficult for example in our

稍微更难一点的东西，比如在我们

21:44

niche former paper which is a foundation

那篇 niche former paper 里，它是一个 foundation

21:46

model for spatialomics we showed we had a

model for spatialomics，我们展示了我们有一个

21:48

new task which we introduced where you

新任务，我们引入了这个任务，其中你

21:50

you say you have a you have the gene

你说你有，你有一个，你有那个 gene

21:52

expression you have sort of this state a

expression，你大概有这样一种状态，啊

21:53

latent space of a single cell can you

一个 single cell 的 latent space，你能不能

21:55

predict the gene expression state of

预测 gene expression state，属于

21:58

its neighboring cells because you know

它 neighboring cells 的，因为你知道

21:59

they're always in communication or maybe

它们一直在 communication，或者也许

22:01

the other way around so like they're always

反过来，就像它们一直

22:02

in communication so because you know

在 communication，所以因为你知道

22:03

they talked each other you sort of see

它们彼此交流，你大概就能看出来

22:05

a remnant of that sort of cell state in

那种 cell state 的残留还留在这一种里面，而且实际上这更难做到

22:07

this one and actually this is harder to do

而且如果你有这些更难的任务，你就需要更强、更 non-linear 的东西

22:09

and if you sort of have these harder tasks

你尤其需要去找更细微的 data，所以你的问题是

22:10

you need to stronger more no linear

你能不能找出更多那种 information-rich 类型的东西，我觉得

22:12

things and you particularly need to look

22:14

for more subtle data so your question

22:16

can you sort of find out more sort of

你能稍微多了解一些吗？

22:19

information rich type of things I think

我觉得是信息量丰富的那类东西

22:21

very much depends on what you sort of ask

很大程度上取决于你具体问的是什么

22:23

at the end of it of course like in the

当然，说到底，就像在

22:24

self supervised learning phase I think

self supervised learning 阶段，我觉得

22:26

we definitely not there yet with the best

我们绝对还没达到最好的

22:29

way how we represent how we learn these

方式，也就是我们如何表示、如何学习这些

22:32

things there's a lot of copying of what

东西，这里面有很多是在照搬

22:34

happens in NLP type of LLM's right and

NLP 类型的 LLM 里发生的那套，对吧，而且

22:36

like massed alternate order so what's a

就像 massed alternate order 那样，所以什么是

22:38

typical way to go but the type of losses

通常的做法是这样，但 loss 的类型

22:41

the type of representations I think

representation 的类型，我觉得

22:42

people are still debating maybe because

人们还在争论，也许是因为

22:45

you know this is sort of often kind of

你知道，这往往有点像是

22:46

tabular data so even just like tabular

tabular data，所以甚至就像 tabular

22:48

foundation models and also being tried out

foundation models，而且也在被尝试

22:50

there's a lot of japa type of approaches

现在有很多 JEPA 类型的做法

22:52

now because there's like we have one

因为现在就像我们有一个

22:53

thing where we use contrastive learning

就是用 contrastive learning 的那种做法

22:54

so just people playing around with the losses

所以就是大家在 losses 上各种折腾

22:56

but also with representations and I think

但也会折腾 representations，而且我觉得

22:59

nowadays reviewers and at this

现在审稿人还有在这个

23:00

I've been advocating for a long time are

我长期以来一直主张的是

23:02

really much more closely looking that you

真的会更仔细地看，就是你

23:03

really properly show scaling loss

真的要把 scaling loss 好好展示出来

23:05

this was what has missed from the early

这正是早期所缺失的东西

23:07

papers of the foundation model and so

foundation model 的论文之类的

23:09

people haven't really shown that you need

人们还没有真正证明你需要

23:10

this complexity to actually sort of

这种复杂性来真正算是

23:12

really scale up but if you go really if you

真正地 scale up，但如果你真的，如果你

23:14

can really demonstrate at scale like across

能真正在 at scale 上展示，比如跨越

23:17

examples and samples that you have this

例子和样本，你有这种

23:18

sort of scaling behavior then I think you

算是 scaling behavior，那么我认为你

23:20

have this bind into the smallest and for

有这种绑定到最小的，并且为了

23:22

these of course there's not so much

当然，这些其实没那么多

23:24

at least inductive biases in these

至少在这些用途里的 inductive biases

23:25

use and so I do think we're gonna

所以我觉得我们确实快要

23:27

gonna hit this regime actually quite

真正进入这个 regime 了，而且实际上相当

23:28

soon and we do see this also in a few

快，我们也在几篇

23:30

papers so for example in our contrastive

论文里看到了这一点，比如在我们的 contrastive

23:32

learning paper we call it SD concept we

learning 论文里，我们把它叫作 SD concept，我们

23:34

do show already quite quite strong

确实已经表现出相当相当强

23:36

scaling behavior and the paper we had

scaling behavior，还有我们那篇论文，

23:38

something for going because we compared

有个东西要推进，因为我们对比了

23:39

with gene form are going to maybe 30

Geneformer，要走到大概 3000 万个 cells，

23:41

million cells but on on hugging phase now we

但在 Hugging Face 上，现在我们有

23:43

have a 350 million cell model with I think

一个 3.5 亿 cells 的 model，我觉得

23:46

going up to maybe 100 million parameters

parameters 会到大概 1 亿，

23:48

so really going to the big scale for us

所以对我们来说真的是进入很大的 scale 了，

23:49

and we do see nice scaling behavior

而且我们确实看到了很好的 scaling behavior

23:51

already starting to emerge just to be

已经开始显现了，先说明白

23:53

clear then at 100 million parameters

然后是在 100 million parameters 的时候

23:55

and three you said I think 300 million

还有三，你说我觉得是 300 million

23:57

cells or something like that that you

cells 之类的，就是那个你

23:59

you get a lift in what kinds of

会在哪些类型的应用里获得提升

24:01

applications in these these sort of

在这些，这些算是

24:03

the niche former kinds of questions

偏 niche 的、以前那类问题里

24:05

where you're saying what is my

你会说，我的……是什么

24:07

environment based on the my transcript

基于我的 transcript 的环境

24:11

yeah I mean you know that the ultimate

是啊，我是说，你知道，我们最终当然想做的是把 perturbation prediction 做得更好，因为现在很多人理解的 virtual surround，很大一部分就是这个；而就这一点来说，我不会说纯 representation learning

24:12

thing of course we want to do is we want

当然我们想做的一件事是，我们想

24:13

to be doing this perturbation prediction

做这个扰动预测

24:16

better because that's a lot of what

更好，因为那其中很大一部分是

24:16

people nowadays understand by virtual

现在人们理解的虚拟

24:18

surround and for this I would not

围绕，而且为此我不会

24:21

claim that pure representation learning

声称纯表征学习

24:23

is the right way to go sorry for making

是正确的做法，抱歉又

24:24

the subtlety again but I think typically

把这点微妙之处说了一遍，但我觉得通常

24:26

you learn the state space with these

你是用这些

24:27

foundation models but then sort of the

foundation models 来学 state space 的，但然后有点像

24:29

way how they behave under an action

它们在一个 action 下如何表现

24:31

on a perturbation this is where you need

在一个 perturbation 下，这就是你需要

24:33

this more world model like type of

这种更像 world model 的

24:35

approaches world model being the very

方法，world model 就是非常

24:36

hip term I'm sure you had this already

这个流行词，我确定你在播客里某个时候已经听过了，

24:38

at some point in your podcast

所以过去一年它出现在所有 machine learning 会议上，但他们把 perturbation prediction 当作一个因素，

24:39

so it's been on all of a machine learning

但更像是 virtual scientist 那种 world model，就 transcriptome 而言，

24:41

conferences in last year but they

或者是在这个 space 上的那类东西，对吧，所以它……

24:42

factor a perturbation prediction but

factor一个perturbation预测但是

24:44

like a virtual scientist kind of a

就像一个虚拟科学家那样的

24:45

world model in terms of transcriptome

transcriptome 意义上的 world model

24:47

or the type of on the space right so it's

或者是在空间上的那种类型，对吧，所以就是

24:49

it's not just a model that is sort of

这不只是一个 model，不是那种

24:50

f of x but f of x comma a right so the

f of x，而是 f of x comma a，对吧，所以就是

24:53

state x plus the action right and then

state x 加上 action，对吧，然后

24:55

sort of predicting the next state x

有点像预测下一个 state x prime，差不多是这样，而且我觉得

24:56

primal something like that and I think

对这个来说，你可能需要其他的

24:58

for this you might need other

组合，所以我们做的是 big new，我们用 for 来学习 representations

25:01

combination so what we do big new we

combination，所以我们做的是大的新东西，我们

25:02

learn the representations with for

学习表征，用

25:04

example as a concept but then

先拿个例子当概念，但接着

25:06

we put a generative model for the

我们放一个 generative model 来

25:09

push forward along the action so for this

沿着 action 往前推，所以对这个

25:11

we use something like flow matching

我们会用类似 flow matching 的东西

25:12

or something like that and in our

或者类似这样的方法，而在我们的

25:13

self-flow paper this is not just

self-flow 论文里，这不只是

25:15

I got accepted for nature methods we

我有一篇被 Nature Methods 接收了，我们

25:16

actually really show that these

其实真的展示了这些

25:20

approaches are much better generating

这些方法在生成方面要好得多

25:22

the effect of of of perturbations but

perturbations 的影响，但是

25:23

they're very much dependent on a good

它们非常依赖于一个好的

25:25

latent space and you know this as a

latent space，而且你知道，这算是一个

25:26

contemplating space in this case for

contemplating space，在这个情况下，对于

25:28

these 350 million cells we actually

这 3.5 亿个细胞，我们实际上

25:30

went into multi species so you have

进入了 multi species，所以你现在有

25:32

now like 16 different species where

大概 16 个不同的物种，其中

25:33

you can potentially look for

你有可能去找

25:34

similarities suddenly here you can maybe

相似之处，突然之间，在这里你也许能

25:36

ask some evolutionary questions that

问一些演化问题，这些

25:37

you couldn't before

你以前问不了

25:38

not that we can do this so well not

不是说我们能做得这么好，不是

25:40

our expertise but you know so if you can

我们的专长，但你知道，所以如果你能

25:42

start doing these fancy sort of cross

开始做这些花哨的、跨

25:45

species cross tissue type of latent

物种、跨组织类型的 latent

25:47

spaces I'm kind of curious at the

我有点好奇这些不同 models 的历史。

25:48

history of some of these different

我觉得你刚才给我们讲了一个挺快的版本。

25:50

models I think you just gave us a

其中有一件事我觉得在社区里还算有点名，

25:52

sort of a rapid version of them the one

就是关于 linear models 是个挺重要的点这件事，

25:55

of the things that like I think was

你能解释一下吗？

25:58

like somewhat famous in the community

在社区里有点出名的是

25:59

was this thing about linear models being

关于线性模型是一个

26:01

kind of a big point can you explain

很大的点这件事，你能解释一下

26:04

specifically what was the like key point

具体来说，关键点到底是什么

26:06

here but I think this is like one of

但我觉得这就像是经典 machine learning 里的那种事

26:07

those things the classic machine learning

一定要始终理解清楚你的 baselines

26:09

make sure you always understand your

你懂的，不然你就会

26:10

baselines well you know otherwise you

花很多时间原地打转

26:12

you spend a lot of time spending your

还自欺欺人，我觉得那具体是一个很有名的 open

26:13

wheels and fool yourself I think it was

wheels and fool yourself 吗，我觉得那是一个

26:15

a specifically a famous open

特别有名的开放

26:17

competition and with the through the

竞争，还有通过那个……

26:19

arc institute that was a sort of a

Arc Institute，那有点像一

26:22

class of like cold water and everyone's

盆冷水，然后大家的

26:25

you know faces can you maybe give a

你知道的表情……你能不能稍微讲一

26:27

little bit of history about that and like

点那段历史，然后像

26:29

why we thought this was exciting and then

为什么我们觉得这很令人兴奋，然后

26:32

where why this had us pivot a bit yeah it

哪里、为什么这让我们稍微转了下方向，对，它

26:35

sounds such an like such an obvious

听起来真的特别、特别明显

26:38

learning right to your baselines properly

把你的 baselines 真正正确地学到位

26:40

but in a field that is rushing towards the

但在这个正冲向

26:42

gold part of the virtual cell people

virtual cell 黄金部分的领域里，人们

26:44

sometimes maybe were not as careful with

有时候可能没那么小心地对待

26:46

the baselines and there's a whole bunch

baselines，而且有一大堆

26:48

of subtleties with respect to hold it and so

关于 hold it 之类的细微之处，等等

26:50

on that you know it wasn't always fully

你知道，这些并不总是被完全

26:51

observant and why some of these things

留意到，以及为什么其中一些事情

26:53

happen so one of the first observations

就发生了，所以最早的几个观察之一

26:55

there was in a thing as basic as

是在一件和

26:57

cell type prediction okay so nothing super

cell type prediction 一样基础的事情上，好吧，所以没什么特别

26:59

fancy okay basically you you give me a

花哨的，好吧，基本上就是你给我一个

27:01

new transcriptome tell me from which

新的 transcriptome，告诉我它来自哪个

27:03

cell type it is right maybe even from

cell type，对吧，甚至可能来自

27:05

organ or something like that and it turns

器官或者类似的东西，而结果发现

27:07

out that even nowadays I think some

即使在今天，我觉得有些

27:10

close to state of the art models I really

接近 state of the art models，我真的

27:12

just logistic aggression right so for

就是 logistic aggression，对吧，所以对于

27:14

such a simple categorization you know

这么简单的一个 categorization，你知道

27:16

this is actually not not so bad there's

这其实还真不算太差，有

27:18

one thing called cell type is for example

一个叫 cell type 的东西，比如说

27:19

which is doing good job and I think we had

它做得挺好，而且我觉得我们有过

27:20

a paper maybe five or six years ago where

一篇论文，大概五六年前，当时

27:22

we had the title very plain deep learning

我们的标题非常朴素，就是 deep learning

27:26

is not needed for cell type prediction

对 cell type prediction 来说是不需要的

27:28

this is true if it's within an organ if you

如果是在一个 organ 内，这确实成立，如果你

27:30

sort of go cross organ suddenly gets a

有点跨 organ 之后，突然就会变得

27:32

bit more subtle because you know so

更微妙一点，因为你知道嘛，所以

27:33

cell types we occur in different organs

cell types 会出现在不同的 organ 里

27:35

and immune cells subtle subtle change

而 immune cells 会有很细微很细微的变化

27:37

depending which tissue there suddenly you

取决于是在哪个 tissue 里，突然你就

27:39

know there you need to do bit more and we

知道在那儿你需要多做一点，然后我们

27:41

like for us where was my own paper called

比如对我们来说，我自己那篇论文叫

27:43

SCTAP like a tabular foundation model

SCTAP，是个像 tabular foundation model

27:45

predictive model type of thing we actually

predictive model 那种东西，我们其实

27:47

sure that this modern area was really

确认了这个现代领域真的

27:48

adding beyond the baseline so that was

在 baseline 之外有增益，所以那算是

27:51

sort of for that class for the

算是针对那一类的

27:53

perturbation modes which then you were

perturbation modes，然后你

27:54

asking Brandon similarly like early on

问 Brandon 类似地，就像早期那样

27:56

many analysis particular perturbacy

很多针对特定 perturbation 的分析

27:58

were done with classical linear models

都是用经典的 linear models 做的

28:00

in pseudo box we had then our SCGEN

当时在 pseudo box 里，我们有了 SCGEN

28:02

that was like something around 2019

那大概是 2019 年前后

28:04

where we started introducing linear shifts

那时我们开始引入 linear shifts

28:06

but in the latent space so sort of non-linear

但是是在 latent space 里，所以有点 non-linear

28:08

encoded and we always compared with subtle

编码后的，而且我们一直拿它跟那种微妙的

28:10

sort of simpler linear shifts but then

算是更简单的 linear shifts 做比较，但后来

28:12

the field getting complex and you know sort of

这个领域变得越来越复杂，你知道，差不多就是 datasets 不断冒出来，大家开始问越来越多 perturbation 类型的问题，这类东西已经有一大堆了。早期有那种 competition，像是常规的一轮 run，比如 MIT 办的；我们一直在办一个常规的 nureps competition，针对 single cell data 里的 prediction problems，有时候……

28:15

datasets popping up people started asking

数据集冒出来，人们开始问

28:17

more more perturbation type of questions

更多更多扰动类型的问题

28:18

there have been a whole bunch of these

已经有一大堆这类东西了

28:19

early on competition for like a regular one

早期有一个像常规竞赛那样的比赛

28:22

run for example by MIT we've been running

比如由MIT举办的我们一直在办

28:24

a regular nureps competition on prediction

一个常规的nureps竞赛关于预测

28:28

problems in single cell data sometimes

单细胞数据中的问题有时候

28:31

for integration data also for perturbation

用于 integration data，也用于 perturbation

28:32

prediction that was very well attended

prediction，那场活动参与度非常高

28:35

and chairing on the cake was most

而最大的锦上添花是，

28:38

definitely that arc competition which

绝对就是那个 ARC competition

28:41

called the thing not just perturbation

它把这东西不只叫作 perturbation

28:43

prediction challenge but virtual cell

prediction challenge，而是 virtual cell

28:44

challenge to sound a bit cooler right

challenge，听起来更酷一点，对吧

28:46

we picked up very well and but with

我们反响很好，而且但是……

28:48

metrics that aerobus were maybe

aerobus 的那些 metrics 可能

28:50

but maybe not not that easy so for actually

但可能不是，没那么容易，所以其实

28:52

our solution in this consortium

我们在这个 consortium 里的解决方案

28:56

where a lot of labs contributed called

有很多 labs 参与，叫做

28:57

open problems was actually that people

open problems，实际上是让人们

28:59

need to submit their own metrics in addition

还需要提交他们自己的 metrics

29:01

so in order to really quantify maybe

所以为了真正量化可能

29:03

subtle things that you might not have thought

那些你可能没想到的细微之处

29:04

before the arc one is sort of fixed a bunch

在 arc one 之前，算是固定了一批

29:06

of metrics that weren't even sort of

metrics，而这些甚至不是那种

29:07

so single cell specific and they could be

那么 single cell specific 的，它们可能

29:09

just just exploited but actually I think

就只是被 exploit 了，但实际我觉得

29:10

the winning model that was then in the end

最后真正胜出的那个 model，当时到最后

29:12

actually sort of crowned as the actual

其实算是被加冕成了真正的

29:13

winning model but essentially it's a

winning model，但本质上它是一个

29:14

flow matching best thing so based on

flow matching 的 best thing，所以基于

29:16

on this idea that you would be but

关于这个想法，也就是你会是，但

29:17

followed with this flow with this self-flow

跟着这个 flow、跟着这个 self-flow 走

29:19

type of idea I think shows you that

这种想法，我觉得它向你表明

29:22

some non-linear things are in this case

一些 non-linear 的东西在这种情况下

29:23

needed so that the actually real winner

是需要的，这样真正的赢家

29:25

so if you sort of remove the

所以如果你某种程度上把那个

29:27

exploitation was actually something that

exploitation 其实是某种东西，

29:28

this is taking some some non-linearities

这需要一些一些 non-linearities

29:30

and I think nowadays we do see that

而且我觉得现在确实能看到

29:33

just linear additive type of models

就是 linear additive type of models

29:35

I'm not cutting the cake for more complex

我不是在替更复杂的东西说话

29:38

things but for I think even nowadays

但我觉得即使在现在

29:40

if you go in a simple

如果你用一个简单的

29:42

with your cell lines and you have sort of

配上你的 cell lines，然后你有那种

29:44

the genocuts then particularly when

genocuts，尤其是当

29:47

the effects are a bit subtle

effects 有点微妙的时候

29:48

all these sort of more complex models

所有这些更复杂的 models

29:50

are sometimes a bit overfitting and that's

有时候会有点 overfitting，这就是

29:51

why linear based on is always very good so

为什么 linear based on 总是非常好，所以

29:53

I always tell my students guys I know

我总是告诉我的学生，伙计们，我知道

29:55

we are guys I mean guys and girls

我们这些家伙——我是说，男生和女生

29:57

we always need to do I know you want to do

我们总是需要去做，我知道你们想做

29:58

the most fancy deep learning thing

最花哨的 deep learning 东西

30:00

but please please you know have the baseline

但拜托拜托，你知道，要有 baseline

30:02

and like say for drugs you know

而且就像，比如说对于药物来说，你知道

30:04

based on as simple as just predicting

基于就像只是做预测一样简单

30:06

getting better at the mean you know

就是在 mean 上做得更好，你知道

30:08

like that's like the hardest base

就像，那就像是最难的 base

30:09

not even linear model like better than the

甚至不是 linear model，就像比那个

30:11

mean sometimes not as hard as the

mean 有时候也没那么难，就像

30:13

so much noise in the data and this stuff

数据里 noise 太多了，而且这些东西

30:14

is super noisy

noise 超级大

30:15

I just wanted to add one point that

我只是想补充一点

30:17

Bo Wang and some of his collaborators

Bo Wang 和他的一些合作者

30:19

actually just had a nature paper

其实刚发了一篇 Nature 论文

30:21

on on how the metric was actually the

讲的是 metric 其实才是那个

30:24

problem in some of those linear

问题，在一些那些 linear

30:26

based lines beating some of the models

based lines 里，还打败了一些模型

30:27

as well so it's clearly it's not only

所以很明显，这不只是

30:30

metric hacking but also just that

metric hacking，还就是这个原因

30:32

the asking the wrong question leads to the

问错问题会导致

30:34

wrong answer or choice of the metric

错误的答案，或者选错 metric

30:36

yeah and I think that you can

对，而且我觉得你可以

30:38

discuss it at some point we're going to

找个时间讨论它，我们之后会

30:39

have agents sort of try and out all kind

让 agents 试着把所有各种

30:41

of combinations of whatever algorithmic

你能想到的 algorithmic 东西的组合都试一遍，其实我们

30:43

thing you can come up with in fact we

确实已经有一些合作了

30:44

actually do have some collaborations

实际上确实有一些合作

30:46

with some of those sort of auto ML

用上其中一些那种 auto ML

30:49

the type of labs try this out

那类实验室会试试这个

30:51

so maybe what is left sounds a bit too

所以也许剩下的听起来有点太

30:52

hard right but like boy maybe our

难了，对吧，但就像，天哪，也许我们

30:54

biggest ideas go into it is actually

最大的想法投入进去，其实

30:56

this method for mediation are jaren

这种做 mediation 的方法就是 jaren

30:58

so one thing I've been really interested in

所以有一件事我一直特别感兴趣

31:00

is the different models of funding

就是不同的融资模式

31:02

and research academia I think Germany has a

在研究学术界，我觉得 Germany 的结构跟 US 从根本上就不一样。

31:05

different structure fundamentally from

你们机构是怎么设置的？

31:07

the US how is your institution set up

那这跟基于 US 的研究模式相比，又是怎么带来一些不一样的机会的？

31:09

and like how does this give like distinct

我非常同意，这些资金对灵活性确实非常关键，但还有那一套……

31:12

opportunities versus like a US based

机会相对于像美国本地的

31:14

research model I very much agree that

研究模型我非常同意

31:16

these funding was very much matter for

这些资金对灵活性非常重要

31:19

the flexibility but also the set of

但还有一组

31:20

teams that you can set up and like the long

你可以组建的团队，还有那种漫长的

31:22

transition that you can have so

过渡，你可以拥有的，所以

31:25

um how else

呃，还有别的什么

31:27

this is not just Germany but I would say

这不只是 Germany，不过我会说

31:29

maybe Europe's largest what's called

也许是 Europe 最大的所谓

31:30

research organization in terms of

研究机构，从

31:32

funding it has a different set of

资金来看，它有一套不同的

31:35

areas that it focuses on health being one

重点领域，健康是其中之一

31:37

of the big ones but then it has sort of

在大项目当中，但它还有点儿

31:39

material information sciences physics

材料、信息科学、物理

31:41

and so on and essentially getting

等等，基本上就是拿到

31:44

federal money to put into a deep research

联邦资金，投入到一项深度研究

31:47

question I've been building up what's

问题里。我一直在搭建所谓的

31:49

called the computational health center

computational health center

31:52

at Hamilton Munich which I think we

在 Hamilton Munich，我觉得我们

31:53

has now something like maybe 45

现在大概有也许 45

31:56

group leaders professors and maybe 500

团队负责人、教授，大概有500

31:59

to 600 people now so we are I think the

到600人，所以我觉得我们是

32:01

largest academic place for healthy eye

欧洲最大的做健康眼

32:03

type of research in Europe where we have

方面研究的学术机构，我们这里有

32:05

people really working on all scales like

人真的在做各个尺度上的研究，比如

32:06

you know I'm sort of focusing on the

你知道，我算是比较专注于

32:07

cellular scale but there's people doing

cellular scale，但也有人在做

32:09

DNA foundation models there's people

DNA foundation models，还有人

32:10

sort of doing interesting things on

算是围绕着做些有意思的事情，在

32:12

protein design there's people doing like

protein design 上，有人在搞像是

32:14

you know patient cohort type of

你知道的，patient cohort 那种

32:16

representation genetics and so on

representation、genetics 等等

32:18

and now you sort of can start putting

而现在你算是可以开始把

32:19

together in one of our big visions is

它们拼在一起，我们的一个大愿景是

32:21

to intersect and to go virtual human

去做交叉结合，走向 virtual human

32:23

and so on but having said that the key

等等，但话虽如此，关键

32:25

thing is in contrast to universities

但这事儿跟大学不一样

32:26

that even though we all have affiliation

就是虽然我们都有 affiliation

32:28

to universities mostly to the technique

跟大学，主要是跟 technique university

32:30

university or the university of Munich

或者 University of Munich

32:33

the two top rank universities interesting

这两所排名最高的大学，挺有意思

32:35

in the European Union so we have this sort

在 European Union，所以我们有这种

32:38

of place where we have these affiliations

地方，我们有这些 affiliation

32:40

at universities right but we don't have

在大学，对吧，但我们没有

32:42

undergraduate teaching essentially just

本科教学基本上就是

32:43

have PhD students and postdocs but we

带博士生和博士后，但我们

32:45

interact very much with the universities

跟大学的互动非常多

32:47

in a particular also the clinics

尤其是也会跟诊所那边

32:48

because you want to have sort of the

因为你想有那种

32:49

translational thing and there

translational 的东西，而且那边

32:51

and increasing and that's maybe an

而且越来越多，这可能是一个

32:52

interesting link so this is maybe a bit

很有意思的联系，所以这可能有点

32:54

similar to NIH type of models like in

类似 US 的 NIH 那种模式

32:56

US but sort of distributed all of the

但多少是分布在全国各地，而不只是集中在某个地方

32:57

country not just sort of in one place

也许有意思的一点、也是关联在于，政府这边不断增加

33:00

and maybe the one interesting

以及 Germany 这边也出现的这种认识，正在某种程度上表明需要有

33:02

and link is that increasing the

以及关联在于增加

33:05

government and this sort of

政府以及这种

33:06

realization also in in Germany is

在德国也有这种认识

33:08

sort of drawing that there needs to be

某种程度上是在描绘需要

33:10

much stronger transfer component not just

强得多的转化环节，不只是

33:12

by going into translation to universities

通过进入大学做转化

33:15

but also into sort of value creation

还进入某种价值创造

33:17

outside of the academic realm which is

在学术领域之外，而这

33:20

obviously going so more to start up so on

显然会更多走向 start up 等等

33:22

which I think the Boston environment has

我觉得 Boston 的环境

33:24

always been one of the key examples but now

一直是一个关键例子，但现在

33:26

of course more so now also the the Bay

当然更是如此，the Bay 也是如此

33:29

area when you know recently I was

33:31

was visiting I believe maybe genetic

33:33

or colleagues and we went to a restaurant

33:35

and then one of the waiters was asking us

33:37

hey are you guys talking about AI and bio

33:39

you know I have been also interested so

33:40

like everyone's thinking about this

33:42

start up by this and so on and I kind

start up by this and so on and so I kind

33:43

of felt that's a good environment and

觉得那是个好环境，而且

33:45

this is something I think our PhDs are

这是我觉得我们的 PhDs 并不是越来越多在做的事，而我们某种程度上也在那里学习，我也觉得在

33:47

not more and more doing and we are sort of

价值创造方面，尤其是 Munich，实际上至少

33:48

learning there and I think also in terms

而且在欧洲大陆，也许是首要之地

33:50

of sort of value creation Munich in

现在是 VC capital 的，但也算是

33:52

particular is actually at least

特别是实际上至少

33:54

and continental maybe the prime place

and continental也许是首选之地

33:56

now for VC capital but also sort of

现在对于VC资本来说，但也有点

33:58

deep tech particularly bio and defense

deep tech，尤其是生物和国防

34:00

area and sort of really exploding

这个领域，可以说真的在爆发式增长

34:02

so this stuff starts to work.

所以这些东西开始奏效了

34:04

One thing the thing we're talking about

有一点，就是我们正在聊的这个事

34:05

start up I think is really fascinating

startup，我觉得真的很有意思

34:07

I've heard that Germany has had some

我听说德国有过一些

34:09

issues with their startup ecosystem

startup 生态系统方面的问题

34:11

it's it's it's hard it's difficult to

这、这、这很难，很难去……

34:13

create a startup there compared to

在那里创业相比

34:14

you know other places in Europe.

你知道欧洲其他地方。

34:16

How like what is the competitive

现在德国的创业公司

34:19

landscape for startups right now

竞争格局是怎样的，具体来说，

34:20

in Germany and specific and how does

德国怎样才能变得更有竞争力

34:23

Germany become more competitive

大概。我对此有投资兴趣，

34:25

probably. I have invested interest in this

因为我可能快要拿到我的

34:28

because I'm probably going to get my

因为我很可能会得到我的

34:30

turn in citizenship.

交出公民身份。

34:31

Welcome let me know if you come

欢迎，你要来的话告诉我

34:32

but come by Munich. So I think

不过来 Munich 吧。所以我觉得

34:35

traditionally Europe but I would say

传统上说 Europe，但我会说

34:38

definitely Germany is quite bureaucratic

Germany 肯定挺官僚的

34:40

and processes are far from as

而且流程远没有

34:42

digital as you would expect in different

你预想的那么 digital，在不同

34:44

places not that I'm saying that US is

地方，倒不是说 US 就

34:45

maybe the best example for fully

也许最好的例子就是完全

34:46

digitized processes but stuff is

数字化的流程，但事情

34:48

definitely faster. The thing is people

肯定更快了。问题是人们

34:50

have realized that now and I think the

现在已经意识到了，而且我觉得

34:52

the particular startup creation and

具体到 startup 的创建和

34:54

implementation is I think more than

实施，我觉得在 Germany 相比 Europe

34:57

competitive in Germany versus Europe

不只是有竞争力

34:58

it's actually sort of the the prime

其实算是那种最顶尖的

35:00

number and in particular like some hot

数量，尤其是像一些热门

35:02

spots burden but in particular Munich

地点，负担，但尤其是 Munich

35:03

sort of the I think is is the

算是，我觉得是，是

35:04

capital for for a high

首都，为了，为了大量

35:07

number of startups and I think even

初创公司，而且我觉得甚至

35:08

out as if you look at to European

出来，如果你看 European Union

35:09

Union there's equally as many startups

有同样多的初创公司

35:12

been created as in US. The thing is where

被创建，就像在 US 一样。问题在于哪里

35:14

Europe and particularly also Germany

欧洲，尤其是德国，

35:16

drops off is at the scaling part. So

掉队的地方是在 scaling 这一块。所以

35:18

you know early VC and also sort of

你知道，早期 VC，还有那种

35:20

not a little bit of money this is there

其实也不算少的钱，这个是有的

35:22

but then when you go into scaling

但当你进入 scaling 阶段时

35:24

these big VC's that sort of take these

那些愿意承担这些

35:26

bigger risks are I think not not

更大风险的大 VC，我觉得并没有

35:27

closely as present. So I saw this again

那么密集地存在。所以我又看到了这一点

35:30

and again that you know if you want to

而且再说，你知道，如果你想

35:31

sort of really go into the sort of

真的有点进入那种

35:34

bigger tickets sort of series A and so on

更大额的、比如 series A 之类的

35:36

the series B you people start going

series B，你们这些人就开始

35:38

more into US also for that but maybe

更多地去美国，也是为了那个，但也许

35:40

I'm sort of simplifying here but this

我在这儿有点简化了，但这

35:42

is at least my current learning

至少是我目前的理解

35:43

Bren. You are on the board or I

Bren。你是在董事会，还是我

35:46

guess part of several local startups

我猜是几家本地 startup 的一部分

35:49

right. What does it take to go from

对。那需要什么才能从

35:51

you know idea to creating a

你知道，从一个想法到创建一家

35:54

you know to creating a new startup in

你知道，到创建一家新的 startup，在

35:56

Germany. I don't know if if you heard

Germany。我不知道你有没有听说

35:57

that the European Union tries to

European Union 试图

35:59

regulate a whole bunch of stuff

监管一大堆东西

36:00

and that can be sometimes an advantage

而有时候这可能是个优势

36:03

you know it's kind of cool like you

你知道，这还挺酷的，就像你知道，现在整个欧洲手机这块是没法再那样了；你可能也听说过，现在大家的手机都得用 USBC 的东西了，这种还挺酷的

36:04

know all across Europe we cannot go

所以接下来会有些好东西，但当然了，也会有些别的情况

36:05

with the mobile phone and you might

用手机，然后你可能

36:07

have heard that everyone needs to have

听说每个人都需要拥有

36:08

USBC things now for their mobile

现在USB-C那些东西是给他们的移动设备用的

36:10

phones this kind of kind of cool stuff

手机这类挺酷的东西

36:12

so there's some some good stuff coming

所以有一些好东西要来了

36:13

but then of course there's something

但当然了，这里面有个东西

36:14

like things like the AI and so on

比如 AI 之类的东西

36:16

that make things a tad more complicated

会让事情稍微复杂一点

36:18

luckily not yet for research

好在研究方面目前还没这样

36:19

but you know for data access across

但你知道，在数据访问方面，跨

36:21

hospitals and so on there's some

医院之类的，会有一些

36:23

you know when we were talking about

你知道，我们之前聊到

36:24

about health deep tech type of stuff

健康 deep tech 这类东西的时候

36:26

right. So there's some systems

对吧。所以有一些系统

36:28

that I think fantastic in Europe and

我觉得这在 Europe 特别棒，而且

36:31

I'm thinking about Finland you know

我在想 Finland，你知道的

36:33

Finjean Denmark and so on with the

Finjean、Denmark 等等，配上那种

36:35

sort of very detailed description

算是非常详细的描述

36:37

of the population terms of sampling

关于人口的，就 sampling 而言

36:39

getting medical records connected far

把医疗记录连接起来要远远

36:40

better than I mean US is also

好得多，我是说 US 也是

36:42

similar mess like like Germany in

类似的烂摊子，就像、就像 Germany 在

36:44

terms of sort of getting all the

就是说得让所有

36:45

systems sort of to talk to each other

系统之间能互相沟通

36:46

like when patients you know sort of

就像病人，你知道，怎么说呢

36:47

go to different

去不同的

36:49

states in US is similar as in

US 各州，跟

36:51

Germany when they go to different

Germany 类似，当他们去不同的

36:52

places it's like very hard to build

地方时，就真的很难建立

36:53

patient trajectories right but there's

patient trajectories，对吧，但是有

36:55

some places that know how to do this

有些地方知道怎么做这个

36:56

and there's a thing like the

而且有这么个东西，就像

36:57

European Union tells you it's called

European Union 告诉你它叫

36:58

the European health data space

37:00

where maybe next two years these

可能未来两年，这些

37:02

things need to be sort of

东西得在某种程度上

37:04

going across boundaries you need to

跨过边界，你需要

37:06

sort of start building these

在某种程度上开始把这些建起来

37:07

health record type of

health record 这种

37:08

combination so I do think

组合，所以我确实觉得

37:10

divisions are starting to happen

分化开始出现了

37:12

but the scale is not as fast so I think

但规模没那么快，所以我觉得

37:15

what you need for a tech startup

做 tech startup，你需要什么

37:18

what I noticed is always the biggest

我注意到的一直是，最大的

37:20

thing is you need to have the best

事情是，你得有最好的

37:21

team for that and you we have

做这个的团队，而且你，我们有

37:23

professors and and PIs that really

那些真的会思考这件事的教授和 PIs，所以比如说，对于我提到的那个系，我说，伙计们，我每一个 PII，我有一群这样的人，但我期望你们至少在未来三年里搞出一个 spin off，看看那会怎么样。

37:25

think about this so for example for

想想这个，比如说

37:27

for the department that I quoted it

对于我引用的那个部门

37:28

I said guys I'm every every PII

我说各位，我每个每个PII

37:30

I have a group of them but expect

我有一组这样的项目，但我期望

37:31

you at least to have one one one

你们至少有一个一个

37:33

spin off in in the next three

在接下来三年内衍生出来

37:35

years let's see how that works

看看那会怎么样

37:37

but you sort of need to sort of get

但你多少得让

37:38

people a bit bit on that track but I

大家稍微往那个方向上靠一靠，不过我

37:40

think this is happening very much

觉得现在这其实很

37:41

from the bottom up nowadays anyway

多是自下而上发生的，反正

37:42

where the students are actually

现在是学生在真正

37:44

pushing for that they do realize

推动这件事，他们确实意识到

37:46

that there's a whole bunch of

目前有非常多

37:47

opportunities at the moment

机会

37:48

particular with with the sort of

特别是那种

37:50

computational angle

computational 角度

37:53

getting yeah I would say push

嗯，我会说，要再推

37:54

even more on the forefront that we

得更往最前沿一些，这样我们

37:56

can now start actually guiding

现在真的可以开始引导

37:57

experience we haven't been spoken

experience，这方面我们还没怎么

37:58

much about right but once you have

谈过，对吧？但一旦你有了

37:59

these perturbation world models

这些 perturbation world models

38:01

you can actually go back and drive

你其实可以回过头去推动

38:02

the experiments right so I do

实验，对，所以我确实

38:03

think computation startups can be

认为 computation startups 可以

38:05

very good also data generation

非常不错，还有 data generation

38:07

and I think this is what will be

而且我觉得这也会

38:08

also turned not just into nice

不只是变成漂亮的

38:10

papers which is sort of mostly our

论文，这差不多是我们的

38:11

business but also I think it's

主要业务，但我也觉得它

38:13

next generation of startups and you

下一代 startups，还有你

38:14

do see a few of these popping up

确实会看到其中几个冒出来

38:16

where you actually get the lab

就是你真的能拿到实验室

38:16

and it will really implement it

而且它会真的把它落地

38:18

what do you see as some of the

你觉得其中一些

38:21

the most promising areas for

最有前景的方向是什么，适合

38:23

commercialization of sort of these

把这些有点像 virtual cell 的

38:25

virtual cell adjacent ideas

Controlling Biology - Fabian Theis, Helmholtz Munich

Play Queue