Controlling Biology - Fabian Theis, Helmholtz Munich

Fabian Theis has been one of the key players in the various iterations of the Virtual Cell concept and initiatives. It is not hyperbole at all to say that his lab has been pushing the frontier of algorithms to understand multi-scale, multi-modal (”mult...

• 0:00Speaker 00

  We can really now start guiding the experimentation. So where before you we built this human cell atlas. We had sort of this mapping aspect. Then we talked about the word model prediction problem from mapping. We went to prediction. But now we can go to actually steering the next experiment. But quite frankly, very often the papers end it with some model, right? Maybe some sort of concept figure, maybe add a little bit validation. But then to be you start the next project, right? And of course you learn a bit, but there's not like a continual learning type of thing. And you think which we call lab compass, which is an inverse generative model that actually allows you, once you have a forward word model or whatever, to tell you what you need to do in order to reach a certain state. Obviously, the ultimate vision being that we want to RL everything. We want to do RL also in this loop. It's a bit out there, but you know, I think we're slowly getting that.

• 0:41Speaker 01

  Welcome to the Latent Space AI for Science Podcast. I'm RJ Hanke, CTO of Miralomics. I'm joined by my co-host Brandon Anderson, who develops RNA therapeutics at Atomic AI. I can't emphasize enough how much Fabian Teas and his labs and companies have contributed to the development of systems biology and various iterations of the virtual cell concept. Papers and models come out of his lab faster than I can read them. And the lab is consistently moving the ball forward on our understanding of how to observe and model the behaviors and the interactions of cells. Just as importantly, his work often focuses on creating the tools that researchers and scientists actually use in their day to day work.

• 1:25Speaker 01

  Most computational biologists have probably used Scampy and N data at some point. And this work has led to the creation of SC verse, which is an amazing collection of open tools that has attracted many, many contributors. Of course, Fabian is one of those academics who has way more honors than you can possibly list. But suffice it to say that he leads a bunch of the AI and computational biology efforts at Helmholtz Munich and related institutions and participates also in many other industry efforts. Fabian, it's really great to meet you. Thanks a lot for joining us.

• 2:00Speaker 00

  Thanks for having me. I'm really looking forward to the discussion, man.

• 2:02Speaker 01

  Yeah, I've been tracking your work for a long time. And so it's a great honor to finally get you on the show and talk to you. Let's kick things off with some of the hype happening right now. There's a ton of interest right now in structural biology. Anthropic and OpenAI biopharma initiatives focus heavily on designing molecules right now. On the other hand, virtual cell and systems biology approaches really haven't gotten as much attention recently. Maybe Zara and Noetic are good counter examples there, but Biohub just announced a $1.8 billion collaboration with the DOE, NIH, Meta and Google to collect an open source, multi-scale biological data that can be used to develop a new generation of virtual cells.

• 2:46Speaker 01

  So my question, do you think that we're about to see renewed interest from the pharma industry in virtual cells based on the activity that we're seeing in academia right now?

• 2:56Speaker 00

  I should disclose before I sort of go to the answer that obviously my opinion is quite academic in the sense that whenever I talk to my friends and colleagues that really sort of day by day sit in the startups, they have like, like what I sort of have as an overview and sort of a bit of intuition, what the different research labs do, they have the same thing for the startups. And then this is sort of the hot big thing that, you know, this and this investments are happening. I don't really have that on that scale. But having said that, for example, you mentioned Biohub's initiative, which by the way is also heavily coupled to Human Cell Atlas, which is sort of this international, really worldwide consortium that sort of pushes this next level of dataset generation. I do think that there's a whole bunch of reasons why we're gonna see that sort of scale up

• 3:41Speaker 00

  from all these models working well in terms of being generative, but also sort of lab in the loop on the molecule level now being exported and sort of really working one level up on the cell level. And I think there's a bunch of reasons for that. And the first one, I think, and maybe the most important one, which you really hit here is the data generation. I think we had that concept of a virtual self for a while. Like my lab, we started this perturbation prediction modeling and using machine learning in 2019 with a paper we called SEGen at the time, which was like, it was a nonlinear model, but essentially it's sort of in a latent space, did some linear transformation and it was actually quite okay. And even nowadays, many of these sort of pseudo linear type of things are still doing a good job.

• 4:21Speaker 00

  But the reason for that being that I think the data sets and the questions that we asked on top were just sort of still on the simplistic side. That's why also sort of it didn't yet take up. So whenever people talk about virtual cells, maybe also a question, I think, what you actually ask is how does, can you predict how a cell would behave under all kind of perturbations? And that's sort of, I think, a quite deep question. If you really think about it, that's maybe even a bit of a holy grail in computational biology, systems biology, sort of this, this input output system that is a cell with memory and like with all the specifications and with all the different sort of levels of molecules and so on, not just RNA. I think being able to map this out, what does even understanding mean?

• 5:02Speaker 00

  And I've been sort of a bit of a meta problem, but I think you understand at least your modeler system very well when you sort of predict its output behavior given different inputs. So I think it's a very good question. And this is why often people nowadays equate a virtual cell with perturbation prediction. I think it's a bit too simplistic, but I think this perturbation prediction type of models and that's sort of being maybe what you mean by virtual cell is taken off a lot. And the startup landscape, I think, is a bit exploding there. We talked, I think, a lot about US, Europe. But also in the Chinese ecosystem, there's a lot of these additional efforts now happening. And I hope that these are not just sort of now a kindling of a flame because everyone sort of adds AI to their things.

• 5:42Speaker 00

  But because pharma is really desperately needing to be cell type context specific in their screens, which at the moment they can't do yet on that scale.

• 5:51Speaker 01

  Maybe you could expound a little bit on what does this, so perturbations right now, like we talked to Zara recently. They're doing these perturbation screens. What does that unlock? And then what do you think the next generation of data is going to unlock in addition to that?

• 6:07Speaker 00

  I mean, for me, if I may stop me when I'm sort of getting too much into these old stories. Like, you know, I'm sort of, I've been trained machine learning scientist for longest time, doing all kinds of things. But you know, I started with single cell early on, QPCR. Maybe initially, if I'm really honest, not so much because it was like the biology that I was interested in, but because we had high sample numbers purely for that, right? Because even early on, it was sort of clear that the number of QPCRs at scale, but then of course transcriptomes was just getting really big. And you know, when I started my career early on, we had like a few bug RNA. Now we sort of had thousands of them or whatever, right? So this is like early times. But now, you know, we go into a regime.

• 6:50Speaker 00

  And that was sort of one of these eye opening moments when, when Parz Bio had, I think they call it the Giga Lab or something like that. And you know, there's sort of similar ones from competing companies now. But when they had like in a single experiment, 10 million transcriptomes measured, right? I thought that was like next level stuff. And you know, this step from just getting more PB and C's as a control tool, like let's actually do something at scale to it, sort of having here the cell space, but then having here the perturbation space, which typically was very scarce. It was like control versus disease or so. So this sort of number of points, what you had here in terms of sampling on the perturbation space was typically much, much more scarce than the number of observations in terms of cells that you had. But now this is exploding. Now you sort of have, I don't know, 100 different cell lines where you do sort of full perturb seek, like full crisp out.

• 7:34Speaker 00

  Or you have like batteries of 1000 drugs or so added. And suddenly this space gets really rich, right? And then you can sort of start doing all these fancy perturbation prediction type of things. And this is currently happening. It's crazy what scale is happening. We are ourselves involved in a whole bunch of things where we actually go beyond cell lines and do organoid perturbations. And maybe if we have time, we can talk a bit more about that. Because I'm quite convinced that we need to go to multi cellular systems to have actually really interesting results. But I think everyone is scaling this, these levels up. And at some point, you know, we're gonna go maybe even to human genetics as a perturbation.

• 8:07Speaker 01

  Can you just like explain what is a perturbation? What do you mean by that? Just for the people who are not familiar.

• 8:13Speaker 02

  Yeah, yeah. Actually, would you mind maybe backing up even one step more than that? Like what do we even mean by like measuring transcriptomics? And you were talking about like PCR based and then also perturb seek and.

• 8:25Speaker 00

  When I keep sort of public talks, I typically say, you know, human genome project, we sort of have the genome for the body, but it's pretty much the same everywhere, right? So like, why do we have all these different cell types, right? What's sort of happening specifically? And also first level, what we typically have, one of the first levels is that, you know, per cell, different stuff is being read off of the genome, right? And these messenger RNAs, which then sort of encode the various downstream proteins. While this is sort of differently expressed, as one says, in different cells. And to say in a human, maybe it's just something like 20,000 different types of these, we're splicing many, many more, but that's sort of the rough level. So what you de facto have is per cell, a 20,000 dimensional vector

• 9:04Speaker 00

  that has a whole bunch of zeros and stuff is not there, but then a whole bunch of sort of numbers, how many of these counts of these messenger RNAs you have? Because these things are called transcripts, right? This is called a transcriptome. Yeah. And de facto, then you have this sort of 20,000 dimensional space that we very often encode in a sort of nicer latent space that sort of has like a sub manifold, like sort of this subspace where these actual cell cloud lives, right? And if the cell cloud then sort of like this sort of 20,000 dimensional vectors, each point then being one cell, right? If they sort of cluster up, then we call these things cell types or something like that. And if they sort of go into a sausage or whatever, right? We call it then trajectories, like a cell differentiation, like the fertilized XL then making all the cell types in the body or something like.

• 9:47Speaker 02

  So like DNA is sort of like a recipe book, or maybe it's a bunch of genes or a bunch of ingredients. And then as you, each cell is its own sort of meal or dish or something, which is a combination of different ingredients. And now the transcriptomics is sort of figuring out. Taking different meals and then sort of decomposing them like, oh, what ingredients went in to make this meal? So if you set a dumb analogy.

• 10:12Speaker 00

  I have a bunch, bunch of foodies as kids. I will use this metaphor a bit more often. I like it.

• 10:17Speaker 02

  Thanks, Brendan.

• 10:18Speaker 01

  I love it. And then, and then perturbations is when you somehow perturb the, the genome or the transcriptome so that it has different transcripts in it. And then typically that's done with a CRISPR knockout where you remove a gene from, from the genome and that causes it to not express that gene, obviously. And then from there, then you see what are the phenotypic or the actual changes in the physical biology that occur because of a perturbation? Is that more or less right?

• 10:49Speaker 00

  Maybe I can take, take a step back there because I think, RJ, you're, you're fully right, but, but this is sort of one of the hardcore type of perturbations, right? Like fully knocking out a gene, sort of in evolution happens a bit, but mostly these things are sort of a tad more subtle. But of course, in order to understand the system, it's a good idea to take out bits and pieces. I think what happens sort of on, on a normal development or normal sort of us being in homeostasis is of course that these, these cells sort of talk to each other all the time, you know, they're sort of living these complex multicellular states, right? So there's all type of communications happen between cells. So they always have stuff coming from the outside and then sort of doing something to them. So for me, a perturbation space, in fact, is just an input space to stuff that happens to a cell, just maybe because it just sits to another one and there's like some signaling happening, right? Or some gradient, some hormones coming in, whatever, right?

• 11:34Speaker 00

  And all this stuff, of course, you cannot sort of tune up or tune down. Maybe even something as basic as when you go, these individual systems where pharmas, drug screens, when you sort of put them in a dish, you know, you can do different, what's called media conditions and they have so much impact, right? And it's like a bit of, for me, still like a bio-driven, maybe even like an alchemy, what type of cocktails people come up with to then sort of maybe take, you know, skin cells and turn them into, into more or less embryo-like state cells, right? Where you can then sort of make new cells to heal this different stuff in the body or these, what are called yamanaka factors, for example, you know, these were Nobel laureate type of stuff, right? So I think having a space of perturbations, essentially of things that you can do to

• 12:18Speaker 00

  a cell, this can be a very rich thing. It can go on as much as actually changes in, in, in the gene, which would be like a full knockout or something more subtle, right? Which is sort of human genetics. And I think all of that, we need to put in a big space of stuff that we can do that has an influence of cell state and, and like a virtual cell model should take all of it, right? And also combinations thereof, of course, and maybe your disease happening, time, whatever, right? That then impacts that.

• 12:40Speaker 01

  That's a great connection to the, the organoids idea. So can you talk a little bit about, okay, so an organoid is, is what? And, and how does that connect to this concept of perturbation space? And what do you do to it in addition to adjust this context?

• 12:58Speaker 00

  Dude, when I, when I first heard about, learned about this concept, I should say, you know, I'm, I'm another trained biologist, so I typically learn about things when I read a bit or when I talk to colleagues and, and friends. And when I first saw that people are able to take essentially cells out of a body, a cancer cell or whatever, right? And let them then grow into small multicellular systems that actually start connecting, do things. I was super, super impressed. So, you know, people can now, for example, take some cells, put them in a dish and then grow little mini brains, essentially set of neurons that essentially really start connecting, talking to each other. And people have shown that they can get those sort of mini brains in a dish to compute. Similarly, people can take heart cells, grow them into little heart organoids. And we have a bunch of local collaborations with fantastic partners that actually start

• 13:43Speaker 00

  beating, like you have beating little heart cell clumps in a dish, right? And then of course, once you have these, and your same thing is, I spoke about brain, there's even simpler systems like the colon where this works very well. There's a whole bunch of cancer systems, but you know, also for, let's say, if you want to study diseases with hard to acquire material, let's say you want to look into the genetics of some type of, of neurodegenerative diseases or so, typically it's very hard to get access to tissue, right? So you take stem cells from the skin, you take cells from the skin, dilute them, turn them into these middle brains and then you can potentially study some, some things there. So I think this concept of replicating little bits of organs, that's why they're called organoids, in a dish is very fascinating. There's all kinds of limitations, obviously.

• 14:25Speaker 00

  Often immune cells, immune system is missing, there's like no, maybe feeding them or something like, so there's all kinds of problems, right? But it's much, much more than just a cell language, it's like one cell type.

• 14:34Speaker 01

  Moving to the modeling side, how does, how do you model that?

• 14:37Speaker 00

  So that's, I think, but the research frontier, so I don't think people have agreed on what the optimal way to model is, particularly also because there's not as much data as one wants. Jay, right, you're, you're involved, if I understand correctly, in a whole bunch of spatialomics slash also sort of H and E, sort of pathology type of effort. So like pathology, there's like so much samples available, right? And now increasingly we sort of do a bit more spatialomics. So I think how we even represent spatially connected data sets. And then you sort of break it down on an organoid level and so on. People are not as clear as, as it is with disassociated cells. So again, taking, taking this a bit simpler, I think if we look at the space of disassociated

• 15:22Speaker 00

  cells and we look at these groups, we have this beautiful metaphor of what we humans call a cell type, which you know, nature doesn't care about, but this is just us putting categories on top of stuff. But it helps us understand this complexity. I don't think we have a similar metaphor for local niches of cells, maybe a little bit, but also in H and E, right? Whenever I, I talk and see what pathologies do, it's often very qualitative and it's maybe not at that level yet. So I think we need to develop a bit of a categorization language for us. Maybe computationally, maybe we don't need this so much, but computationally we also need to sort of think a bit how we deal with this. At the moment, we essentially deal with this as distributions of, of, of cells. So this is sort of the solution at the moment, but I think it can be much more subtle, like graphs or something like that.

• 16:04Speaker 01

  So on the one hand, you have the bitter lesson and, you know, sort of you, you remove inductive bias and you end up being able to have a larger data set that's applicable to your problem. And then that helps you to get more accurate models because your, your distribution is wider and whatever. So that's like the one, and then, but then there's this like countervailing tendency, which is like, maybe you could call it the anti-bitter lesson or like the other, the flip side of the bitter lesson where when you're in data poor regimes, for one thing, then you have to induce, introduce a lot of inductive bias to get enough statistical power to make good inferences. But there's also a second side of that, which I think John Platt, in our discussion with

• 16:48Speaker 01

  him not too long ago, brought up was that you have these, these like predictive and then descriptive models. And what you really want as a scientist is to get to a simplification, a descriptive model where you can say, oh, well, I've tested all the apples and the gravity works in such a way on apples. And then, but I'm modeling this as a universal concept with simply so that we know that it works on planets as well. Right? And, and so that I want to understand from you, what are the simplifications or, you know, sort of maybe simplifications, maybe insights, maybe actual mathematical or otherwise scientific models that are becoming apparent, that are going to help to reduce all the

• 17:32Speaker 01

  complexity so that we're able to actually make meaningful, gain meaningful understanding from these really large data sets?

• 17:39Speaker 00

  Yeah, that's, that's a, that's a very important, very, I think, kind of deep point, which I think sort of breaks the field into maybe sort of a bit two different groups of people with answers, right? So how much do we need to put inductive biases, maybe priors into a model to maybe make it better explainable, right? But potentially also more robust in going out of distribution? Or how much are we going to just rely on bigger data to sort of have the model be more expressive and so on? And I think, and there, I, I need to be very clear before we may go into the philosophy.

• 18:19Speaker 00

  I think literature has really clearly shown over, I think now decades, at, at least also very many data sets, that if we build in a priori inductive biases too strongly, that we might lose predictive power, right? So it comes at a cost. So and if you're now, let's say you're a physician and you just want to, or maybe even if you're a patient and you, you, maybe you might have cancer or not, maybe you want to have the best model as a predictor. Maybe if you're getting treated, you want to have a model that is maybe a bit less good, but you understand it a bit more. But for the prediction, at least I think you, you want to have the best. And this is, I think, I think not entirely clear. And I think in our field with this perturbation space scale, the cell scale, I think is big, but perturbation space scale, like the stuff that we do to it not being as big, I don't

• 19:04Speaker 00

  think we're yet in this super data rich regime. So I think in our case, these hybrid models might be still the thing to go for, but more and more of course, and this is, I think this, I don't know if you can call it sad, but the honest learning from, from big scale machine learning is that at some point data beats human made additional concepts in these models. And I, I think this one simple example I often put is AlphaFold, where you know, the first models sort of made this extraordinary effort to, to make, you know, the molecules rotation invariant, they sort of build it into the learning process and only in later versions, you know, sort of rotation symmetry was automatically found by the model just with more data and was a bit more flexible because whatever physical constraint you put in, it's not always as rigid and then you sort of, you gain something if you let the model learn from this.

• 19:46Speaker 00

  So I think in the future, the bigger part, at least of the sort of more used models will be those more flexible, very much database ones, where you then a posteriori put the interpretation on top.

• 19:59Speaker 01

  This also brings up the, the question for me. So you actually, I think recently wrote a, or co-authored a paper on the statistical power that the current generation of virtual cells have in terms of the data, the, the lift that you get from additional data in those current generation of virtual cell models. And I think you had said basically that the objective function or the training methodology is potentially the problem there and not necessarily that the information is not in the data. So like, what, can you explain that a little bit and why you think that maybe there's plenty of untapped information in the information? And then like what, where does that get us and what, what do we need to do to get further?

• 20:43Speaker 00

  Yeah, I, I think it's fair to say that. You know, early on, we had some sort of very small, sort of more limited prediction models of learning, let's say latent spaces first, then also sort of perturbations on top. And then, you know, when like this really large scale foundation model representation came out, everyone was sort of buying into the hype and maybe potentially sort of a bit over buying. Then, you know, obviously after this, you know, comes what's it called, the sobering up value or something like that. Yeah. And then there were a whole bunch of papers coming up, where they were saying, you know, hey, maybe even more or less linear model, we're doing a good job. And actually, in our SEgen paper, early on, we had already some of these benchmarks showing it's not so, not so super easy in simple tasks to improve on that.

• 21:23Speaker 00

  Um, I think, I think there's a truth to it, that in particular, the current data state for some things, let's say you want to predict the effect of a CRISPR in a cell, and you might not need the most fancy model. Whereas if you ask, or if you want to predict a cell type or something like that, you know, for that most logistic aggression is fine. But if you want to predict something a bit more difficult, for example, in our niche former paper, which is a foundation model for spatialomics, we showed, we had a new task which we introduced where you say you have a, you have the gene expression, you have sort of this state, a latent space of a single cell. Can you predict the gene expression state of its neighboring cells? Because, you know, they're always in communication, or maybe the other way around. So like, they're always in communication. So because, you know, they talk to each other, you sort of see a remnant of that sort of cell state in this one.

• 22:07Speaker 00

  And actually, this is harder to do. And if you sort of have these harder tasks, you need these stronger, more nonlinear things. And you particularly need to look for more subtle data. So your question, can you sort of find out more sort of information, which type of things? I think very much depends on what you sort of ask at the end of it. Of course, like in the self-supervised learning phase, I think we're definitely not there yet with the best way how we represent, how we learn these things. There's a lot of copying of what happens in NLP type of LLMs, right? And like masked autoencoder, so was the typical way to go. But the type of losses, the type of representations, I think people are still debating. Maybe because, you know, this is sort of often kind of tabular data, so even just like tabular foundation models are now sort of being tried out.

• 22:50Speaker 00

  There's a lot of JEPAR type of approaches now, because it's like, we have one thing where we use contrastive learning. So just people playing around with the losses, but also with the representations. And I think nowadays, reviewers, and at least I've been advocating this for a long time, are really much more closer looking that you really properly show scaling laws. This was what has missed from the early papers of the foundation models. And so people haven't really shown that you need this complexity to actually sort of really scale up. But if you go really, if you can really demonstrate at scale, like across examples and samples that you have this sort of scaling behavior, then I think you have this bind into these models. And for these, of course, there's not so much at least deductive biases in these use. And so I do think we're gonna, gonna hit this regime actually quite soon, and we do see this also in a few papers. So for example, in our contrastive learning paper, we call it SEConcept, we do show already

• 23:34Speaker 00

  quite, quite, quite strong scaling behavior. In the paper, we had something for going, because we compared with gene former going to maybe thirty million cells, but on, on hugging face now, we have a three hundred fifty million cell model with, I think, going up to maybe a hundred million parameters. So really going to the big scale for us. And we do see nice scaling behavior already starting to emerge.

• 23:52Speaker 01

  Just to be clear, then, at a hundred million parameters, and you said, I think, three hundred million cells or something like that, that you, you get a lift in what kinds of applications? In these, these sort of, the niche former kinds of questions where you're saying, what is my environment based on my transcriptome?

• 24:10Speaker 00

  Yeah, I mean, you know, the, the ultimate thing, of course, we want to do is we want to be doing this perturbation prediction better, because that's sort of what people nowadays understand by virtual cell, right? And for this, I would not claim that pure representation learning is the right way to go. Sorry for making this subtlety again, but I think typically you learn the state space with these foundation models, but then sort of the way how they behave under an action, under perturbation, this is where you need this more world model like type of approaches. World model being the very hip term, I'm sure you had this already at some point in your podcast, so to be on all of a machine learning conferences in last year. But de facto, a perturbation prediction, but like a virtual cell, is kind of a world model in terms of transcriptome or the type of omic space, right? So it's, it's not just a model that is sort of F of X, but F of X comma A, right? So the state X plus the action, right?

• 24:55Speaker 00

  And then sort of predicting the next state X prime or something like that. And I think for this, you might need other combinations. So what we typically do, we learn the representations with, for example, SEConcept. But then we put a generative model for the push forward along the action. So for this, we use something like flow matching or something like that. And in our self flow paper, this is not just got accepted for nature methods, we actually really show that these approaches are much better generating the effect of, of, of perturbations, but they're very much dependent on a good latent space. And you know, in this SEConcept latent space, in this case, for this three hundred fifty million cells, we actually went into multi-species. So you have now like sixteen different species where you can potentially look for similarities. Suddenly, here you can maybe ask some evolutionary questions that you couldn't before.

• 25:39Speaker 00

  Not that we can do this so well, it's not our core expertise, but you know, sort of you can start doing these fancy, um, sort of cross-species, cross-tissue type of latent spaces.

• 25:48Speaker 02

  I'm kind of curious about the history of some of these different models. I think you just gave us a sort of a rapid version of them. The, one of the things that, like, I think was like somewhat famous in the community was this thing about linear models being kind of a, a big point. Can you explain specifically, what was the like key point here? Because I think this is like one of those things in classic machine learning, make sure you always understand your baselines well, you know, otherwise you, you spend a lot of time spinning your wheels and fool yourself. I think there was a specifically a famous open competition on, with the, through the Arc Institute, that was a sort of a class of like cold water on everyone's, you know, faces. Can you maybe give a little bit of history about that and like why we thought this was

• 26:31Speaker 02

  exciting and then where, why this had us pivot a bit?

• 26:35Speaker 00

  Yeah. It sounds such an, like such an obvious learning, right? Do your baselines properly, but in a field that is rushing towards the gold part of the virtual cell, people sometimes maybe were not as careful with the baselines. And there's a whole bunch of subtleties with respect to holdout and so on, that, you know, wasn't always fully observed and why some of these things happen. So one of the first observations there was in a thing as basic as cell type prediction, okay? So nothing super fancy, okay? Basically, you, you, you give me a new transcriptome, tell me from which cell type it is, right? Maybe even from which organ or something like that. And it turns out that even nowadays, I think some close to state of the art models are really just logistic aggression, right? So for such a simple categorization, you know, this is actually not, not so bad. There's one thing called cell type, which is for example, which is doing a good job.

• 27:20Speaker 00

  And I think we had a paper maybe five or six years ago where we had the title very plain, deep learning is not needed for cell type prediction. This is true if it's within an organ. If you sort of go cross organs, suddenly it gets a bit more subtle, because you know, sort of cell types reoccur in different organs and immune cells, it'll subtle change depending which tissue they are. Suddenly, you know, there you need to do a bit more. And we, like for us, there was one paper called SETab, like a tabular foundation model, predictive model type of thing. We actually showed that this nonlinearity was really adding beyond the baseline. So that was sort of for, for, for that class, for the perturbation models, which then you were asking Brendan similarly, like early on, many analysis, particular perturbation were done with classical linear models in Sudubok. We had then our, our SEgen, that was like something around twenty nineteen, where we

• 28:05Speaker 00

  started introducing linear shifts, but in the latent space, so sort of nonlinear encoded. And we always compared with subtle, sort of simpler linear shifts. But then the field getting complex and, you know, sort of data sets popping up, people started asking more and more perturbation type of questions. There have been a whole bunch of these early on competitions, like a regular one run, for example, by MIT. We've been running a regular NeurIPS competition on prediction problems in single cell data, sometimes for integration data, also for perturbation prediction. Um, that was very well attended. The cherry on the, on, on, on the cake was most definitely that, that, that ARC competition, which called the thing not just perturbation prediction challenge, but virtual cell challenge. This sounded a bit cooler, right? We picked up very well, but with metrics that are robust, were maybe, but maybe not,

• 28:50Speaker 00

  not that easy. So for, actually, actually, our solution in, in this, this, this consortium, where a lot of labs contributed called open problems, was actually that people need to submit their own metrics in addition, so to really quantify maybe subtle things that you might not have thought before. The ARC one, it sort of fixed a bunch of metrics that weren't even so, so single cell specific, and they could be just, just exploited. So actually, I think the winning model that was then in the end actually sort of crowned as the actual winning model, but essentially it's a flow matching based thing, so based on, on this idea that people pour forward with this flow, with this self flow type of idea. I think shows you that some nonlinear things are in this case needed, so that the actually real winner, so if you sort of remove the exploitation, was actually something that is, this is taking some, some nonlinearities.

• 29:30Speaker 00

  And I think nowadays, we do see that just linear or additive type of models are not cutting the, the, the cake for more complex things, but for, I think even nowadays, if you go in a simple, if you sell lines and you have sort of G knockouts, then particularly when the effects are a bit subtle, all these sort of more complex models are sometimes a bit overfitting, and that's why linear baseline is always very good. So I always tell my students, guys, I know we, we, guys, I mean, guys and girls, we always need to do, I know you want to do the most fancy deep learning thing, but please, please, you know, have the baseline. And like, say for drugs, you know, baseline as simple as just predicting, getting better at the mean, you know, like that's like the hardest base, like not even linear model, like better than the mean, sometimes not as hard if there's so much noise in the data. And this stuff is super noisy.

• 30:15Speaker 01

  I just wanted to add one point that Bo Wang and some of his collaborators actually just had a nature paper on, on how the metric was actually the problem in some of the, those linear baselines beating some of the models as well. So it's clearly, it's not only metric hacking, but also just that the, asking the wrong question leads to the wrong answer.

• 30:34Speaker 00

  Or choice of the metric, yeah. And I, I think that they could, you can discuss it at some point, we're gonna have agents sort of trying out all kind of combinations of whatever algorithmic thing you can come up with. In fact, we actually do have some collaborations with some of, of those sort of auto ML type of labs to try this out. So maybe what is left sounds a bit too harsh, right? But like, boy, maybe our biggest ideas go into is actually this metric formulation, RJ, right?

• 30:58Speaker 02

  So one thing I've been really interested in is the different models of funding and research academia. I think Germany has a different structure fundamentally from the US. How is your institution set up? And like, how does this give like distinct opportunities versus like a US-based research model?

• 31:15Speaker 00

  I, I very much agree that these funding models very much matter for the flexibility, but also the set of teams that you can set up and like the long-term vision that you can have. So, um, Helmholtz?

• 31:27Speaker 00

  This is not just Germany, but I would say maybe Europe's largest, what's called research organization in terms of funding. It has a different set of areas that it focuses on, health being one of the big ones, but then it has sort of material, information sciences, physics and so on. And essentially getting federal money to put into a deep research question. I've been building up what's called the computational health center at Helmut Munich, which I think has now something like maybe forty five group leaders, professors and maybe five hundred to six hundred people now. So we are, I think, the largest academic place for health AI type of research in Europe, where we have people really working on all scales. Like, you know, I'm sort of focusing on the cellular scale, but you know, there's people

• 32:09Speaker 00

  doing DNA foundation models, there's people sort of doing interesting things on protein design, there's people doing like, you know, patient cohort type of representation genetics and so on. And now you sort of can start putting them together and one of our big visions is to intersect and to go virtual human and so on. But having said that, the key thing is in contrast to universities, that even though we all have affiliation to universities, mostly to the Technische Universität or the University of Munich, the two top ranked universities interesting in the European Union. So we have this sort of place where we have these affiliations at universities, right? But we don't have undergraduate teaching, we essentially just have PhD students and postdocs, but we interact very much with the universities and in particular also the clinics, because you want to have sort of this translational thing in there.

• 32:51Speaker 00

  And increasingly, and that's maybe an interesting link, so this is maybe a bit similar to NIH type of models like in US, but sort of distributed all over the country, not just sort of in one place. And maybe the one interesting link is that increasingly the government and this sort of realization also in Germany is sort of dawning that there needs to be much stronger transfer component, not just by going into translation to universities, but also into sort of value creation outside of the academic realm, which is obviously going so more to startups and so on, which I think the Boston environment has always been one of the key examples, but now of course more so now also the Bay Area, when, you know, recently I was visiting, I believe maybe Genetek or colleagues, and we went to a restaurant and

• 33:35Speaker 00

  then one of the waiters was asking us, hey, are you guys talking about AI and bio, you know, I have been also interested. So like everyone's sort of thinking about this startup ideas and so on. And I kind of felt that's a good environment and this is something I think our PhDs are now more and more doing and we're sort of learning there. And I think also in terms of sort of value creation, Munich in particular is actually at least in continental Europe, maybe the prime place now for VC capital, but also sort of deep tech, particularly in the bio and defense area and sort of pretty exploding. So this stuff starts to work.

• 34:04Speaker 02

  One thing, one, the thing we're talking about startup, I think, is really fascinating. I've heard that Germany has had some issues with their startup ecosystem. It's, it's, it's difficult to create a startup there compared to even other places in Europe. How like, what is the competitive landscape for startups right now in Germany and specific? And how does Germany become more competitive broadly? I have a vested interest in this because I'm probably going to get my German citizenship soon.

• 34:31Speaker 00

  Welcome. Let me know if you come, come by Munich. So I think traditionally Europe, but I would say definitely Germany is quite bureaucratic and processes are far from as digital as you would expect in different places. Not that I'm saying that US is maybe the best example for fully digitized processes, but stuff is definitely faster. The thing is people have realized that now, and I think the particular startup creation and implementation is, I think, more than competitive in Germany versus Europe, is actually sort of the, the, the prime number and in particular like some hotspots burden, but in particular Munich sort of the, I think, is the capital for, for highest number of startups. And I think even nowadays, if you look at to European Union, there's equally as many startups being created as in US.

• 35:13Speaker 00

  The thing is where Europe and particularly also Germany drops off is at the scaling part. So you know, early VC and also sort of non-dilutive type of money, this is there, but then when you go into scaling, these big VCs that sort of take these bigger risks are, I think, not closely as present. So I saw this again and again, that, you know, if you want to sort of really go into the sort of bigger tickets, sort of serious A and so on, serious B, you, people start going more into US also for that. But maybe I'm sort of a bit simplifying here, but this is at least my current learning brain.

• 35:44Speaker 02

  Okay, you are on the board or I guess part of several local startups, right? What, what does it take to go from, you know, idea to creating a, you know, to creating a new startup in Germany?

• 35:56Speaker 00

  I, I don't know if you heard that the European Union tries to regulate a whole bunch of stuff and that can be sometimes an advantage, you know, it's kind of cool, like, you know, all across Europe, we cannot go with a mobile phone. And you might have heard that everyone needs to have USB C things now for their mobile phones. This is kind of, kind of cool stuff. So there's some, some good stuff coming, but then of course there's something like things like the AI act and so on, that makes things a tad more complicated. Luckily not yet for research, but you know, for data access across hospitals and so on, there's some, you know, and we were talking about, about health, deep tech type of stuff, right? So there's some systems that are, I think, fantastic in Europe. And I'm thinking about Finland, you know, Finjan, Denmark and so on, with this sort of very detailed description of the population in terms of sampling, getting medical records

• 36:40Speaker 00

  connected, far better than, I mean, of US is also similar to MS like, like Germany in terms of sort of getting all the systems sort of to talk to each other. Like when patients, you know, sort of go to different states in US is similar as in Germany, when they go to different places, it's like very hard to build patient trajectories, right? But there's some places that know how to do this and there's a thing like the European Union tells you, it's called the European health data space, where back maybe next two years, these things need to be sort of going across boundaries, you need to sort of start building these, these health record type of combination. So I do think divisions are starting to happen, but the scale is, is not as fast. So I think what you need for a tech startup, what I noticed is, is always the biggest thing is you need to have the best team for that.

• 37:22Speaker 00

  And you, we have professors and PIs that really think about this. So for example, for, for the department that I called it, I said, guys, every, every PI, every group leader would expect you at least to have one, one, one spin off in, in the next three years and see how that works, right? But you sort of, you sort of get people a bit, a bit on that track. But I think this is happening very much from the bottom up nowadays anyway, where the students are actually pushing for that. They do realize that there's a whole bunch of opportunities at the moment, particularly with, with this sort of computational angle getting, yeah, I would say pushed even more on the forefront, that we can now start actually guiding experiments. We haven't really spoken so much about it, right? But once you have these perturbation world models, you can actually go back and drive the experiments, right? So I do think computational startups can be very good also data generation.

• 38:07Speaker 00

  And I think this is what will be also turned not just into nice papers, which is sort of mostly our business, but also I think in this next generation of startups. And you do see a few of these popping up where you actually get the lab and the loop really implemented.

• 38:18Speaker 01

  What do you see as some of the, the, the most promising areas for commercialization of sort of these virtual cell adjacent ideas?

• 38:28Speaker 00

  It's interesting, right? Like if you looked at earlier, and of course this is what people always reflect to what OpenAI and Fop and so on wanted to do, like they didn't really know how to make money. And some people might argue they don't even know nowadays, right? And there's a few startups that, you know, got really quite strong investment that also said, we build a foundation model of biology, we decode biology and so on, without actually saying what you want to do with it. And I've been seeing this again and again, that these platform type of ideas are super attractive, you know, they sort of fit into what we know. I mean, you spoke about some of the tools and platforms that we build like Scanpy and SCW and so on, right? But turning this actually into a monetary thing is not as obvious. I mean, there might be some sort of consulting business, but if you say you build a virtual cell platform, typically what investors understand, and this is maybe even more so

• 39:12Speaker 00

  in the more conservative VC ecosystem in Europe than in US, is that at some point you develop a drug and then you know, it's a clear thing. So I saw so many of these platform companies pivoting at some point to we make a drug. And yeah, decide, of course, you have our, our sort of virtual cell, whatever business, but you know, we have two drugs that we found and now this, so if we put a lot of money in, maybe you have some other counting something. There's a few companies that are sort of still stick with the scaling up of, of their AI business, but it's not as clear how you would sort of directly turn this into a viable sort of really scalable business model and if there's even room for many of these. And of course, then there's also this angst that at some point, like Astra 2.0 might be doing parts of this already anyway.

• 39:53Speaker 01

  Yeah, I mean, I, I think that it's actually, this has been a thesis of mine that up until recently, every single company that builds a model and builds a something ends up just building drugs because that's the path to monetization over time, right? And that, like, whereas you, you have to like land sales with pharma that's very difficult or some other big, big pocket spender, if you're a tools company, if you say, I'm going to build a port, like an asset and then I can sell that off, that's really attractive to an investor.

• 40:27Speaker 00

  Yes.

• 40:28Speaker 01

  But, but we've seen actually the opposite happening this year, right? So you have like, you have Instructural, you have a couple, you have Bolts, you have, Brandon will know them better, but there's a couple of those. And then, oh, Chai is one that we talked to. And then we've interviewed a few that are doing this, doing reinforcement learning environment plays as well. I would love to see the activity level in structure that, that we see in structural, also doing things like what Noetic is doing.

• 40:56Speaker 00

  I mean, I, I think Noetic is, is a good example in the sense that they also have, I think, a rather nice scale of data generation. And a big part of that is also access to data. So, you know, we talked before about these organoid perturbation type of things. And we've been setting up actually in collaboration with a whole bunch of pharma, a consortium which we call the Cancer Organoid Perturbation Atlas, where then pharma actually buys in because they want to get early access to some of these sort of perturbation type of data sets, right? And I think some of these deals sort of go in these, these directions. So I think if you couple it with data generation, not just with the modeling part, I think this is where sweet spot hits, where then also actually pharma gets buying. I think more and more of the companies I see, they sort of build their mode not just around a virtual cell model, but in combination with sort of unique, so the specific data, this is then of course also more value creation.

• 41:40Speaker 01

  I don't know if this is exactly correct, but I read that you said somewhere that you have your students follow what, what you, I think, called a Y model, in which you have to have a methods and a wet lab advisor. That sounds super interesting, super relevant to the conversation we're having right now about data being so important and all the. Attention on lab work and that being moat for a lot of companies. Can you talk a little bit about that?

• 42:07Speaker 00

  Yeah, thanks for raising this. So I've been setting up and since then it's been sort of become quite a big thing. What's called the Munich School for Data Science, right? So this is beyond just bios. So we had a whole bunch of other disciplines. And there we had this supervisor model that is always two PIs need to come together to make a proposal. Then sort of proposes to the students and a graduate student type of model, which has this, this Y, so there's like a methody more domain one, right? And then the supervisor student. And sometimes we had two students, another Y model, so Y plus Y equals H. You get it? But, but I think this idea is somewhat even nowadays a bit trivial, right? But I, I thought it was good to bake this in so hard so that students really ideally go to two lab meetings. One lab meeting where they need to really learn and understand some deeper biological problem, understand the language, maybe of a medical setup or something like that.

• 42:52Speaker 00

  Then they go to computation or typically they're all computation data science type of students, right? But then they go to the computation one, they sort of get push the method aspect of it. And thereby sucking up a whole bunch of, of problems that are sort of not exactly precisely yet formulated in, in that bio lab. That sort of can't directly put yet in a function, but then because they are sort of methody, right? And they're quantitative, they can actually put this into a metric or whatever that then turns into a model and in a paper. And I, I think this is a way how, uh, I mean, we've been doing this now for maybe eight years or so, this, this type of school. I really, a whole bunch of papers came out where I, I, I learned about systems that I had no idea about this before, right? You know, people approached me, you know, could we do something like this? And then, you know, I learned something about the lung or the brain that, you know, I had no idea about.

• 43:37Speaker 00

  But then, uh, I said, oh, hey, cool data. I think our model types could fit. Let's build a project around it. And then those students became the experts in my lab for that sort of detailed biological thing together with a partner. And then we, we drive this forward. And this actually nowadays extends also across the data types. Early on, a lot of things have been around just RNA. But, you know, since then, you know, we had a very nice paper in Science last year on single set proteomics. We've been extending things to spatial observations, you know, there's sort of all these additional modalities and things that you can, can, can measure. And I think then interesting questions come up in the particular domain where, you know, these essays work very well.

• 44:11Speaker 01

  Yeah, I mean, this is my natural inclination is to like gathering cross domain stuff. It, I think it's all often ripe for a good, good generation of really good and interesting ideas. Have you, just as a follow up, these, you've been doing this eight years now, I guess. So what kinds of, you know, sort of longer term outcome benefits have you seen from this?

• 44:33Speaker 00

  I think one thing that, this is also, I think, how we got students to even join in this maybe initially. It sounds like more work, right? You kind of need to learn two things and you go to two lab meetings and so on. But I think once your students are done in, in, in, in Germany, Europe, I think in general, but bigger. I think in Europe in general, people sort of typically do a master's first before they go into a PhD, right? So not just a bachelor's in US, right? So you do like one or two year more of master. And, and then they sort of know science a little bit, but they don't exactly know what sort of interests them. So in these combination type of thing, you know, they can sort of sample a bit because in this grad school they get different proposals. And so they can sample a bit what they're interested in terms of domain. But then they can also play around a bit with the method side. So I think that was kind of attractive.

• 45:12Speaker 00

  Yeah, I, I think with this sort of trying out, uh, uh, different things, I had the impression that this sort of turned the students into very good data analysts, but then also modeling people. And then a downstream, you know, the sort of areas where they, they went into was also equally diverse, you know? So some people went like now in the biomedical domain into pharma, but you know, some obviously data research. Some went more into some sort of startup direction. Some went into very pure AI type, type of work. I think Nvidia just a few months ago hired in one month, two of my people going very much into sort of the computation aspect. And also a bit of a bio part that Nvidia is now building up and there's similar of the big companies. So I think it, it gives them an, an, an interesting profile. I think nowadays when you get interviewed, you know, it's, it's a good idea to say, you know, I, I didn't only do like this fancy model

• 45:57Speaker 00

  that improved maybe 2% in a NERUB paper type of thing, but you also sort of had the real sort of more complex, deeper application that like, you know, I had to help find as, as a use case.

• 46:07Speaker 02

  There's been a movement amongst agentic AI, like several guests we've had, or we've had a blog post about how the wet lab scientists have started basically becoming computational thanks to agentic AI. I don't know if this is, I was, I'm kind of curious, like as this evolves, as basically now agentic AI enables any scientist to be now become their own computational. What do you think the evolution of this sort of like co advising strategy is? And also what do you think the evolution of like integration between wet lab biologists and computational teams are? I mean, I also can think of like, I've heard several, you know, famous CEOs say things like, our strength is integration between computational and wet lab teams.

• 46:48Speaker 02

  So I'm, I'm curious to see how you think that evolves and how you also like set up culture to exploit, you know, these new technologies.

• 46:56Speaker 00

  Let, let me maybe sort of keep the, the culture part a bit out because I think it's a hard one to answer, but a very good one. So let me first sort of talk about the, I think, very cool, cool point that I, I think you asked about where, where these things happen, sort of goes with deep integration. I think one of the most exciting frontiers we currently face, and I think maybe have as an opportunity is we can really now start guiding the experimentation. So where before you, we built this human cell atlas, we had sort of this mapping aspect. Then we talked about the word model prediction problem, right? So from mapping, we went to prediction. But now we can go to actually steering the next experiment, right? And this has always been a scientific method, you know, sort of putting our hypothesis in the loop and then doing something with it. But quite frankly, very often the papers end it with some model, right? Maybe some sort of concept figure, maybe at a little bit validation.

• 47:42Speaker 00

  But then typically you start the next project, right? And of course you learn a bit, but there's not like, not like a continual learning type of thing. But now, you know, I think we can actually really loop back more and more. This is happening very much so in a molecule design or material design and so on. You know, sort of the lab and the loop is really working out. And even though some big pharma has been putting this out as sort of a general, general big theme. There's not many examples here, but really people do multiple loops of this integration of an experiment, a computation model, then going back, whether this is now two person or one, right? But like this integration is hard to set up. And it's hard to set up once because of modeling, because actually sort of once you have a model actually deciding what's an optimal next question is not trivial.

• 48:22Speaker 00

  And we actually brought out, I think it should be on, on bio archive today or tomorrow, a new thing which we call lab compass, which is an inverse generative model that actually allows you, once you have a forward word model or whatever, to tell you what you need to do in order to reach a certain state. So it's super exciting for us, right? And I think a big part of the work was not just the model, but actually the work with experimental people. In this case, a good friend and colleague from University of Cambridge, who was doing hematopoiesis in addition, had a sort of a rapid cycle system where we actually set up three, four lab in the loops, where we iteratively improved on a model. And this is, I think, a really hard thing. And this sort of goes into your second question, Brandon, right? Like, how do you bring up a culture where once your computational person or model or whatever predicts something

• 49:05Speaker 00

  that you actually get to convince experimental people to put their lifetime in? Because these experiments often take some time, right? To actually buy internet, actually measure some of these follow ups. And sometimes, you know, this works a bit better if it's sort of molecules because then, you know, so the stakes are a lot is high. But you know, if you like say, now you need to do this new seller and you need to add these very strange things. Do you really want to do this? Like, shouldn't we maybe not take a bit of our gut feeling and then, you know, this sort of these intermediate solutions come? So I think really setting this up is far from trivial. And that needs culture and it might be the same person doing it. A lot of people then start throwing around robotics and so on as well. But I, I think this is a tad too early because the, the cycle times in terms of our essays are still a bit on the long side. But I do think this will come and, you know, obviously the, the ultimate vision being that we want to RL everything. We want to do RL also in, in, in this loop. It's a bit out there, but you know, I think we're slowly getting there. At the moment it's more maybe actively.

• 49:51Speaker 01

  I think that very much like we talked to Lila, there, that is their, their, kind of their core thesis is, is, is that basically the lab is part of the RL environment, kind of as well. Yeah, no, that's great. Fabian, it is late at night for you now. I really appreciate you taking the time out of your evening to, to talk to us. Look forward to, to watching your research. And when I say you, I mean, your lab and colleagues and you as well. I really look forward to watching this progress and keeping careful track of the things that we talked about today.

• 50:23Speaker 00

  Yeah, thanks, thanks so much, RJ, Brandon, for having me. I, I think these links are important. This sort of look in beyond what maybe you do like in your own sort of smaller disciplines and these discussions helped me a lot. I really enjoyed the discussion, uh, had a lot of fun. Thanks.

• 50:37Speaker 01

  Great. Thanks. Take care.

• 50:39Speaker 00

  Thank you.