# Steve Quake and Charlotte Bunne: The Holy Grail of Biology

An episode of [Ground Truths](https://invisiblenetwork.ai/podcast/1728526108), 18 Mar 2025, 44 min.

Audio: https://api.substack.com/feed/podcast/159078596/b5fa2eff1498b796c9abdff6eb453c4c.mp3

“Eventually, my dream would be to simulate a virtual cell.”—Demis Hassabis The aspiration to build the virtual cell is considered to be equivalent to a moonshot for digital biology. Recently, 42 leading life scientists published a paper in Cell on why this is so vital, and how it may ultimately be accomplished. This conversation is with 2 of the authors, Charlotte Bunne, now at EPFL and Steve Quake, a Professor at Stanford University, who heads up science at the Chan-Zuckerberg Initiative The audio (above) is available on iTunes and Spotify. The full video is linked here, at the top, and also can be found on YouTube. TRANSCRIPT WITH LINKS TO AUDIO Eric Topol (00:06): Hello, it's Eric Topol with Ground Truths and we've got a really hot topic today, the virtual cell. And what I think is extraordinarily important futuristic paper that recently appeared in the journal Cell and the first author, Charlotte Bunne from EPFL, previously at Stanford’s Computer Science. And Steve Quake, a young friend of mine for many years who heads up the Chan Zuckerberg Initiative (CZI) as well as a professor at Stanford. So welcome, Charlotte and Steve. Steve Quake (00:42): Thanks, Eric. It's great to be here. Charlotte Bunne: Thanks for having me. Eric Topol (00:45): Yeah. So you wrote this article that Charlotte, the first author, and Steve, one of the senior authors, appeared in Cell in December and it just grabbed me, “How to build the virtual cell with artificial intelligence: Priorities and opportunities.” It's the holy grail of biology. We're in this era of digital biology and as you point out in the paper, it's a convergence of what's happening in AI, which is just moving at a velocity that's just so extraordinary and what's happening in biology. So maybe we can start off by, you had some 42 authors that I assume they congregated for a conference or something or how did you get 42 people to agree to the words in this paper? Steve Quake (01:33): We did. We had a meeting at CZI to bring community members together from many different parts of the community, from computer science to bioinformatics, AI experts, biologists who don't trust any of this. We wanted to have some real contrarians in the mix as well and have them have a conversation together about is there an opportunity here? What's the shape of it? What's realistic to expect? And that was sort of the genesis of the article. Eric Topol (02:02): And Charlotte, how did you get to be drafting the paper? Charlotte Bunne (02:09): So I did my postdoc with Aviv Regev at Genentech and Jure Leskovec at CZI and Jure was part of the residency program of CZI. And so, this is how we got involved and you had also prior work with Steve on the universal cell embedding. So this is how everything got started. Eric Topol (02:29): And it's actually amazing because it's a who's who of people who work in life science, AI and digital biology and omics. I mean it's pretty darn impressive. So I thought I'd start off with a quote in the article because it kind of tells a story of where this could go. So the quote was in the paper, “AIVC (artificial intelligence virtual cell) has the potential to revolutionize the scientific process, leading to future breakthroughs in biomedical research, personalized medicine, drug discovery, cell engineering, and programmable biology.” That's a pretty big statement. So maybe we can just kind of toss that around a bit and maybe give it a little more thoughts and color as to what you were positing there. Steve Quake (03:19): Yeah, Charlotte, you want me to take the first shot at that? Okay. So Eric, it is a bold claim and we have a really bold ambition here. We view that over the course of a decade, AI is going to provide the ability to make a transformative computational tool for biology. Right now, cell biology is 90% experimental and 10% computational, roughly speaking. And you've got to do just all kinds of tedious, expensive, challenging lab work to get…

## Transcript

[0:06] SPEAKER_02: Hello,
[0:07] SPEAKER_02: it's Eric Topol with Ground Truths,
[0:08] SPEAKER_02: and we've got a really hot topic today,
[0:11] SPEAKER_02: the virtual cell,
[0:14] and what I think is an extraordinarily important futuristic
[0:17] paper that recently
[0:18] SPEAKER_02: appeared in the journal Cell,
[0:20] SPEAKER_02: and the first author,
[0:21] SPEAKER_02: Charlotte Boone from
[0:24] SPEAKER_02: EPFL, previously at Computer Science in Stanford.
[0:28] SPEAKER_02: And Steve Quake,
[0:31] SPEAKER_02: a young friend of mine for many years,
[0:34] who heads up the Chan Zuckerberg Institute as well as a
[0:37] professor at Stanford.
[0:38] SPEAKER_02: So welcome, Charlotte and Steve.
[0:42] SPEAKER_01: Thanks, Eric.
[0:42] SPEAKER_01: It's great to be here.
[0:44] SPEAKER_00: Thanks for having me.
[0:45] SPEAKER_02: Yeah,
[0:46] SPEAKER_02: so you wrote this article that Charlotte,
[0:49] SPEAKER_02: the first author,
[0:50] SPEAKER_02: and Steve,
[0:50] SPEAKER_02: one of the senior authors,
[0:53] SPEAKER_02: appeared in Cell in December.
[0:55] And it just grabbed me because how to build the virtual cell
[0:58] with artificial intelligence,
[1:00] SPEAKER_02: priorities and opportunities.
[1:03] SPEAKER_02: It's the holy grail of biology.
[1:05] SPEAKER_02: We're in this era of digital biology.
[1:07] SPEAKER_02: And as you point out in the paper,
[1:09] SPEAKER_02: it's a convergence of what's happening in AI,
[1:12] which is just moving at a velocity that's just so extraordinary
[1:16] and what's
[1:16] SPEAKER_02: happening in biology.
[1:19] So maybe we can start off by you had some 42 authors that I
[1:24] assume they congregated
[1:26] SPEAKER_02: for a conference or something.
[1:27] SPEAKER_02: Or how did you get 42 people to agree to the words in this paper?
[1:33] SPEAKER_01: We did.
[1:33] We had a meeting at CZI to bring community members together
[1:39] from many different
[1:40] SPEAKER_01: parts of the community,
[1:41] SPEAKER_01: from computer science to bioinformatics,
[1:44] SPEAKER_01: AI experts,
[1:46] SPEAKER_01: biologists.
[1:47] SPEAKER_01: We don't trust any of this.
[1:48] SPEAKER_01: We wanted to have some real contrarians in the mix as well.
[1:52] And have them have a conversation together about,
[1:55] is there an opportunity here?
[1:57] SPEAKER_01: What's the shape of it?
[1:58] SPEAKER_01: What's realistic to expect?
[2:00] SPEAKER_01: And that was sort of the genesis of the article.
[2:04] SPEAKER_02: And Gerald, how did you get to be drafting the paper?
[2:09] So I did my postdoc with Aviv Regev at Genentech and Jure
[2:13] Leskowicz at CCI.
[2:14] SPEAKER_00: And Jure was part of the residency program of CCI.
[2:17] SPEAKER_00: And so this is how we got involved.
[2:19] SPEAKER_00: And Jure had also prior work with Steve on the universal cell embedding.
[2:26] SPEAKER_00: So this is how everything got started.
[2:28] SPEAKER_02: Yeah,
[2:29] and it's actually amazing because it's a who's who of people
[2:32] who work in life science,
[2:33] SPEAKER_02: AI,
[2:34] SPEAKER_02: and digital biology,
[2:36] SPEAKER_02: and omics.
[2:37] SPEAKER_02: I mean, it's pretty darn impressive.
[2:39] So I thought I'd start off with a quote in the article because
[2:43] it kind of tells a
[2:44] SPEAKER_02: story of where this could go.
[2:47] SPEAKER_02: So the quote was in the paper,
[2:50] SPEAKER_02: AIVC,
[2:50] SPEAKER_02: that's Artificial Intelligence Virtual Cell,
[2:54] SPEAKER_02: has the potential to revolutionize the scientific process of
[2:59] SPEAKER_02: leading to future breakthroughs in biomedical research,
[3:01] SPEAKER_02: personalized medicine,
[3:02] SPEAKER_02: drug discovery,
[3:03] SPEAKER_02: cell engineering,
[3:04] SPEAKER_02: and programmable biology.
[3:06] SPEAKER_02: That's a pretty big statement.
[3:08] So maybe we can just kind of toss that around a bit and maybe
[3:12] give it a little more
[3:13] SPEAKER_02: thoughts and color as to what you were positing there.
[3:19] SPEAKER_01: Yeah.
[3:19] SPEAKER_01: Charlie, you want me to take the first shot at that?
[3:23] SPEAKER_01: Okay.
[3:24] So Eric,
[3:26] it is a bold claim and,
[3:27] you know,
[3:28] we have a really bold ambition here.
[3:30] We view that over the course of a decade,
[3:33] AI is going to provide the ability to
[3:35] SPEAKER_01: make a transformative computational tool for biology.
[3:39] Right now,
[3:39] cell biology is 90% experimental and 10% computational,
[3:44] roughly speaking.
[3:46] SPEAKER_01: And you've got to do just,
[3:47] SPEAKER_01: you know,
[3:48] SPEAKER_01: all kinds of tedious,
[3:49] SPEAKER_01: expensive,
[3:50] SPEAKER_01: challenging lab work to get to the answer.
[3:52] And I don't think AI is going to replace that,
[3:55] but it can invert the ratio.
[3:57] So within 10 years,
[3:58] I think we can get to biology being 90% computational and 10% experimental.
[4:03] SPEAKER_01: And the goal of the virtual cell is to build a tool that'll do that.
[4:09] SPEAKER_02: Yeah,
[4:09] and I think a lot of people may not understand why it is
[4:12] considered the Holy Grail,
[4:13] SPEAKER_02: because it is the fundamental unit of life,
[4:17] SPEAKER_02: and it's incredibly complex.
[4:20] It's not just all the things happening in the cell with atoms
[4:24] and molecules and
[4:26] SPEAKER_02: organelles and everything inside,
[4:28] SPEAKER_02: but then there's also the interactions to other cells on the outside.
[4:34] SPEAKER_02: tissue and world.
[4:35] So,
[4:35] I mean,
[4:35] it's really quite extraordinary challenge that you've taken on here.
[4:39] And I guess there's some debate,
[4:44] of course,
[4:45] as to do we have the right foundation?
[4:49] SPEAKER_02: We're getting into foundation models in a second.
[4:53] SPEAKER_02: A good friend of mine and part of this whole,
[4:56] SPEAKER_02: I think,
[4:58] SPEAKER_02: process that you got together,
[4:59] SPEAKER_02: Aaron Siegel from Israel,
[5:01] SPEAKER_02: he said,
[5:01] SPEAKER_02: we're at this tipping point.
[5:04] SPEAKER_02: All the stars are lying and we have all the different components,
[5:06] SPEAKER_02: the data,
[5:08] SPEAKER_02: the compute,
[5:08] SPEAKER_02: the modeling.
[5:10] SPEAKER_02: And in the paper,
[5:10] you describe how we have over the last couple of decades have
[5:14] so many different
[5:15] SPEAKER_02: data sets that are rich,
[5:16] SPEAKER_02: that are global initiatives.
[5:19] SPEAKER_02: But then there's also questions.
[5:21] SPEAKER_02: Do we really have the data?
[5:22] SPEAKER_02: I think Bo Wang especially asked about that.
[5:26] SPEAKER_02: Maybe Charlotte, what are your thoughts about data deficiency?
[5:30] SPEAKER_02: There's a lot of data, but do you really have what we need before we...
[5:35] bring them all together for this kind of single model that will
[5:38] get us someday to
[5:39] SPEAKER_02: the virtual cell.
[5:41] SPEAKER_00: So I think,
[5:41] SPEAKER_00: I mean,
[5:41] SPEAKER_00: one core idea of building this AIVC is that we basically can leverage all
[5:46] SPEAKER_00: experimental data that is overall collected.
[5:48] SPEAKER_00: So this also goes back to the point Steve just made.
[5:51] So meaning that we basically can integrate across many
[5:54] different studies data
[5:56] because we have AI algorithms or the architectures that power
[6:00] such an AIVC are able
[6:02] SPEAKER_00: to integrate basically data sets on many different scales.
[6:05] SPEAKER_00: So we are going a bit away from this dogma.
[6:08] SPEAKER_00: I'm designing one algorithm from one data set to this idea of I have an
[6:13] SPEAKER_00: architecture that can take in multiple data set on multiple scales.
[6:17] So this will help us a bit in like being somewhat efficient
[6:21] with the type of
[6:21] experiments that we need to make and the type of experiments we
[6:24] need to conduct.
[6:25] SPEAKER_00: And again, what Steve just said, right?
[6:27] SPEAKER_00: Like ultimately we can very much steer which experiments we need to even
[6:31] SPEAKER_00: like experiments, like data sets we need to collect.
[6:34] SPEAKER_00: Currently, of course, we don't have all the data that is sufficient.
[6:38] I mean,
[6:38] in particular,
[6:39] I think most of the tissues we have,
[6:41] they are healthy tissues,
[6:42] right?
[6:43] SPEAKER_00: We don't have all the disease phenotypes that we would like to measure.
[6:46] SPEAKER_00: Having patient data is always a very tricky case.
[6:49] SPEAKER_00: mostly non-interventional data,
[6:52] meaning we have very limited understanding of somehow the
[6:55] effect of different perturbations,
[6:58] perturbations that happen on many different scales in many
[7:01] different environments.
[7:03] SPEAKER_00: So we need to collect a lot here.
[7:05] SPEAKER_00: I think the overall...
[7:07] journey that we are going with is that we take the data that we
[7:11] have we make clever
[7:13] decisions on the data that we will collect in the future and we
[7:16] have this also
[7:17] self-improving entity that is aware of what it doesn't know
[7:21] right so we need to be
[7:22] able to understand like how well can I predict something in
[7:26] this on this somewhat
[7:27] regime if I cannot then we should focus our data collection
[7:31] effort into this so I
[7:34] SPEAKER_00: think that's
[7:35] Of course,
[7:36] not present state,
[7:37] but this will basically also guide the future collection.
[7:41] SPEAKER_02: Speaking of data,
[7:42] one of the things I think that's fascinating is we saw how
[7:46] AlphaFold II really
[7:48] SPEAKER_02: revolutionized predicting proteins.
[7:52] SPEAKER_02: But remember,
[7:53] SPEAKER_02: that was based on this extraordinary resource that had been built,
[7:56] SPEAKER_02: the Protein Data Bank,
[7:58] SPEAKER_02: that enabled that.
[7:59] SPEAKER_02: And for the virtual cell, there's no such thing as a Protein Data Bank.
[8:04] SPEAKER_02: It's so much more,
[8:06] SPEAKER_02: as you emphasize,
[8:06] SPEAKER_02: Charlotte,
[8:07] SPEAKER_02: it's so much dynamic and these perturbations that are just,
[8:11] SPEAKER_02: you know,
[8:11] SPEAKER_02: all across the board,
[8:12] SPEAKER_02: as you emphasize that.
[8:15] SPEAKER_02: Now,
[8:16] SPEAKER_02: the Human Cell Atlas,
[8:18] SPEAKER_02: which is currently some tens of millions,
[8:22] SPEAKER_02: but going into a billion cells,
[8:24] SPEAKER_02: we learned that it used to be 200 cell types.
[8:27] SPEAKER_02: Now, I guess it's well over 5,000.
[8:29] And that we have 37 trillion cells approximately in the average
[8:34] person's adult's body.
[8:37] SPEAKER_02: is a formidable map that's being made now.
[8:42] SPEAKER_02: And I guess the idea that you're advancing is that we used to,
[8:47] SPEAKER_02: and this goes back to a statement you made earlier,
[8:49] SPEAKER_02: Steve,
[8:50] SPEAKER_02: everything we did in science was hypothesis driven.
[8:54] SPEAKER_02: But if we could get this computational model of the virtual cell,
[8:58] SPEAKER_02: then we can have AI exploration of the whole field.
[9:03] SPEAKER_02: Is that really the nuts of this?
[9:06] SPEAKER_01: Yes.
[9:07] SPEAKER_01: And, you know, a couple of thoughts on that, maybe.
[9:12] SPEAKER_01: You know,
[9:14] SPEAKER_01: Theo Karalezzos,
[9:15] SPEAKER_01: our lead AI person at CZI says,
[9:18] machine learning is the formalism through which we understand
[9:21] high dimensional data.
[9:23] SPEAKER_01: And I think that's a very deep statement.
[9:25] And,
[9:26] you know,
[9:26] biological systems are intrinsically very high dimensional.
[9:29] SPEAKER_01: You've got 20,000 genes in the human genome in these cell atlases.
[9:33] SPEAKER_01: You're measuring all of them at the same time in each single cell.
[9:37] And there's a lot of structure in the relationships of their
[9:39] gene expression there
[9:40] SPEAKER_01: that is just not evident to the human eye.
[9:43] SPEAKER_01: And for example,
[9:45] SPEAKER_01: a cell by gene,
[9:46] SPEAKER_01: our database that collects all the aggregates,
[9:49] SPEAKER_01: all of the single cell transcriptomic data is now over 100 million cells.
[9:55] SPEAKER_01: And as you mentioned,
[9:57] we're seeing ways to increase that by an order of magnitude in
[10:00] the near future.
[10:02] The project that Yuri Leskovich and I worked on together that
[10:05] Charlotte referenced
[10:06] earlier was like a first attempt to build a foundational model
[10:10] on that data to
[10:11] SPEAKER_01: discover some of the correlations and structure that was there.
[10:15] SPEAKER_01: And so with a subset,
[10:16] SPEAKER_01: I think it was the 20 or 30 million cells,
[10:19] SPEAKER_01: we built a large language model.
[10:22] And began asking it,
[10:25] what do you understand about the structure of this data?
[10:27] SPEAKER_01: And it kind of discovered lineage relationships without us teaching it.
[10:31] You know,
[10:31] we trained on a matrix of numbers,
[10:33] no biological information there.
[10:35] And it learned a lot about the relationships between cell type
[10:38] and lineage.
[10:39] And that emerged from that high dimensional structure,
[10:41] which was super pleasing to us.
[10:44] SPEAKER_01: And really,
[10:45] SPEAKER_01: I mean,
[10:45] SPEAKER_01: for me personally,
[10:47] SPEAKER_01: gave me the confidence to say,
[10:49] SPEAKER_01: this stuff is going to work out.
[10:50] SPEAKER_01: There is a future for the virtual cell.
[10:53] SPEAKER_01: It's not some made up thing.
[10:55] SPEAKER_01: There is real substance there.
[10:57] And this is worth investing an enormous amount of CZI resources
[11:00] in going forward
[11:01] SPEAKER_01: and trying to rally the community around as a project.
[11:04] Well,
[11:04] yeah,
[11:04] I mean,
[11:05] the premise here is that there is a language of life.
[11:09] SPEAKER_02: And you just made a good case that there is.
[11:12] SPEAKER_02: If you can predict,
[11:15] SPEAKER_02: if you can query,
[11:16] SPEAKER_02: if you can generate like that,
[11:19] SPEAKER_02: it's reminiscent of the famous Go game of Lee Soto,
[11:24] SPEAKER_02: the world champion player.
[11:26] SPEAKER_02: and how the machine came up with a move many,
[11:30] SPEAKER_02: many years ago that no human would have anticipated.
[11:32] And I think that's what you're getting at and the ability for
[11:36] inference and reason
[11:37] SPEAKER_02: now to add to this.
[11:39] SPEAKER_02: So,
[11:41] SPEAKER_02: you know,
[11:41] SPEAKER_02: Charlotte,
[11:41] SPEAKER_02: one of the things,
[11:42] SPEAKER_02: of course,
[11:44] SPEAKER_02: is about,
[11:45] SPEAKER_02: well,
[11:46] there's two terms in here that are unfamiliar to many of the
[11:50] listeners or viewers
[11:52] SPEAKER_02: of this podcast.
[11:55] SPEAKER_02: Universal representations and virtual instruments that you make a pretty
[12:04] SPEAKER_02: significant part of how you are going about this virtual cell model.
[12:09] SPEAKER_02: So could you describe that and also the embeddings as part of the UR,
[12:16] SPEAKER_02: the universal representation?
[12:18] Because I think embeddings or these meaningful relationships
[12:21] are key to what Steve
[12:22] SPEAKER_02: was just talking about.
[12:25] SPEAKER_00: Yes.
[12:25] SPEAKER_00: So in order to somewhat leverage like very different modalities,
[12:29] in order to leverage basically modalities that we take
[12:32] measurements across
[12:33] SPEAKER_00: different scales,
[12:34] SPEAKER_00: like the idea is that we have large,
[12:37] SPEAKER_00: maybe transformer models that might be very different.
[12:40] SPEAKER_00: If I have imaging data, I have a vision transformer.
[12:43] SPEAKER_00: If I have a text data,
[12:44] SPEAKER_00: I have large language models that are designed,
[12:46] SPEAKER_00: of course,
[12:47] SPEAKER_00: for DNA,
[12:47] SPEAKER_00: then they have a very wide context and so on and so forth.
[12:50] SPEAKER_00: But so the idea is somewhat that we have a
[12:53] SPEAKER_00: models that are connected through the scales of biology,
[12:57] SPEAKER_00: because those scales we know,
[12:58] we know which components are somewhat involved or in
[13:01] measurements that are
[13:02] SPEAKER_00: happening upstream.
[13:03] So we have the somewhat interconnection of very large model
[13:07] that will be trained on
[13:08] SPEAKER_00: many different data.
[13:09] And we have this internal model representation that somewhat
[13:12] capture everything they've seen.
[13:15] And so this is what we call those universal representation that
[13:18] will exist across
[13:20] SPEAKER_00: the scales of biology.
[13:22] SPEAKER_00: And what is great about AI,
[13:23] SPEAKER_00: and so I think this is a bit like a history of AI in short,
[13:27] SPEAKER_00: is like the ability to predict,
[13:28] SPEAKER_00: right?
[13:29] SPEAKER_00: The last years, the ability to...
[13:32] generate right we can generate new hypothesis we can generate
[13:35] modalities that we
[13:36] are missing we can potentially generate certain cellular state
[13:40] molecular state and
[13:41] have a certain certain property but i think what's really
[13:45] coming is this ability to
[13:46] reason so we see this in those very large language models the
[13:49] ability to reason
[13:50] SPEAKER_00: about hypothesis how we can test it and so this is
[13:55] SPEAKER_00: what those instruments ultimately need to do, right?
[13:57] So we need to be able to simulate the change of a perturbation
[14:02] on a cellular phenotype,
[14:03] SPEAKER_00: so on the internal representation,
[14:05] SPEAKER_00: the universal representation of a cell state.
[14:07] We need to simulate the effect the mutation has downstream and
[14:10] how this would
[14:11] SPEAKER_00: propagate in our representations upstream.
[14:14] And we need to build many different types of virtual
[14:17] instruments that allow us to basically
[14:20] design and build all those capabilities that ultimately the AI
[14:24] virtual cell needs
[14:25] SPEAKER_00: to possess that will then allow us to reason,
[14:27] SPEAKER_00: to generate hypothesis,
[14:29] SPEAKER_00: to basically predict the next experiment to conduct,
[14:32] SPEAKER_00: to predict the outcome of a perturbation experiment,
[14:35] SPEAKER_00: to in silico design cellular states,
[14:37] SPEAKER_00: molecular states,
[14:38] SPEAKER_00: things like that.
[14:39] SPEAKER_00: So,
[14:39] and this is why we make the separation between like internal
[14:42] representation as well
[14:44] SPEAKER_00: as those instruments that operate on those representations.
[14:47] SPEAKER_02: Yeah,
[14:47] that's what I really liked is that you basically described the architecture,
[14:50] SPEAKER_02: how you're going to do this by putting these URs into the VIs,
[14:56] SPEAKER_02: having a decoder and a manipulator.
[14:58] SPEAKER_02: And you basically got the idea.
[15:00] If you can bring all these different integrations about,
[15:04] which,
[15:05] of course,
[15:06] is pending.
[15:07] Now,
[15:08] there are obviously many naysayers here that this is impossible
[15:12] to do.
[15:13] SPEAKER_02: One of them is this guy, Philip Ball.
[15:16] SPEAKER_02: I don't know if you read the language, How Life Works.
[15:20] SPEAKER_02: Now, he's a science journalist, and he writes, he's a prolific writer.
[15:27] He says,
[15:27] comparing life to a machine,
[15:28] a robot,
[15:29] a computer,
[15:30] sells it short.
[15:32] uh life is a cascade of processes each with a distinct integral
[15:36] integrity and
[15:36] autonomy the logic of which has no parallel outside the living
[15:41] world um is he right
[15:44] there's no way to bottle this it's silly it's too complex you
[15:49] you know we don't
[15:51] know all right and you know it's great that there's naysayers
[15:55] if everyone agreed
[15:56] SPEAKER_01: this was doable
[15:57] SPEAKER_01: Would it be worth doing?
[15:58] SPEAKER_01: I mean,
[15:59] the whole point is to take risks and get out and do something
[16:01] really challenging in
[16:02] SPEAKER_01: the frontier where you don't know the answer.
[16:05] SPEAKER_01: If we knew that it was doable, I wouldn't be interested in doing it.
[16:09] SPEAKER_01: So I personally am happy that there's not a consensus that...
[16:15] SPEAKER_02: Well,
[16:16] SPEAKER_02: I mean,
[16:17] SPEAKER_02: to capture people's imagination here,
[16:20] SPEAKER_02: if you're successful and you marshal a global effort,
[16:23] I don't know who's going to pay for it because it's a lot of
[16:25] work coming here going forward.
[16:28] SPEAKER_02: But if you can do it,
[16:29] SPEAKER_02: the question here is,
[16:30] SPEAKER_02: you know,
[16:31] SPEAKER_02: right today we talk about,
[16:32] SPEAKER_02: oh,
[16:32] SPEAKER_02: let's make an organoid.
[16:34] So we can figure out how to treat this person's cancer or
[16:37] understand this person's
[16:38] SPEAKER_02: rare disease or whatever.
[16:40] And instead of having to wait weeks for this culture and all
[16:43] the expense and whatnot,
[16:44] SPEAKER_02: you could just do it in a computer and silico.
[16:47] And you have this virtual twin of a person's cells and their
[16:51] tissue and whatnot.
[16:52] SPEAKER_02: So the opportunity here is,
[16:54] SPEAKER_02: I don't know if people get this,
[16:56] SPEAKER_02: is just extraordinary and quick and cheap if you can get there.
[17:02] SPEAKER_02: I guess, and it's such a bold initiative idea.
[17:06] SPEAKER_02: Who will pay for this, you think?
[17:08] SPEAKER_01: Well, CZI is putting an enormous amount of resources into it.
[17:12] SPEAKER_01: And it's a major project for us.
[17:16] SPEAKER_01: We have been laying the groundwork for it.
[17:19] SPEAKER_01: We recently put together what I think is,
[17:22] SPEAKER_01: if not the largest,
[17:23] one of the largest GPU supercomputer clusters for nonprofit
[17:27] basic science research.
[17:29] SPEAKER_01: That came online at the end of last year.
[17:31] SPEAKER_01: And in fact, in December, we put out an RFA
[17:35] SPEAKER_01: for the scientific community to propose using it to build models.
[17:39] SPEAKER_01: And so we're sharing that resource within the scientific community.
[17:43] SPEAKER_01: As I think you appreciate,
[17:45] one of the real challenges in the field has been access to
[17:47] compute resources and
[17:49] SPEAKER_01: industry has it,
[17:50] SPEAKER_01: academia at a much lower level.
[17:52] SPEAKER_01: We're able to be somewhere in between,
[17:54] SPEAKER_01: not quite at the level of a private company,
[17:55] SPEAKER_01: but the tech company,
[17:57] SPEAKER_01: but at a level beyond what most universities are able to do.
[18:00] SPEAKER_01: And we're trying to use that to drive the field forward.
[18:02] We're also planning on launching RFAs this year to help drive
[18:07] this project forward
[18:08] SPEAKER_01: and funding people globally on that.
[18:10] And we're building a substantial internal effort within CZI to
[18:14] help drive this project forward.
[18:16] SPEAKER_02: You know,
[18:17] SPEAKER_02: I think it has the looks of like the Human Genome Project,
[18:21] SPEAKER_02: which at the time,
[18:22] SPEAKER_02: as you know,
[18:23] SPEAKER_02: when it was originally launched,
[18:25] SPEAKER_02: people thought,
[18:25] SPEAKER_02: oh,
[18:25] SPEAKER_02: this is impossible.
[18:27] SPEAKER_02: And look what happened.
[18:28] SPEAKER_02: It got done.
[18:29] SPEAKER_02: And now,
[18:30] SPEAKER_02: you know,
[18:30] SPEAKER_02: the sequence of genome is just a,
[18:32] SPEAKER_02: you know,
[18:32] SPEAKER_02: commodity,
[18:33] SPEAKER_02: very relatively very inexpensive.
[18:37] SPEAKER_01: I think a lot about those parallels.
[18:39] SPEAKER_01: And I will say one thing, Philip Ball, I will concede him the point.
[18:44] SPEAKER_01: The cells are very complicated.
[18:46] SPEAKER_01: The genome project.
[18:47] I mean,
[18:47] the sort of genius there was to turn it from a biology problem
[18:50] into a chemistry problem.
[18:51] There is a test tube with a chemical in it,
[18:53] work out the structure of that chemical.
[18:55] SPEAKER_01: And if you can do that, the problem is solved.
[18:57] I think what it means to have the virtual cell is much more
[19:01] complex and ambiguous
[19:02] SPEAKER_01: in terms of defining what it's going to do and when you're done.
[19:07] SPEAKER_01: And so we have our work cut out for us there to try to do that.
[19:11] SPEAKER_01: And that's why a little bit,
[19:14] I established our North Star and CZI for the next decade as
[19:17] understanding the
[19:18] SPEAKER_01: mysteries of the cell.
[19:19] SPEAKER_01: And that word mystery is like very important to me.
[19:22] You know,
[19:22] I think the molecules,
[19:23] as you pointed out earlier,
[19:24] are understood.
[19:25] SPEAKER_01: Genome sequenced, protein structure, solved or predicted.
[19:29] SPEAKER_01: We know a lot about the molecules.
[19:31] SPEAKER_01: Those are, if not solved problems, pretty close to being solved.
[19:35] And the real mystery is how do they work together to create
[19:37] life in the cell?
[19:39] SPEAKER_01: And that's what we're trying to answer with this virtual cell project.
[19:43] SPEAKER_02: Yeah,
[19:43] SPEAKER_02: I think another thing that,
[19:44] SPEAKER_02: of course,
[19:45] SPEAKER_02: is happening concurrently to add the,
[19:49] SPEAKER_02: I think,
[19:50] likelihood that you'll be successful is we've never seen the
[19:54] foundation models
[19:55] SPEAKER_02: coming out in life science as they have in recent weeks,
[19:59] SPEAKER_02: months,
[19:59] SPEAKER_02: never.
[19:59] SPEAKER_02: I mean,
[20:00] SPEAKER_02: I have a paper in science this tomorrow coming out,
[20:05] SPEAKER_02: summarizing the progress about not just RNA,
[20:09] SPEAKER_02: DNA,
[20:10] uh ligands i mean the the whole idea alpha fold three but now
[20:15] bolts and so many
[20:17] others it's just amazing how fast the torrent of new foundation
[20:20] models so charlotte
[20:22] what do you think accounts for this um this this is
[20:26] unprecedented in life science
[20:29] to see foundation models coming out of this clip uh on
[20:32] evolution uh on i mean you
[20:35] SPEAKER_02: name it design uh of every different
[20:39] SPEAKER_02: molecule of life or of course in cells included in that.
[20:44] SPEAKER_02: What do you think is going on here?
[20:47] SPEAKER_00: So,
[20:47] SPEAKER_00: I mean,
[20:47] SPEAKER_00: on the one hand,
[20:47] SPEAKER_00: of course,
[20:48] SPEAKER_00: we benefit,
[20:49] SPEAKER_00: profit,
[20:49] and inherit from all the tremendous efforts that have been made
[20:53] in the last decades
[20:54] SPEAKER_00: on assembling those data sets that are very,
[20:57] SPEAKER_00: very standardized,
[20:57] SPEAKER_00: right?
[20:58] CellX gene is like very,
[21:00] like somehow AI friendly,
[21:01] as you can say it,
[21:02] right?
[21:02] SPEAKER_00: Like it is somewhat a platform that is easy to feed into algorithms.
[21:06] SPEAKER_00: But at the same time, we actually also see really new...
[21:10] SPEAKER_00: building mechanisms, design principles of AI algorithms in itself.
[21:15] SPEAKER_00: So I think we have understood that in order to really make progress,
[21:18] SPEAKER_00: build those systems that work well,
[21:21] SPEAKER_00: we need to build AI tools that are designed for biological data.
[21:25] SPEAKER_00: So to give you...
[21:27] SPEAKER_00: an easy example, right?
[21:28] SPEAKER_00: Like if I use a large language model on text,
[21:31] it's not going to work out of the box for DNA because we have
[21:35] different reading directions,
[21:36] SPEAKER_00: different context lens,
[21:38] SPEAKER_00: and many,
[21:38] SPEAKER_00: many,
[21:38] SPEAKER_00: many,
[21:39] SPEAKER_00: many more.
[21:40] SPEAKER_00: And if I look at standard computer vision,
[21:43] SPEAKER_00: right,
[21:44] where we can say AI really excels and I'm applying standard
[21:48] computer vision,
[21:49] SPEAKER_00: vision transformers on multiplex images,
[21:52] SPEAKER_00: They are not going to work because normal computer vision architectures,
[21:57] SPEAKER_00: they always expect the same three inputs,
[21:59] SPEAKER_00: RGB,
[22:00] SPEAKER_00: right?
[22:00] SPEAKER_00: In multiplex images,
[22:02] SPEAKER_00: I'm measuring up to 150 proteins potentially in a single experiment,
[22:07] SPEAKER_00: but every study will measure different proteins.
[22:09] SPEAKER_00: So I deal with many different scales, right?
[22:12] SPEAKER_00: Like larger scales than I used to.
[22:15] attention mechanisms that we have in usual computer vision
[22:18] transformers are not
[22:19] SPEAKER_00: going to work anymore.
[22:20] SPEAKER_00: They are not going to scale.
[22:22] SPEAKER_00: And at the same time,
[22:22] I need to be completely flexible in whatever input combination
[22:26] of channel I'm just
[22:27] SPEAKER_00: going to face in this experiment.
[22:29] SPEAKER_00: So I mean, so this is what
[22:32] SPEAKER_00: we right now did,
[22:33] SPEAKER_00: for example,
[22:33] SPEAKER_00: in our very first work,
[22:35] SPEAKER_00: inheriting the design principle that we laid out in the paper,
[22:38] SPEAKER_00: AI virtual cell,
[22:40] and then come up with new AI architectures that are dealing
[22:43] with these very special
[22:44] SPEAKER_00: requirements that biological data have.
[22:47] SPEAKER_00: We have now a lot of computer scientists that work very,
[22:51] SPEAKER_00: very closely,
[22:51] SPEAKER_00: have a very good understanding of biologists,
[22:53] SPEAKER_00: biologists that are getting much and much more into the computer science.
[22:58] So people who are fluent in both languages somewhat that are
[23:02] able to now build
[23:03] SPEAKER_00: models that are adapted and designed for biological data.
[23:06] And we don't just take basically computer vision architectures
[23:09] that work well on
[23:10] SPEAKER_00: street scenes and try to apply them on biological data.
[23:13] SPEAKER_00: So it's just a...
[23:15] SPEAKER_00: Very different way of thinking about it,
[23:17] SPEAKER_00: starting,
[23:18] SPEAKER_00: constructing basically specialized architectures.
[23:21] Besides,
[23:21] of course,
[23:22] the tremendous data efforts that have happened in the past.
[23:24] SPEAKER_02: Yeah,
[23:25] SPEAKER_02: I mean,
[23:25] SPEAKER_02: it's pretty,
[23:25] and we're not even talking about just sequence because we've
[23:28] also got imaging,
[23:30] SPEAKER_02: which has gone to a revolution,
[23:32] be able to image subcellular without having to use any types of
[23:36] stains or think
[23:37] SPEAKER_02: that would disrupt cells.
[23:38] SPEAKER_02: That's another part of kind of the deep learning era that came along.
[23:43] SPEAKER_02: One thing I thought was fascinating in the paper in Cell,
[23:46] SPEAKER_02: you wrote,
[23:48] SPEAKER_02: for instance,
[23:48] the short read archive of biological sequence data holds over
[23:52] 14 petabytes of information,
[23:54] which is a thousand times larger than the data set used to
[23:57] train chat GPT.
[23:59] SPEAKER_02: I mean, that's a lot of tokens.
[24:02] SPEAKER_02: That's a lot of stuff.
[24:06] SPEAKER_02: Compute resources.
[24:07] SPEAKER_02: It's almost like you're going to need a deep seek information
[24:11] SPEAKER_02: type of way to get this,
[24:13] SPEAKER_02: I mean,
[24:14] SPEAKER_02: not that DeepSeek is as its claim to be so much more economical,
[24:19] but there's a data challenge here in terms of working with that
[24:24] massive amount that
[24:25] SPEAKER_02: is different than the language of a human language,
[24:30] SPEAKER_02: that is,
[24:30] SPEAKER_02: our language,
[24:31] SPEAKER_02: wouldn't you say?
[24:32] SPEAKER_02: Yeah.
[24:34] SPEAKER_01: So,
[24:35] SPEAKER_01: you know,
[24:35] SPEAKER_01: Eric,
[24:35] SPEAKER_01: it brings to mind one of my favorite quotes from Sidney Brenner,
[24:38] SPEAKER_01: who was such a wit.
[24:40] SPEAKER_01: And,
[24:41] SPEAKER_01: you know,
[24:41] SPEAKER_01: in 2000,
[24:43] SPEAKER_01: at the sort of early,
[24:45] SPEAKER_01: you know,
[24:46] SPEAKER_01: first flush of success in genomics,
[24:48] SPEAKER_01: he said...
[24:50] uh biology is drowning in a sea of data and starving for
[24:53] knowledge um yeah a very
[24:56] deep statement right and there's a little bit what that the
[24:59] motivation was for
[25:00] putting the short read archive statistic into the paper there
[25:03] um and you know again
[25:05] for me um part of the value of this endeavor of creating a
[25:09] virtual cell is it's a
[25:11] SPEAKER_01: tool to help us translate data into knowledge
[25:15] SPEAKER_02: Well, there's two, I think, phenomenal figures in your cell paper.
[25:20] SPEAKER_02: The first one that takes across the capabilities of the virtual cell,
[25:25] and the second that compares the virtual cell to the real or
[25:28] the physical cell.
[25:29] SPEAKER_02: And we'll link that with this in the transcript.
[25:32] SPEAKER_02: And the other thing we'll link is that there's a nice Atlantic article.
[25:36] SPEAKER_02: A virtual cell is a...
[25:38] SPEAKER_02: is a holy grail of science.
[25:40] SPEAKER_02: It's getting closer.
[25:41] That may not be quite close as like next week or year,
[25:45] but it's getting close.
[25:46] And that's good for people who are not well grounded in this
[25:49] because it's much more
[25:52] SPEAKER_02: taken out of the technical realm.
[25:55] SPEAKER_02: This is really exciting, what you're onto here.
[26:00] SPEAKER_02: And what's interesting,
[26:03] SPEAKER_02: Steve,
[26:03] SPEAKER_02: since I've known you for so many years,
[26:05] SPEAKER_02: earlier in your career,
[26:06] SPEAKER_02: you really worked on omics,
[26:08] SPEAKER_02: that is being DNA and RNA.
[26:11] SPEAKER_02: And in recent times, you've made this switch to cells.
[26:16] SPEAKER_02: Is that just because you're trying to anticipate the field?
[26:20] SPEAKER_02: Tell us a little bit about your migration.
[26:23] SPEAKER_01: Yeah,
[26:23] SPEAKER_01: I mean,
[26:24] so a big part of my career has been trying to develop new
[26:28] measurement technologies
[26:29] SPEAKER_01: that will provide insight into biology.
[26:31] SPEAKER_01: And, you know, decades ago, that was understanding molecules.
[26:34] SPEAKER_01: Now it's understanding more complex biological things like cells.
[26:38] SPEAKER_01: And it was like a natural progression.
[26:41] SPEAKER_01: I mean, we built the sequencers.
[26:44] SPEAKER_01: sequence the genomes done, right?
[26:46] And it was clear that people were just going to do that at
[26:48] scale then and,
[26:49] SPEAKER_01: you know,
[26:51] SPEAKER_01: create lots of data.
[26:52] SPEAKER_01: Hopefully knowledge would get out of that.
[26:53] SPEAKER_01: But for me as an academic,
[26:55] SPEAKER_01: I never wanted to,
[26:56] SPEAKER_01: never thought I'd be in the position I'm in now,
[26:58] SPEAKER_01: let's put it that way.
[26:59] SPEAKER_01: I just wanted to keep running a small research group.
[27:01] SPEAKER_01: So I realized
[27:02] SPEAKER_01: I would have to get out of the genome thing and find the next frontier.
[27:06] SPEAKER_01: And it became this intersection of microfluidics and genomics,
[27:10] SPEAKER_01: which is,
[27:10] SPEAKER_01: you know,
[27:10] I spent a lot of time developing microfluidic tools to analyze
[27:13] cells and try to do
[27:15] SPEAKER_01: single cell biology to understand their heterogeneity.
[27:19] And that through a winding path led me to all these cell
[27:22] atlases and to where we are now.
[27:26] SPEAKER_02: Well, we're fortunate for that.
[27:27] SPEAKER_02: And also with your work with CGI to help propel that forward and
[27:32] I think it sounds like we're going to need a lot of help to get
[27:34] this thing done.
[27:35] SPEAKER_02: Now,
[27:35] SPEAKER_02: Charlotte,
[27:36] SPEAKER_02: as a computer scientist now at EPFL,
[27:40] SPEAKER_02: what are you going to do to keep working on this?
[27:44] And what's your career advice for people in computer science
[27:48] who have an interest
[27:48] SPEAKER_02: in digital biology?
[27:51] So I work in particular on the prospect of using this to build
[27:56] diagnostic tools and
[27:57] SPEAKER_00: to make diagnostics in the clinic easier,
[28:00] because ultimately we have somewhat limited capabilities in the
[28:04] hospital to run
[28:05] SPEAKER_00: like deep omics.
[28:06] SPEAKER_00: But the idea of being able to somewhat map with a cheaper and lighter
[28:12] modality or like somehow diagnostic test into something much
[28:17] richer because a model
[28:18] has been seeing all those different data and can basically
[28:21] contextualize it.
[28:22] SPEAKER_00: It's very interesting, right?
[28:24] SPEAKER_00: We've seen all those pathology foundation models.
[28:26] If I can always run an H&E but then decide when to run deeper
[28:30] diagnostics to have a
[28:31] SPEAKER_00: better or more accurate
[28:33] SPEAKER_00: prediction that is very powerful and ultimately like reducing the costs,
[28:37] SPEAKER_00: but the precision that we have in hospitals.
[28:39] So my faculty position right now is co-located between the
[28:43] School of Life Sciences,
[28:44] SPEAKER_00: School of Computer Science.
[28:46] So I have a dual affiliation and I'm affiliated to the
[28:48] hospitals to actually make this possible.
[28:51] and as a career advice i think don't be shy and stick to your
[28:56] discipline i think i
[28:57] have a bachelor's in biology but i never only did biology i
[29:01] have a phd in computer
[29:02] science which kind of you would think a bachelor in biology not
[29:06] necessarily
[29:06] qualifies you through so i think this interdisciplinarity also
[29:09] requires you to be
[29:10] very fluent very comfortable in reading many different styles
[29:15] of papers and
[29:15] SPEAKER_00: publications because of the course
[29:17] SPEAKER_00: publication in a computer science venue will be very,
[29:20] SPEAKER_00: very different from the way we write in biology.
[29:23] SPEAKER_00: So don't stick to your study program,
[29:26] but just be free in selecting whatever course gets you closer
[29:29] to the knowledge you
[29:30] need in order to do the research or whatever task you're
[29:35] building and working on.
[29:39] SPEAKER_02: Well, you know, Charlotte, the way you're set up there,
[29:43] with this coalescence of life science and computer science is
[29:48] so ideal and so
[29:49] SPEAKER_02: unusual here in the U.S.
[29:51] SPEAKER_02: So that's fantastic.
[29:54] SPEAKER_02: And that's kind of what we need.
[29:55] And that's really the underpinning of how you're going to get
[29:58] to the virtual cells,
[30:00] SPEAKER_02: getting these two communities together.
[30:01] SPEAKER_02: And Steve,
[30:02] SPEAKER_02: likewise,
[30:03] you were an engineer and somehow you became one of the pioneers
[30:08] of digital biology
[30:09] SPEAKER_02: way back before you had that term.
[30:12] SPEAKER_02: This interdisciplinary,
[30:14] SPEAKER_02: transdisciplinary,
[30:16] SPEAKER_02: we need so much of that in order for you all to be successful,
[30:18] SPEAKER_02: right?
[30:20] SPEAKER_01: Absolutely.
[30:21] I mean,
[30:21] there's so much great discovery to be done on the boundary
[30:25] between fields.
[30:26] SPEAKER_01: I trained as a physicist and,
[30:28] SPEAKER_01: you know,
[30:28] kind of made my career this boundary between physics and
[30:31] biology and technology development.
[30:35] SPEAKER_01: And it's just sort of been the gift that keeps on giving.
[30:39] SPEAKER_01: You know, there's a new way to measure something.
[30:42] You discover something new scientifically,
[30:44] and it just all suggests new things to measure.
[30:48] SPEAKER_01: It's very self-reinforcing.
[30:50] SPEAKER_02: Now,
[30:52] a couple of people who you know well have made some pretty big
[30:55] statements about
[30:57] SPEAKER_02: this whole era of digital biology.
[30:59] SPEAKER_02: And I think,
[31:00] SPEAKER_02: you know,
[31:01] the virtual cell is perhaps the biggest initiative of all the
[31:07] digital biology
[31:08] SPEAKER_02: ongoing efforts.
[31:09] SPEAKER_02: But Jensen Wong wrote,
[31:10] SPEAKER_02: the first time in human history,
[31:12] SPEAKER_02: biology has the opportunity to be engineering,
[31:14] SPEAKER_02: not science.
[31:16] SPEAKER_02: And Demis Hassabis wrote,
[31:19] SPEAKER_02: said, we're seeing engineering science.
[31:21] SPEAKER_02: You have to build the artifact of interest first,
[31:24] then use the scientific method to reduce it down and understand
[31:28] its components.
[31:29] SPEAKER_02: Well, here, there's a lot to do to understand its components.
[31:34] SPEAKER_02: And if we can do that,
[31:35] SPEAKER_02: like,
[31:35] SPEAKER_02: for example,
[31:36] SPEAKER_02: right now,
[31:37] SPEAKER_02: as both of you know,
[31:38] SPEAKER_02: AI drug discoveries and high gear,
[31:40] SPEAKER_02: and there's umpteen numbers of companies working on it.
[31:44] SPEAKER_02: But it doesn't account for the self.
[31:47] SPEAKER_02: I mean, it basically is protein, protein-ligand interactions.
[31:52] SPEAKER_02: What if we had drug discovery that was cell-based?
[31:57] SPEAKER_02: Could you comment about that?
[31:58] SPEAKER_02: Because that doesn't even exist right now.
[32:02] SPEAKER_01: Yeah, I mean, I can say something first.
[32:05] SPEAKER_01: Charlotte, if you've got thoughts, I'm curious to hear them.
[32:09] So I do think AI approaches are going to be very useful
[32:12] designing molecules.
[32:14] SPEAKER_01: And so from the perspective of designing new therapeutics,
[32:17] SPEAKER_01: whether they're small molecules or antibodies,
[32:19] SPEAKER_01: yeah,
[32:20] SPEAKER_01: I mean,
[32:20] SPEAKER_01: there's a ton of investment in that area.
[32:22] That is like a near-term fruit,
[32:23] perfect thing for venture people to invest in.
[32:28] SPEAKER_01: And there's opportunity there.
[32:30] SPEAKER_01: There's been enough proof of principle.
[32:32] um however um i i do agree with you that if you want to really
[32:38] understand what
[32:39] happens when you drug a target you're going to want to have
[32:42] some model of the cell
[32:44] um and maybe not just the cell but all the different cell types
[32:48] of the body to
[32:49] understand where toxicity will come from if you have on target
[32:52] toxicity and whether
[32:53] SPEAKER_01: you get efficacy on the thing you're trying to do and so
[32:56] We really hope that people will use the virtual cell models
[33:00] we're going to build as
[33:01] SPEAKER_01: part of the drug discovery development process.
[33:03] And we would think it's going to,
[33:05] you know,
[33:05] I agree with you in a little of a blind spot.
[33:08] SPEAKER_01: And we think if we make something useful, people will be using it.
[33:11] The other thing I'll say on that point is I'm very enthusiastic
[33:15] about the future of
[33:16] SPEAKER_01: cellular therapies.
[33:17] SPEAKER_01: And one of our big bets at CZI has been starting the New York Biohub,
[33:22] SPEAKER_01: which is aimed at,
[33:23] SPEAKER_01: you know,
[33:24] really being very ambitious about establishing the engineering
[33:27] and scientific
[33:27] foundations of how to engineer completely radically more
[33:32] powerful cellular therapies.
[33:34] SPEAKER_01: And the virtual cell is going to help them do that.
[33:36] SPEAKER_01: Right.
[33:36] SPEAKER_01: It's going to be essential for them to achieve that mission.
[33:40] I think you're pointing out one of the most important things
[33:42] going on in medicine
[33:43] today is how we didn't anticipate that live cell therapy of
[33:49] engineered cells and
[33:50] SPEAKER_02: ideally off the shelf or in vivo,
[33:53] not just having to take them out and work on them outside the
[33:57] body is a revolution ongoing.
[33:59] SPEAKER_02: And it's not just in cancer, it's in autoimmune diseases and many others.
[34:03] SPEAKER_02: So it's part of the virtual cell need.
[34:06] SPEAKER_02: We need this.
[34:08] SPEAKER_02: One of the things that's a misnomer I want you both to comment on,
[34:12] SPEAKER_02: we keep talking about single cell,
[34:13] SPEAKER_02: single cell.
[34:15] SPEAKER_02: And,
[34:15] SPEAKER_02: you know,
[34:16] SPEAKER_02: there's a paper spatial multiomics this week,
[34:19] SPEAKER_02: five different single cell,
[34:21] SPEAKER_02: you know,
[34:23] SPEAKER_02: scales,
[34:24] SPEAKER_02: all integrated.
[34:25] SPEAKER_02: It's great.
[34:25] SPEAKER_02: But we don't get to single cell.
[34:27] SPEAKER_02: We're basically looking at 50 cells, 100 cells.
[34:30] SPEAKER_02: We're not doing single cell because we're not going deep enough.
[34:34] SPEAKER_02: Is that just a matter of time when we actually are doing it?
[34:36] SPEAKER_02: Of course,
[34:37] SPEAKER_02: the more we do get down to the single or a few cells,
[34:40] SPEAKER_02: the more insights we're going to get.
[34:42] SPEAKER_02: Would you comment about that?
[34:43] SPEAKER_02: Because,
[34:44] SPEAKER_02: you know,
[34:45] SPEAKER_02: we have all this literature on single cell comes out every day,
[34:48] SPEAKER_02: but we're not really there yet.
[34:53] Charlotte,
[34:53] you want to take a first pass at that and then I can say something?
[34:56] SPEAKER_00: Yes, so...
[35:00] SPEAKER_00: So it depends, right?
[35:00] So I think if we look at certain spatial proteomics,
[35:03] we still have subcellular resolution.
[35:05] SPEAKER_00: So of course,
[35:06] SPEAKER_00: we always measure many different cells,
[35:07] but we are able to somewhat get down to a resolution where we
[35:10] can look at certain
[35:12] SPEAKER_00: co-localization of proteins.
[35:14] SPEAKER_00: This also goes back to the point just made before, right?
[35:17] SPEAKER_00: Like having this very good environment to study like drugs, right?
[35:21] SPEAKER_00: Like if I want to build a new drug,
[35:23] SPEAKER_00: if I want to build a new protein,
[35:25] SPEAKER_00: the idea of building this multi-scale model,
[35:27] SPEAKER_00: right,
[35:27] SPEAKER_00: allows us to actually
[35:29] simulate like different like somehow binding changes and
[35:32] binding because we
[35:33] simulate the effect of a drug ultimately the readers we have
[35:36] they are subcellular
[35:37] so of course we we often like in the spatial biology we often
[35:41] have a bit like
[35:42] methods that are rather coarse they have a spot that averages
[35:45] over certain like
[35:47] SPEAKER_00: some cells, right, like hundreds of cells or a few cells.
[35:50] But I think we also have more and more technologies that are
[35:54] zooming in that are
[35:55] subcellular where we can actually tag or have like those probe
[35:58] based methods that
[36:00] SPEAKER_00: allow us to zoom in.
[36:01] SPEAKER_00: There's microscopy of individual cells to really capture them in 3D.
[36:06] SPEAKER_00: They are, of course, not very high throughput yet.
[36:09] but it gives us also an idea of the morphology and how
[36:13] ultimately morphology
[36:15] determines certain like somehow cellular properties or cellular phenotype.
[36:18] So I think there's lots of progress also on the experimental
[36:23] and that ultimately
[36:24] SPEAKER_00: will back feed into the
[36:27] SPEAKER_00: AI virtual cell, those models will be fed by those data.
[36:30] SPEAKER_00: Similarly,
[36:31] SPEAKER_00: looking at dynamics,
[36:32] SPEAKER_00: right,
[36:33] SPEAKER_00: looking at live imaging of individual cells,
[36:36] SPEAKER_00: of their morphological changes.
[36:38] SPEAKER_00: Also,
[36:38] this ultimately is data that we will need to get a better
[36:41] understanding of disease mechanisms,
[36:43] SPEAKER_00: cellular phenotypes,
[36:44] SPEAKER_00: functions,
[36:46] SPEAKER_00: perturbation responses.
[36:47] SPEAKER_02: Right.
[36:48] SPEAKER_02: Yeah,
[36:49] SPEAKER_02: Steve,
[36:50] let me comment on that and the amazing progress that we have
[36:53] made with space and time,
[36:55] SPEAKER_02: spatial temporal resolution,
[36:57] SPEAKER_02: spatial omics over these years,
[36:59] but that we still could go deeper in terms of getting to
[37:03] individual cells,
[37:04] SPEAKER_02: right?
[37:06] SPEAKER_01: So, what can we do with a single cell?
[37:11] I'd say we are very mature in our ability to amplify and
[37:15] sequence the genome of a
[37:16] SPEAKER_01: single cell,
[37:18] SPEAKER_01: amplify and sequence the transcriptome of a single cell,
[37:22] SPEAKER_01: So you can ask, is one cell enough to make a biological conclusion?
[37:27] And maybe I think what you're referring to is people want to
[37:29] see replicates.
[37:30] And so you can ask how many cells do you need to see to have
[37:32] confidence in any
[37:33] SPEAKER_01: given biological conclusion,
[37:35] SPEAKER_01: which is a reasonable thing.
[37:36] SPEAKER_01: It's a statistical question in good science.
[37:39] um i think uh i've been very impressed with how the mass spec
[37:45] people have been
[37:45] doing recently i think they've finally cracked the ability to
[37:49] look at proteins from
[37:49] single cells and they can look at a couple thousand proteins i
[37:53] mean that that was
[37:54] you know i think one of these nature method of the year things
[37:56] at the end of last
[37:57] SPEAKER_01: year and the visual proteomics yes yeah they are over the heart now um
[38:03] SPEAKER_01: Yeah, they are over the hump with single cell measurements.
[38:08] SPEAKER_01: Part of what's missing right now,
[38:10] SPEAKER_01: I think,
[38:11] SPEAKER_01: is the ability to reliably do all of that on the same cell.
[38:14] SPEAKER_01: So this is what Charlotte was referring to,
[38:16] SPEAKER_01: be able to do sort of multi-molded measurements on single cells.
[38:18] SPEAKER_01: That's kind of in its infancy.
[38:20] And there's a few examples,
[38:21] but there's a lot more work to be done on that.
[38:24] And I think also the fact that these measurements are all
[38:26] destructive right now.
[38:28] SPEAKER_01: And so you're losing the ability to look how the cells evolve over time.
[38:31] SPEAKER_01: You've got to say this time point,
[38:33] I'm going to dissect this thing and look at a state and I don't
[38:36] get to see what
[38:36] SPEAKER_01: happens further down the road.
[38:38] SPEAKER_01: So that's another future, I think, measurement challenge to be addressed.
[38:42] SPEAKER_02: Yeah,
[38:42] SPEAKER_02: and I think,
[38:43] SPEAKER_02: you know,
[38:43] SPEAKER_02: I'm just trying to identify some of the multitude of challenges in this
[38:49] extraordinarily bold initiative because there are no shortages
[38:53] and that's what's
[38:53] SPEAKER_02: good about it.
[38:55] It's given people,
[38:56] you know,
[38:57] lots of work to do to overcome,
[39:00] override some of these challenges.
[39:03] SPEAKER_02: Before we wrap up,
[39:04] besides the fact that you point out that all the work has to be
[39:07] done and be
[39:08] SPEAKER_02: validated in real experiments,
[39:10] SPEAKER_02: not just live in a virtual AI world,
[39:12] but you also comment about the safety and ethics of this work
[39:18] and assuming you're
[39:19] SPEAKER_02: going to gradually get there and be successful.
[39:21] SPEAKER_02: So could either or both of you comment about that?
[39:25] Because it's very thoughtful that you're that you're thinking
[39:27] already about that.
[39:30] SPEAKER_01: as scientists and members of the sort of larger community,
[39:33] we want to be careful and ensure that we're interacting with
[39:40] people who set policy
[39:41] in a way that ensures that these tools are being used to
[39:44] advance the cause of
[39:45] science and not do things that are detrimental to human health
[39:47] and are used in a
[39:49] SPEAKER_01: way that respects patient privacy.
[39:51] And so the ethics around how you use all this with respect to
[39:56] individuals is going
[39:58] SPEAKER_01: to be important to be thoughtful about from the beginning.
[40:00] And I also think there's an ethical question around what it
[40:05] means to be publishing papers.
[40:07] You don't want people to be forging papers using data from the
[40:11] virtual cell without
[40:12] being clear about where that came from and pretending that it
[40:15] was a real experiment.
[40:16] So there's issues around those sorts of ethics as well that
[40:18] need to be considered.
[40:21] SPEAKER_02: And of those 40-some authors around the world,
[40:24] SPEAKER_02: do you have the sense that you all work together to achieve this
[40:30] SPEAKER_02: Is there kind of a global bonding here that's going to collaborate?
[40:36] SPEAKER_01: I think this effort is going to go way beyond those 40 authors.
[40:39] SPEAKER_01: It's going to include a much larger set of people.
[40:41] SPEAKER_01: And I'm really excited to see that evolve with time.
[40:45] SPEAKER_02: Yeah, no, it's really quite extraordinary how you kick this thing off.
[40:49] SPEAKER_02: And the paper is,
[40:50] SPEAKER_02: you know,
[40:51] the blueprint for something that we all can anticipate that
[40:55] could change a lot of
[40:57] SPEAKER_02: science and medicine.
[40:58] I mean,
[40:58] we saw,
[40:59] as you mentioned,
[40:59] Steve,
[41:00] how that deep visual proteomics saved lives.
[41:03] SPEAKER_02: It was what I wrote as spatial medicine, no longer spatial biology.
[41:07] And so the way that this can change the future of medicine,
[41:11] I think a lot of people...
[41:12] just have to have a little bit of imagination that once we get
[41:15] there with this AIVC,
[41:17] SPEAKER_02: that there's a lot in store that's really quite exciting.
[41:21] SPEAKER_02: Well,
[41:22] I think this has been an invigorating review of that paper and
[41:27] some of the issues
[41:29] SPEAKER_02: surrounding it.
[41:30] I couldn't be more enthusiastic for your success and ultimately
[41:37] where this could take us.
[41:39] Did I miss anything during the discussion that we should touch
[41:42] on before we wrap up?
[41:45] SPEAKER_01: Not from my perspective.
[41:46] SPEAKER_01: It was a pleasure as always, Eric, and a fun discussion.
[41:51] SPEAKER_02: Yeah, thanks so much.
[41:52] SPEAKER_02: Well, thank you both and all of the co-authors of this paper.
[41:56] SPEAKER_02: We're going to be following this with with great interest.
[41:59] And I think for most people listening,
[42:02] they may not know that this is in store for the future.
[42:06] SPEAKER_02: Someday we will get there.
[42:08] I think one of the things to point out right now is the models
[42:11] we have today that
[42:12] SPEAKER_02: largely the large language models based on transformer architecture
[42:18] SPEAKER_02: they're going to continue to evolve.
[42:19] We're already seeing so much in inference and ability for
[42:23] reasoning to be exploited
[42:27] SPEAKER_02: and not asking for prompts with immediate answers,
[42:30] but waiting for days to get back a lot more work from a lot
[42:34] more computing resources.
[42:37] SPEAKER_02: But we're going to get models in the future to fold this together.
[42:41] I think that's one of the things that you've touched on in the
[42:44] paper so that
[42:45] SPEAKER_02: whatever we have today in concert with what you've laid out,
[42:50] SPEAKER_02: AI is just going to keep getting better.
[42:52] The biology that these foundation models are going to get
[42:56] broader and more
[42:58] SPEAKER_02: compelling as to their use cases.
[43:01] SPEAKER_02: That's why I believe in this.
[43:04] SPEAKER_02: I don't see this as a static situation right now.
[43:07] I just think that you're anticipating the future and we will
[43:10] have better models to
[43:11] be able to integrate this massive amount of what some people
[43:16] would consider
[43:16] SPEAKER_02: disparate data sources.
[43:18] SPEAKER_02: So thank you both and all your colleagues for writing this paper.
[43:23] I don't know how you got the 42 authors to agree to it all,
[43:27] which is great.
[43:28] SPEAKER_02: And it's just the beginning of something that's a new frontier.
[43:31] SPEAKER_02: Thanks very much.
[43:33] SPEAKER_01: Thank you, Eric.

Shows are listed from Apple Podcasts' charts and directory, and Podcast Index where it is used. Every show and episode belongs to its publisher and plays from the publisher's own feed; this site hosts none of them.