Controlling Biology - Fabian Theis, Helmholtz Munich

“We can really now start guiding the experimentation. So where before you, we built these human cell atlases, we had sort of this mapping aspect. Then we talked about the world models 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 ended with some model, right? Maybe some sort of concept figure, maybe add a little bit validation.”

Claim: Researchers can now use models to guide experimentation, progressing from mapping biological systems and predicting outcomes to steering the next experiment.

“From mapping we went to prediction. But now we can go to actually steering the next experiment. But quite frankly, very often the papers ended with some model, right? Maybe some sort of concept figure, maybe add a little bit validation.”

Claim: In the speaker's experience, research papers often concluded with a model, sometimes accompanied by a concept figure and limited validation.

“But then typically you start the next project, right? And of course you learn a bit, but it's not like, not like a continual learning type of thing. A new thing we ca-- which we call Lab Compass, which is an inverse generative model that actually allows you, once you have a forward world model or whatever, to tell you what you need to do in order to reach a certain state.”

Claim: Lab Compass is an inverse generative model that uses a forward world model to identify actions needed to reach a specified state.

“But then typically you start the next project, right? And of course you learn a bit, but it's not like, not like a continual learning type of thing. A new thing we ca-- which we call Lab Compass, which is an inverse generative model that actually allows you, once you have a forward world model or whatever, to tell you what you need to do in order to reach a certain state.”

Claim: The speaker says research projects typically proceed to a new project, with some learning but not a continual-learning process.

“Obviously, the, the ultimate vision being that we want to RL everything. We want to do RL also in, in this loop. It's a bit out there, but you know, I think we're s-- we're slowly getting there.”

Claim: The speaker's ultimate vision is to apply reinforcement learning broadly, including within the experimental loop.

“Obviously, the, the ultimate vision being that we want to RL everything. We want to do RL also in, in this loop. It's a bit out there, but you know, I think we're s-- we're slowly getting there.”

Claim: The speaker considers applying reinforcement learning throughout the experimental loop an ambitious goal, but believes progress toward it is gradual.

“I can't emphasize enough how much Fabian Theis and his labs and companies have contributed to the development of systems biology and various iterations of the virtual cell concept.”

Claim: The interviewer says Fabian Theis and his labs and companies have made substantial contributions to systems biology and to versions 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.”

Claim: The interviewer says Fabian Theis's lab consistently advances understanding of how to observe and model cellular behavior and interactions.

“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.”

Claim: The interviewer says Fabian Theis's lab publishes papers and releases models at a pace faster than the interviewer can keep up with.

“Just as importantly, his work often focuses on creating the tools that researchers and scientists actually use in their day-to-day work. Most computational biologists have probably used Scanpy and Anndata at some point, and this work has led to the creation of scVerse, which is an amazing collection of open tools that has attracted many, many contributors.”

Claim: The interviewer says the work led to scVerse, a collection of open tools that has attracted many contributors.

“Just as importantly, his work often focuses on creating the tools that researchers and scientists actually use in their day-to-day work. Most computational biologists have probably used Scanpy and Anndata at some point, and this work has led to the creation of scVerse, which is an amazing collection of open tools that has attracted many, many contributors.”

Claim: The interviewer says Fabian Theis's work often produces tools used in researchers' and scientists' daily work.

“Just as importantly, his work often focuses on creating the tools that researchers and scientists actually use in their day-to-day work. Most computational biologists have probably used Scanpy and Anndata at some point, and this work has led to the creation of scVerse, which is an amazing collection of open tools that has attracted many, many contributors.”

Claim: The interviewer estimates that most computational biologists have used Scanpy and Anndata at some point.

“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.”

Claim: The interviewer says Fabian Theis leads multiple AI and computational biology efforts at Helmholtz Munich and related institutions and participates in other industry efforts.

“There's a ton of interest right now in structural biology. Anthropic and OpenAI biopharma initiatives focus heavily on designing molecules right now.”

Claim: The interviewer says there is currently substantial interest in structural biology.

“There's a ton of interest right now in structural biology. Anthropic and OpenAI biopharma initiatives focus heavily on designing molecules right now.”

Claim: The interviewer says Anthropic's and OpenAI's biopharma initiatives currently focus heavily on molecule design.

“On the other hand, virtual cell and systems biology approaches really haven't gotten as much attention recently. Maybe Xaira and Noetic are good counter examples there.”

Claim: The interviewer says virtual-cell and systems-biology approaches have received comparatively less attention recently, while suggesting Xaira and Noetic may be counterexamples.

“But Biohub just announced a one point eight billion dollar collaboration with the DOE, NIH, Meta and Google to collect and open source multi-scale biological data that can be used to develop a new generation of virtual cells.”

Claim: The interviewer says Biohub announced a $1.8 billion collaboration with the DOE, NIH, Meta, and Google to collect and open-source multiscale biological data for developing a new generation of virtual cells.

“I, I, I should disclose before I sort of go to, to the answer that obviously my, my opinion is quite academic in the sense that whenever I talk to, to my, my friends and colleagues that really sort of day by day sit in, 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.”

Claim: Fabian Theis described his perspective on startup activity as academic, based on an overview and intuition rather than day-to-day experience inside startups.

“And you know, this is sort of the, the hot big thing that, you know, this and this investments are happening. I don't really have that on, on, on that scale. But having said that, for example, you, you, you me-- you mentioned, uh, 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, uh, level of dataset generation.”

Claim: Fabian Theis said Biohub's initiative is closely connected to the Human Cell Atlas, an international consortium promoting the next level of biological dataset generation.

“I do think that there's a whole bunch of reasons why we're gonna see that sort of scale up 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, on, on the cell level.”

Claim: Fabian Theis expects generative models and molecule-level lab-in-the-loop approaches to scale up to work at the cell level.

“And I, I think there's a bunch of reasons for that. And the first one, I think maybe the most important one, which you really hit here, is the data generation. I think we had that concept of a virtual cell for a while.”

Claim: Fabian Theis considers data generation a particularly important reason virtual-cell research is expected to scale up.

“Like my, my lab, we started this perturbation prediction modeling and, and using machine learning in twenty nineteen with a paper we called SEGen at the time, which was like it was a nonlinear model, but essentially sort of in the latent space, did some linear transformation, and it was actually quite okay.”

Claim: Fabian Theis described SEGen as a nonlinear model that applied a linear transformation in latent space and performed reasonably well.

“Like my, my lab, we started this perturbation prediction modeling and, and using machine learning in twenty nineteen with a paper we called SEGen at the time, which was like it was a nonlinear model, but essentially sort of in the latent space, did some linear transformation, and it was actually quite okay.”

Claim: Fabian Theis said his lab began machine-learning-based perturbation-prediction modeling in 2019 with a paper called SEGen.

“And even nowadays, many of these sort of pseudo-linear type of things are still doing a good job. But the reason for that being that I think the datasets 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 your question, I think what you actually ask is how does-- can you predict how a cell would behave under all kind of perturbations?”

Claim: Fabian Theis attributed the lack of earlier momentum in virtual-cell research to relatively simplistic datasets and research questions.

“And even nowadays, many of these sort of pseudo-linear type of things are still doing a good job. But the reason for that being that I think the datasets and the questions that we asked on top were just sort of still on the simplistic side.”

Claim: Fabian Theis said many pseudo-linear approaches still perform well because the datasets and questions they address have remained relatively simplistic.

“That's why also sort of it didn't yet take up. So whenever people talk about virtual cells, maybe also your question, I think what you actually ask is how does-- can you predict how a cell would behave under all kind of perturbations?”

Claim: Fabian Theis described predicting a cell's behavior under different perturbations as a deep question and possibly a holy grail of computational and systems biology.

“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 its specifications and with all the different sort of levels of molecules and so on, not just RNA.”

Claim: Fabian Theis characterized a cell as an input-output system with memory and multiple levels of molecular specifications, extending beyond RNA.

“I, I think being able to map this out, what does even understanding mean? And I've been the sort of bit of a meta problem, but I, I think you understand, or at least you model a system very well when you sort of predict its output behavior given different inputs.”

Claim: Fabian Theis considers a system well understood or modeled when its output behavior can be predicted from different inputs.

“So I think it's a very good question, and this is why often people nowadays equate a virtual cell with, um, perturbation prediction. I think it's a bit too simplistic, but I think these perturbation prediction type of models, and that sort of being, uh, maybe what you mean by virtual cell, is taking off a lot.”

Claim: Fabian Theis considers it somewhat too simplistic to equate a virtual cell with perturbation prediction.

“So I think it's a very good question, and this is why often people nowadays equate a virtual cell with, um, perturbation prediction. I think it's a bit too simplistic, but I think these perturbation prediction type of models, and that sort of being, uh, maybe what you mean by virtual cell, is taking off a lot.”

Claim: Fabian Theis said perturbation-prediction models are gaining substantial momentum.

“And the startup landscape, I think, is a bit exploding there. We, we talk, I think, a lot about just Europe, but also in, in the Chinese ecosystem, there's a lot of, of these, these additional efforts now happening.”

Claim: Fabian Theis said startup activity in virtual-cell and perturbation-prediction efforts is expanding in Europe and China.

“And I, I hope that these are not just sort of now a kindling of, of a flame because everyone sort of adds AI to their things, but be- because pharma is really desperately, uh, needing to be cell type context specific in their screens, which at the moment they can't do yet on that scale.”

Claim: Fabian Theis hopes the current growth in virtual-cell activity is not merely a temporary trend caused by companies adding AI to their offerings.

“And I, I hope that these are not just sort of now a kindling of, of a flame because everyone sort of adds AI to their things, but be- because pharma is really desperately, uh, needing to be cell type context specific in their screens, which at the moment they can't do yet on that scale.”

Claim: Fabian Theis said pharmaceutical companies urgently need cell-type-specific screening, but currently cannot perform it at the required scale.

“I mean, f-for, for me, if I may stop me when I'm sort of go-- getting t-bit too, too much into these, these, these old stories. Like, you know, I'm sort of... I'm, I'm, I'm-- I've been trained machine learning scientist for the longest time doing all kinds of things, but, you know, I started with single cell early on qPCR.”

Claim: Fabian Theis began his career as a machine-learning scientist and started working with single-cell data, including qPCR, early in his career.

“Maybe initially, if I'm really honest, not so much because it was like the bi-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.”

Claim: Fabian Theis says he initially pursued single-cell biology partly because it offered high sample numbers, not primarily because he was interested in the biology itself.

“Maybe initially, if I'm really honest, not so much because it was like the bi-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.”

Claim: The number of qPCR measurements at scale and, later, transcriptomes was growing substantially.

“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, we go into to a regime, and that was sort of one of these eye-opening moments when, when Pars Bio had like, I think they call it the GigaLab or something like that, and you know, there's sort of similar ones from competing companies now.”

Claim: Fabian Theis says that early in his career he had a few RNA measurements, while later he had thousands.

“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, we go into to a regime, and that was sort of one of these eye-opening moments when, when Pars Bio had like, I think they call it the GigaLab or something like that, and you know, there's sort of similar ones from competing companies now.”

Claim: Fabian Theis recalls that Parse Biosciences had an initiative he thinks was called the GigaLab, and says competing companies now have similar efforts.

“But when they had like in a single experiment, ten million transcriptomes measured, right? I thought that was like, like next level stuff and, you know, this step from just getting more PBMCs as a control to 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.”

Claim: Historically, the number of perturbations sampled was much smaller than the number of cell observations.

“But when they had like in a single experiment, ten million transcriptomes measured, right? I thought that was like, like next level stuff and, you know, this step from just getting more PBMCs as a control to 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.”

Claim: Fabian Theis considered it next-level work when ten million transcriptomes were measured in a single experiment.

“But now this is exploding. Now you sort of have, I don't know, a hundred different cell lines where you do sort of full perturb, like full crisp out, or you have like batteries of a thousand drugs or so added, and suddenly this space gets really rich, right?”

Claim: Perturbation datasets are expanding rapidly, with examples including roughly one hundred cell lines receiving broad CRISPR perturbations or panels of roughly one thousand drugs.

“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, can talk a bit about more about that.”

Claim: Fabian Theis and his colleagues are involved in projects that perform perturbation experiments in organoids, beyond cell lines.

“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, can talk a bit about more about that.”

Claim: The increasing scale of perturbation data makes it possible to begin performing perturbation prediction.

“Because I, I'm quite convinced that we need to go to multicellular systems to have actually really interesting readouts. But I think everyone is scaling this, these, these levels up, and at some point, you know, we're gonna go maybe even to human genetics as a perturbation.”

Claim: Researchers are scaling up the different levels of perturbation experiments.

“Because I, I'm quite convinced that we need to go to multicellular systems to have actually really interesting readouts. But I think everyone is scaling this, these, these levels up, and at some point, you know, we're gonna go maybe even to human genetics as a perturbation.”

Claim: Fabian Theis suggests that human genetics may eventually be used as a perturbation.

“Because I, I'm quite convinced that we need to go to multicellular systems to have actually really interesting readouts. But I think everyone is scaling this, these, these levels up, and at some point, you know, we're gonna go maybe even to human genetics as a perturbation.”

Claim: Fabian Theis is quite convinced that researchers need to use multicellular systems to obtain genuinely interesting readouts.

“When I give sort of public talks, I, I typically say, you know, human genome project, we sort of have the genome for, for, for the body, but it's pretty much the same everywhere, right? So like, why do we have all these different cell types, right?”

Claim: The human genome is largely the same across different cells in the body.

“What's sort of happening specifically? And so 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, well, these are sort of differently expressed as one says in, in, in different cells.”

Claim: Different cells express different messenger RNAs transcribed from the genome; these RNAs encode downstream proteins and vary across cell types.

“And to say in a human, maybe it's just something like twenty thousand different types of these, we're splicing many, many more, but they're sort of the, the rough level. So what you de facto have is per cell a twenty thousand dimensional vector that has a whole bunch of zeros when stuff is not there, but then a whole bunch of sort of numbers, how many of these counts of these messenger RNAs you have.”

Claim: At a rough level of description, humans have about twenty thousand types of messenger RNA.

“And to say in a human, maybe it's just something like twenty thousand different types of these, we're splicing many, many more, but they're sort of the, the rough level. So what you de facto have is per cell a twenty thousand dimensional vector that has a whole bunch of zeros when stuff is not there, but then a whole bunch of sort of numbers, how many of these counts of these messenger RNAs you have.”

Claim: A cell's transcript measurements can be represented as a roughly twenty-thousand-dimensional vector of messenger RNA counts, including zeros for transcripts that are absent.

“Because these things are called transcripts, right? This is called then the transcriptome. Yeah. And, and, and de facto then you have this sort of twenty thousand dimension space that we very often encode in a sort of nicer latent space that sort of has like an, a sub manifold, like sort of this, this subspace where these actual cell cloud lives, right?”

Claim: Researchers often represent the roughly twenty-thousand-dimensional transcript space in a lower-dimensional latent space, with cells lying on a submanifold.

“Because these things are called transcripts, right? This is called then the transcriptome. Yeah. And, and, and de facto then you have this sort of twenty thousand dimension space that we very often encode in a sort of nicer latent space that sort of has like an, a sub manifold, like sort of this, this subspace where these actual cell cloud lives, right?”

Claim: Messenger RNAs are called transcripts, and their collection in a cell is called the transcriptome.

“And if the cell cloud then sort of like this sort of twenty thousand 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.”

Claim: Clusters of cells in transcriptomic space are called cell types.

“And if they sort of go into a sort of sausage or whatever, right, we call them then trajectories, like a cell differentiation, like the fertilized egg cell then making all the cell types in the body or something like that.”

Claim: A trajectory through cell-state space can represent cell differentiation, such as a fertilized egg developing into the body's different cell types.

“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 a, 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.”

Claim: During normal development and homeostasis, cells continuously communicate with one another.

“You know, they sort of live in these complex multicellular states, right? So there's all type of communications happening between cells. So they always have stuff coming from the outside and then sort of doing something to them.”

Claim: Cells in complex multicellular states receive external inputs and respond to them.

“So for me, a perturbation space, in fact, is just an input space to stuff that happens to a cell, just maybe-”

Claim: A perturbation space is the input space of influences or events acting on a cell.

“... because it just sits to another one and there's like some signaling happening, right? Or some gradient, some hormones coming in, whatever, right? And all this stuff, of course, you cannot sort of tune up or tune down.”

Claim: Inputs to cells can include neighboring-cell signals, gradients, and hormones, and these influences cannot all be experimentally tuned up or down.

“Maybe even something as basic as when you go these in vitro systems where pharma does drug screens, when you sort of put them in, in a dish, you know, you can do different what's called media conditions, and they have so much impact, right?”

Claim: Media conditions can substantially affect in-vitro systems used for pharmaceutical drug screening.

“And it's like a bit of, for me, still like a, 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 em-- into more or less embryo-like state cells, right?”

Claim: The speaker characterizes the selection of biological cocktails for converting skin cells into embryo-like cells as bio-driven and possibly akin to alchemy.

“Where you can then sort of make new cells to heal different stuff in your body or These what are called Yamanaka factors, for example, you know, these were, uh, uh, um, Nobel laureate type of stuff, right?”

Claim: The speaker says that newly made cells can be used to heal different conditions in the body and cites Yamanaka factors as Nobel-laureate-level work.

“So I think having a space of perturbations essentially of things that you can do to a cell, this can be a very rich thing. It could-- can go on as much as actually changes in, in, in a gene, which would be like a full knockout or something more subtle, right?”

Claim: The space of possible cellular perturbations ranges from full gene knockouts to more subtle changes.

“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?”

Claim: A virtual cell model should account for a broad space of influences on cell state, including human genetic variation.

“And also combinations thereof, of course, and maybe, you know, disease happening, time, whatever, right? That then inf-- uh, um, impacts that.”

Claim: A virtual cell model should account for combinations of perturbations and may also need to represent disease and time.

“And when I first saw that people are able to take essentially cells out of a body, out of a cancer cell or whatever, right, and let them then grow into small multicellular systems that actually sort of start, connect, and do things, I was super, super impressed.”

Claim: The speaker was impressed that cells taken from a body, including cancer cells, can grow into small multicellular systems that connect and carry out activities.

“So, you know, people can now, for example, take some cells, put them in a dish, and then grow little mini brains, essentially a set of neurons that essentially really start connecting, talking to each other.”

Claim: Researchers can grow mini-brains in a dish from cells, with neurons that connect and communicate.

“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, a bunch of local collaborations with fantastic partners that actually start beating.”

Claim: Researchers have demonstrated that mini-brains grown in a dish can perform computation.

“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, a bunch of local collaborations with fantastic partners that actually start beating.”

Claim: Heart cells can be grown into organoids that begin beating.

“Like you have beating little heart cell clumps in a dish, right? And yeah, then of course, once you have these, and the same thing is I spoke about brain. There's even simpler systems like the colon, where this works very well.”

Claim: Organoid methods work well for simpler systems such as the colon.

“There's a whole bunch of cancer systems, but you know, also for, let's say, if you wanna study diseases with hard to acquire material, let's say you wanna look into the genetics of some type of, of neurodegenerative diseases or so. Like typically it's very hard to get access to tissue, right? So you take stem cells from the sk-- you take cells from the skin, differentiate them, make-- turn them into these mid little brains, and then you can potentially study some, some things there.”

Claim: Organoids may enable disease research, including research on neurodegenerative diseases, when the relevant tissue is difficult to obtain.

“So I think this concept of replicating little bits of organs, it's why they're called organoids in a dish, is very fascinating. There's all kinds of limitations obviously.”

Claim: The speaker finds the replication of small parts of organs in a dish using organoids fascinating.

“Often immune cells, immune system is missing. There's like no, uh, maybe feeding them or something like-- So there's all kind of problems, right? But it's much, much more than just a cell line, which is like one cell type.”

Claim: Organoids have limitations, often including the absence of immune cells or an immune system and insufficient feeding or support.

“Often immune cells, immune system is missing. There's like no, uh, maybe feeding them or something like-- So there's all kind of problems, right? But it's much, much more than just a cell line, which is like one cell type.”

Claim: Organoids are more complex than cell lines, which consist of a single cell type.

“So that's I think bit 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.”

Claim: Organoid modeling remains a research frontier; there is no agreed optimal approach, partly because available data are insufficient.

“So that's I think bit 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.”

Claim: Researchers have not agreed on the optimal modeling approach, partly because available data is insufficient.

“So, so like pathology, there's like so much samples available, right? And now increasingly we sort of do a bit more spatial omics. 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.”

Claim: Pathology provides many samples, and the use of spatial omics is increasing.

“So, so like pathology, there's like so much samples available, right? And now increasingly we sort of do a bit more spatial omics. 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 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, right?”

Claim: Methods for representing spatially connected datasets, including at the organoid level, are less established than methods for dissociated cells.

“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 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, right? 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.”

Claim: Cell types are human-imposed categories that help people understand biological complexity; nature does not inherently care about those categories.

“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 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, right? 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.”

Claim: Cell types are human-imposed categories that help people understand biological complexity, rather than categories used by nature itself.

“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.”

Claim: There is not yet a widely useful conceptual metaphor for local niches of cells comparable to the cell-type concept.

“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.”

Claim: Researchers do not yet have a metaphor for local cellular niches comparable to the cell-type metaphor.

“Maybe a little bit, but also in H&E, right? Whenever I re-- I, I talk and see what pathologists do, it's often very qualitative and it's maybe not on that level yet.”

Claim: Pathology work using H&E is often qualitative and may not yet provide the desired level of detail.

“Maybe a little bit, but also in H&E, right? Whenever I re-- I, I talk and see what pathologists do, it's often very qualitative and it's maybe not on that level yet.”

Claim: Pathologists' work with H&E is often qualitative and may not yet provide the level of characterization needed.

“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, to sort of think a bit how we deal with this.”

Claim: Researchers may need a categorization language for local cellular structures, although such categories may not be necessary for computational methods.

“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, to sort of think a bit how we deal with this.”

Claim: Researchers may need to develop a categorization language for local cellular niches, although such categories may be less necessary for computational approaches.

“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 more, much more subtle, like graphs or something like that.”

Claim: Current methods represent spatial cellular data as cell distributions, but richer representations such as graphs may capture it more subtly.

“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 more, much more subtle, like graphs or something like that.”

Claim: Current approaches represent the data as cell distributions, but richer representations such as graphs may capture more subtle structure.

“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.”

Claim: Reducing inductive bias can make a larger dataset applicable to a problem and improve model accuracy by broadening the data distribution.

“Um, 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, you have to induce-- introduce a lot of inductive bias to get enough statistical power to make good inferences.”

Claim: When data is scarce, introducing substantial inductive bias may be necessary to achieve enough statistical power for reliable inferences.

“And it-- 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 simply so that we know that it works on planets as well," right? And, and so that I wanna 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 gonna help to reduce all the complexity so that we're able to actually make meaningful-- gain meaningful understanding from these really large data sets.”

Claim: Scientists seek simplified descriptive models that capture general principles and generalize beyond the examples used to establish them.

“And I think-- And there I, I, I need to be very clear before we may go into the phil-philosophy. I think literature has really clearly shown over, um I think now decades at, at least for also our very main data sets, that if we build in a priori inductive biases too strongly that we might lose predictive power, right?”

Claim: Research over decades indicates that imposing strong a priori inductive biases can reduce predictive power, including on major datasets.

“And I think-- And there I, I, I need to be very clear before we may go into the phil-philosophy. I think literature has really clearly shown over, um I think now decades at, at least for also our very main data sets, that if we build in a priori inductive biases too strongly that we might lose predictive power, right?”

Claim: Research over decades, including on major datasets, indicates that strong a priori inductive biases may reduce predictive power.

“So it comes at a cost. So and if you're now, let's say, a physician and you just wanna... 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.”

Claim: Using strong inductive biases can come at the cost of predictive power.

“So it comes at a cost. So and if you're now, let's say, a physician and you just wanna... 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.”

Claim: For prediction, a physician or patient should want the best available predictive model.

“So it comes at a cost. So and if you're now, let's say, a physician and you just wanna... 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 wanna 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 wanna have the best. And this is, I think, I think not, not entirely clear.”

Claim: The speaker thinks that for prediction a physician or patient would want the best-performing model, whereas a person receiving treatment might prefer a somewhat less accurate but more understandable model.

“Maybe if you're getting treated, you wanna 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 wanna have the best. And this is, I think, I think not, not entirely clear.”

Claim: For treatment decisions, a person may prefer a somewhat less accurate model if it is more understandable.

“Maybe if you're getting treated, you wanna 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 wanna have the best. And this is, I think, I think not, not entirely clear.”

Claim: The speaker considers it not entirely clear whether the best predictive model should always be preferred.

“And I think in our field with this perturbation space scale, the cell scale I think is big, perturbation ba- space scale, like the stuff that we do to it not being as big, I don't think we're yet in this super data-rich regime.”

Claim: The speaker believes their field is not yet in a super data-rich regime, considering the scale of cells and perturbations in their work.

“And I think in our field with this perturbation space scale, the cell scale I think is big, perturbation ba- space scale, like the stuff that we do to it not being as big, I don't think we're yet in this super data-rich regime.”

Claim: The speaker's field is not yet in a very data-rich regime for perturbations, even though the cell-scale space is large.

“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.”

Claim: The speaker thinks hybrid models might still be the best approach in their field.

“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.”

Claim: At sufficiently large scale, data can outperform human-designed concepts incorporated into models.

“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.”

Claim: Hybrid models may remain the preferred approach in the speaker's field, given its current data 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.”

Claim: The speaker's lesson from large-scale machine learning is that, at some point, data outperforms human-designed concepts added to models.

“And I, you know, at least 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.”

Claim: Early AlphaFold models explicitly incorporated rotation invariance into their learning process.

“And I, you know, at least 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.”

Claim: Early AlphaFold models incorporated rotation invariance directly into their learning process.

“They sort of built 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.”

Claim: Later AlphaFold versions learned rotation symmetry from more data, allowing greater flexibility than imposing rigid physical constraints.

“They sort of built 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.”

Claim: The speaker says later AlphaFold versions learned rotation symmetry from more data rather than relying on a rigid imposed constraint, making the approach more flexible.

“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 data-based ones where you then a posteriori put the interpretation on top.”

Claim: The speaker expects that, in the future, most widely used models will be flexible and data-driven, with interpretation applied after training.

“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 data-based ones where you then a posteriori put the interpretation on top.”

Claim: The speaker predicts that, in the future, most more widely used models will be flexible and data-driven, with interpretation applied after training.

“And then there were a whole bunch of papers coming up where they're basically saying, you know, hey, maybe even more or less linear model were doing a good job. And actually in our SCGen paper early on, we had already some of these benchmarks showing you it's not so, not so super easy and simple task actually to improve on that.”

Claim: Early benchmarks in the SCGen paper showed that improving on relatively simple models for the evaluated tasks was not easy or straightforward.

“Um, I think, I think there's a truth to it that particularly the current data state for some things, let's say you wanna predict the effect of a CRISPR in a cell and you might not need the most fancy model.”

Claim: Given current data, predicting the effect of a CRISPR intervention in a cell may not require a highly sophisticated model.

“Whereas if you ask a bit or if you wanna predict a cell type or something like that, you know, for that mostly logistic regression is fine. But if you wanna predict something a bit more difficult, for example, in our NicheFormer paper, which is a foundation model for spatial omics, 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.”

Claim: The NicheFormer paper introduced a task of predicting neighboring cells' gene-expression states from a cell's gene-expression state.

“Whereas if you ask a bit or if you wanna predict a cell type or something like that, you know, for that mostly logistic regression is fine. But if you wanna predict something a bit more difficult, for example, in our NicheFormer paper, which is a foundation model for spatial omics, 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.”

Claim: Logistic regression is mostly sufficient for predicting a cell type.

“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.”

Claim: Communication between neighboring cells can leave a remnant of one cell's state in another cell's state.

“And actually this is harder to do. And if you sort of have these harder tasks, you need these stronger, more, 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 rich type of things?”

Claim: Harder prediction tasks require stronger, more nonlinear models and analysis of subtler data.

“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, how we learn these things.”

Claim: The best methods for representing and learning biological data through self-supervised learning have not yet been established.

“There's a lot of copying of what happens in NLP type of LLMs, right? And like masked autoencoder. So what's 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.”

Claim: Researchers are adapting NLP and large-language-model approaches such as masked autoencoders, but continue to debate loss functions and representations, partly because the data is often tabular.

“So even just like tabular foundation models are now sort of being tried out. There's a lot of JEPA type of approaches now because there's like we have one thing where we use contrastive learning. So it's just people playing around with the losses but also with the representations.”

Claim: Researchers are exploring tabular foundation models and JEPA-style approaches, experimenting with loss functions and representations such as contrastive learning.

“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 model.”

Claim: Reviewers now scrutinize whether scaling laws are properly demonstrated, a feature the speaker says was missing from early foundation-model papers.

“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, and samples that you have this sort of scaling behavior, then I think you have this buy-in into these models.”

Claim: Researchers have not yet convincingly shown that greater model complexity is necessary for scaling; demonstrating scaling behavior across examples and samples can build confidence in the models.

“And for these of course there's not so much at least inductive biases and issues. So I, I do think we're gonna, gonna hit this regime actually quite soon and we do see this also in a few papers.”

Claim: The speaker expects the relevant scaling regime to be reached quite soon and says that a few papers already show it.

“So for example, in our contrastive learning paper, we call it SC concept, we do show already quite, quite, quite strong scaling behavior. In the paper we had something for going because we compared with Geneformer going to maybe thirty million cells.”

Claim: The SC concept contrastive-learning paper reported strong scaling behavior and compared results extending to about 30 million cells with Geneformer.

“But on, on Hugging Face now we have a three hundred and fifty million cell model with, I think going up to maybe a hundred million parameters. So really going to the big scale for us then we do see nice scaling behavior really starting to emerge.”

Claim: The speaker says a Hugging Face model trained on 350 million cells has up to about 100 million parameters, and that clear scaling behavior is beginning to emerge at this scale.

“Yeah. I m- I mean, you know, the, the ultimate thing of course we wanna do is we wanna be doing this perturbation prediction better 'cause 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.”

Claim: The speaker's ultimate goal is to improve perturbation prediction, which the speaker says people currently understand as a virtual cell.

“Yeah. I m- I mean, you know, the, the ultimate thing of course we wanna do is we wanna be doing this perturbation prediction better 'cause 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.”

Claim: The speaker does not claim that pure representation learning is the right approach to perturbation prediction.

“Sorry for making the 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.”

Claim: Foundation models typically learn a state space, while predicting how the state changes under an action or perturbation requires more world-model-like approaches.

“... all over machine learning conferences in last year. But de facto, a perturbation prediction model like a virtual cell is kind of a world model in terms of transcriptome or other type of omics space, right?”

Claim: A virtual-cell perturbation-prediction model functions as a world model over transcriptome or other omics state space.

“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? And then sort of predicting the next state X prime or something like that.”

Claim: A perturbation model maps a current state and an action to a predicted next state, rather than mapping the current state alone.

“And I think for this you might need other combinations. So what we typically do, we learn the representations with, for example, scConsort. But then we put a generative model for the push forward along the action.”

Claim: The speaker's approach is to learn representations, for example with scConsort, and use a generative model to predict the state transition caused by an action.

“So for this we use something like flow matching or something like that. And in our cell flow paper, this is not just, uh, got accepted for nature methods, we actually really show that these, uh, approaches are much better generating the effect of, of, of, of perturbations, but they're very much dependent on a good latent space.”

Claim: The speaker's group uses methods such as flow matching for generative perturbation modeling; the CellFlow paper reports substantially better generation of perturbation effects, but the methods depend heavily on a good latent space.

“And, you know, in this scConsort latent space, in this case, for these three hundred and fifty million cells, we actually went into multi-species. So you have now like sixteen different species where you can potentially look for similarities.”

Claim: The scConsort latent space for the 350-million-cell model includes 16 species, enabling potential comparisons of similarities across species.

“Suddenly here you can maybe ask some evolutionary questions that you couldn't before. 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.”

Claim: Cross-species, cross-tissue latent spaces may enable new evolutionary questions, though the speaker says his group cannot yet do this well and it is outside their core expertise.

“... 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, a famous open competition on-- with the, through the Arc Institute that was a sort of a glass of like cold water on everyone's, you know, faces.”

Claim: Researchers should understand their baselines well because failing to do so can waste time and lead them to fool themselves.

“It sounds such an, like such an obvious learning, right? Do your baselines properly, but in a field that is rushing towards the gold pot 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.”

Claim: In research pursuing a virtual cell, researchers may not have handled baselines carefully enough, and subtleties such as holdout design were not always fully accounted for.

“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 regression, right?”

Claim: Some near-state-of-the-art models for cell-type prediction are essentially logistic regression models.

“So for such a simple categorization, you know, this is actually not, not so bad. There's one thing called Cell Typist, for example, which is doing a good job, and I think we had a paper maybe five or six years ago where we had the ta-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 organ, suddenly it gets a bit more subtle because, you know, sort of cell types reoccur in different organs and immune cells sort of su-subtle change depending which tissue they are.”

Claim: For cell-type prediction within an organ, deep learning is not needed, according to a paper by the speaker's group.

“This is true if it's within an organ. If you sort of go cross organ, suddenly it gets a bit more subtle because, you know, sort of cell types reoccur in different organs and immune cells sort of su-subtle change depending which tissue they are.”

Claim: Cross-organ cell-type prediction is more subtle because cell types recur across organs and immune cells vary somewhat by tissue.

“Suddenly, you know, there you need to do a bit more. And we-- like for us, there was one paper called scTab, like a tabular foundation model, uh, predictive model type of thing. We actually showed that this non-linearity was really adding be-beyond the, the baseline.”

Claim: The speaker's group showed that nonlinearities in its scTab tabular predictive model improved on the baseline for more complex cell-type prediction.

“So that was sort of f-f-for, for that class. For the perturbation models, which then you were asking Brendan similarly, like early on many analysis, particularly Perturb-seq, were done with classical linear models in pseudo bulk.”

Claim: Early perturbation-data analyses, especially Perturb-seq analyses, often used classical linear models on pseudobulk data.

“We had then our, our scGen that was like something around twenty nineteen, where we started introducing linear shifts, but in the latent space, so sort of nonlinear encoded, and we always compared with subtle, uh, sort of simpler, uh, linear shifts.”

Claim: Around 2019, the speaker's group introduced scGen, applying linear shifts in a nonlinearly encoded latent space and comparing it with simpler linear-shift methods.

“Uh, we've been running a regular Europe's competition on, uh, prediction problems in single-cell data, sometimes for integration data, also for perturbation prediction, um, that was very well attended.”

Claim: The speaker's group has run well-attended European competitions on single-cell prediction tasks, including data integration and perturbation prediction.

“And the cherry in the, uh, on, 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.”

Claim: The Arc Institute competition was named the Virtual Cell Challenge rather than the Perturbation Prediction Challenge.

“That sound a bit cooler, right? We picked up very well and but of-- with metrics that are robust were maybe, but maybe not, not that easy. So for ac-ac-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 in order to really quantify maybe subtle things that you might not have thought before.”

Claim: Open Problems required participants to submit their own metrics alongside the provided metrics so that subtle aspects of model performance could be quantified.

“The Arc one, you know, sort of fixed a bunch of metrics that weren't even so, so single cell specific, and they could be just ex-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.”

Claim: The Arc competition used fixed metrics that were not necessarily tailored to single-cell data and could be exploited.

“The Arc one, you know, sort of fixed a bunch of metrics that weren't even so, so single cell specific, and they could be just ex-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.”

Claim: The model ultimately declared the Arc competition's winner was based on flow matching.

“So based on, on this idea that, you know, people pro-forward with this flow, with this cell flow type of idea. I think shows you that some nonlinear things are, in this case, needed so that the actual real winner, so if you sort of remove the, the exploitation, was actually something that is, this is taking some, some nonlinearities.”

Claim: After metric exploitation was removed from consideration, the actual winning approach in the Arc competition used nonlinearities, suggesting they were needed for that task.

“And I think nowadays we do see that this linear or additive type of models are not cutting the, the, the cake for more complex things.”

Claim: Linear or additive models are insufficient for more complex tasks in current practice.

“But for, I think even nowadays, if you go in a simple, a few cell lines and you have sort of, uh, um, gene knockouts, then particularly when the effects are a bit subtle or these sort of more complex models are sometimes a bit over fitting, and that's why linear baseline is always very good.”

Claim: For gene-knockout prediction in a small number of cell lines, complex models can overfit, particularly when effects are subtle, making a linear baseline useful.

“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 wanna do the most fancy deep learning thing, but please, please, you know, have the baseline."”

Claim: The speaker advises students to establish a baseline even when they want to use an advanced deep-learning method.

“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 baseline, not even linear model, like better than the mean.”

Claim: For drug-response prediction, a simple baseline is to test whether a model can outperform the mean prediction.

“Sometimes not as hard if there aren't so much noise in the data, and this stuff is super noisy.”

Claim: The data in these prediction settings are extremely noisy.

“... some of the models as well. So it, it's clearly it's not only metric hacking, but also just that the asking the wrong question leads to the wrong answer.”

Claim: Asking the wrong research question leads to an incorrect answer.

“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.”

Claim: The speaker expects agents eventually to test many combinations of possible algorithms.

“In fact, we actually do have some collaborations with some of, of those sort of AutoML type of labs to try this out. So maybe what is left sounds a bit too harsh, right?”

Claim: The speaker's group collaborates with AutoML labs to test automated exploration of algorithmic combinations.

“But like where maybe our biggest ideas go into it is actually this metric formulation, RJ, right?”

Claim: The speaker identifies metric formulation as an area where their group contributes its most important ideas.

“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.”

Claim: Research funding models affect institutions’ flexibility, the teams they can establish, and the long-term vision they can pursue.

“So, um, Helmholtz, this is not just Germany, but I would say maybe Europe's largest what's called research organization in terms of funding.”

Claim: The speaker describes Helmholtz as possibly Europe’s largest research organization by funding and says it is not exclusively German.

“It has 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 fed-federal money to put into a deep research question.”

Claim: Helmholtz focuses on areas including health, materials and information sciences, and physics, using federal funding to pursue deep research questions.

“I've been building up what's called the Computational Health Center at Helmholtz Munich, which I think has now something like maybe forty-five group leaders and professors and maybe five hundred to six hundred people now.”

Claim: The Computational Health Center at Helmholtz Munich has, by the speaker’s estimate, about 45 group leaders and professors and 500–600 people.

“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 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 sort of to, to intersect and to go virtual human and so on.”

Claim: The center’s researchers work across biological and health-data scales, including cellular research, DNA foundation models, protein design, patient-cohort representations, and genetics.

“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 doing DNA foundation models.”

Claim: The speaker believes his institution is the largest academic center for health-AI research in Europe.

“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 sort of to, to intersect and to go virtual human and so on.”

Claim: A major vision at the center is to integrate its research areas toward creating a virtual human.

“But having said that, the key thing is, in contrast to universities, that even though we all have affiliation to universities, mostly to the Technical University or the University of Munich, the two top-ranked universities interestingly 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 we want to have sort of this translational thing in there.”

Claim: The institution has university affiliations but no undergraduate teaching; it mainly hosts PhD students and postdocs and collaborates with universities and clinics on translational research.

“And increasingly, and that's maybe an interesting link to-- 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.”

Claim: The speaker compares the institution’s model to the U.S. NIH model, while noting that it is distributed nationwide rather than concentrated in one location.

“And maybe the one interesting, um, um, um, link is that increasingly the government and this sort of realization also in, in Germany is, is sort of dawning that there needs to be much stronger transfer component, not just by going into translation to universities, but also in, into sort of value creation outside of the academic realm, which is obviously going so more, more to startups and so on, which I think the Boston environment has always been one of the key examples.”

Claim: The German government is increasingly recognizing the need to transfer research into value creation outside academia, including through startups.

“You know, I have been also interested." So, like, everyone's sort of thinking about this startup ideas and so on. And I, I kind of felt that's a-- it's a good environment, and this is something I think our PhDs are now more and more doing, and, and we're, we're sort of learning there.”

Claim: The speaker viewed the Bay Area as a favorable environment for developing AI-and-biology startup ideas.

“And I think also in terms of sort of value creation, Munich in particular is actually at least in continentally or maybe the prime place now for VC capital, but also sort of deep tech, uh, particularly bio and, and defense area, um, sort of really exploding.”

Claim: The speaker describes Munich as a leading continental European location for venture capital and says deep tech, particularly biotech and defense, is rapidly expanding there.

“Welcome. Let me know if you come, come by Munich. So, um, I think traditionally Europe, but I would say definitely Germany is, is quite bureaucratic and processes are far from as digital as you would expect in different places.”

Claim: The speaker characterizes Germany as traditionally bureaucratic, with processes that are less digital than expected in other 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, the particularly startup creation and implementation is, I think, more than competitive in Germany versus Europe.”

Claim: The speaker considers Germany’s startup creation and implementation more than competitive relative to other European countries.

“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, the particularly startup creation and implementation is, I think, more than competitive in Germany versus Europe.”

Claim: The speaker says processes are faster in Germany than in the United States, while not claiming that the United States is a model of full digitization.

“It's actually sort of the, the, the, the prime number and in particular like some hotspots, Berlin, but in particular Munich, sort of the, I th- I think is, is the capital for, for, um, um, highest number of startups.”

Claim: The speaker identifies Berlin and especially Munich as German startup hotspots and suggests Munich may have the highest number of startups.

“And I think even nowadays, if you look into European Union, there's equally as many startups being created as in US. The thing is, where Europe in particular also Germany drops off is at the scaling part.”

Claim: The speaker says Europe, particularly Germany, lags in scaling startups.

“And I think even nowadays, if you look into European Union, there's equally as many startups being created as in US. The thing is, where Europe in particular also Germany drops off is at the scaling part.”

Claim: The speaker says that the European Union and the United States currently create equal numbers of startups.

“So, you know, early VC and also sort of, um, 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, I think, not, not closely as present.”

Claim: The speaker says Europe has early-stage VC and non-dilutive funding, but fewer large, risk-taking VCs available to finance startup scaling.

“So I've, I've, I saw this again and again that, you know, if you wanna sort of really go into the, um, the sort of bigger tickets, sort of series A and so on, series B, you-- people start going more into US also for that.”

Claim: The speaker says founders seeking larger rounds such as Series A or Series B often look to the United States for funding.

“I, I don't know if, if you heard that the European Union tries to regulate a whole bunch of stuff, and that can be, uh, sometimes an advantage. You know, it's kinda cool, like, you know, all across Europe we can now go with a mobile phone.”

Claim: The speaker believes some European Union regulation can be advantageous, citing the ability to use a mobile phone across Europe as a positive example.

“And you might have heard that everyone needs to have USB-C, uh, things now for their mobile phones. These are 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 make things a tad more complicated.”

Claim: The speaker considers the EU requirement for mobile phones to use USB-C a positive development.

“And you might have heard that everyone needs to have USB-C, uh, things now for their mobile phones. These are 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 make things a tad more complicated.”

Claim: The speaker believes the AI Act makes things more complicated.

“Luckily, not yet for research, but you know, for data access across hospitals and so on, there's some, uh, you know-- and we were talking about, about health deep tech type of stuff, right? So there are some, um, systems that are, I think, fantastic in Europe, and I'm thinking about Finland, you know, Fingene, um, Denmark and so on, with this sort of very detailed description- Of the population in terms of sampling, getting medical records connected far better than-- I mean, of US is also similar mess like, like Germany in terms of sort of getting all the systems sort of to talk to each other.”

Claim: The speaker says Finland and Denmark have systems for population sampling and medical-record linkage that are far better than Germany's, while the United States is similarly disorganized to Germany in connecting systems.

“Luckily, not yet for research, but you know, for data access across hospitals and so on, there's some, uh, you know-- and we were talking about, about health deep tech type of stuff, right? So there are some, um, systems that are, I think, fantastic in Europe, and I'm thinking about Finland, you know, Fingene, um, Denmark and so on, with this sort of very detailed description- Of the population in terms of sampling, getting medical records connected far better than-- I mean, of US is also similar mess like, like Germany in terms of sort of getting all the systems sort of to talk to each other.”

Claim: The speaker says the AI Act has not yet complicated research.

“Like when patients, you know, sort of go to, 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 are 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 in like maybe next two years, these things need to be sort of going across boundaries.”

Claim: The speaker says it is very difficult to build patient trajectories across different locations in the United States or Germany because medical records are difficult to connect across those locations.

“Like when patients, you know, sort of go to, 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 are 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 in like maybe next two years, these things need to be sort of going across boundaries.”

Claim: The speaker says the European Health Data Space is intended to enable health data to be shared across borders, perhaps within the next two years.

“You need to sort of start building these, these health record type of combinations. So I do think the visions are starting to happen, but the scale is, is not as fast.”

Claim: The speaker believes health-record integration systems need to be built and that the vision is beginning to materialize, but not at sufficient scale or speed.

“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.”

Claim: The speaker believes that having the best team for its purpose is the most important requirement for a tech startup.

“And, you know, we have professors and, and PIs that really think about this. So for example, for, for the department that I coordinated, I said, "Guys, every, every PI, every group leader would expect you at least to have one, one, one spinoff in, in, in the next, uh, three years.”

Claim: The speaker set an expectation that every PI and group leader in the department he coordinated would have at least one spinoff within three years.

“Let's see how that works," right? But you sort of need to sort of get people a, 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.”

Claim: The speaker believes startup creation is increasingly driven from the bottom up, with students pushing for it.

“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, I would say pushed even more in the forefront that we can now start actually guiding experiments.”

Claim: The speaker says computational approaches are increasingly prominent and can enable researchers to guide 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, 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.”

Claim: The speaker believes computational startups can be effective at generating data.

“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, 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.”

Claim: The speaker expects computational and data-generation work to be commercialized in a next generation of startups, not only published as research papers.

“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, 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.”

Claim: The speaker says perturbation world models can be used to guide and drive experiments.

“And you do see a few of these popping up where you actually get the lab and the loop really implemented.”

Claim: The speaker says a few companies are emerging with lab-in-the-loop systems implemented.

“It, it's interesting, right? Like if, if you looked at earlier, and of course, this is what people always reflect to what OpenAI, Anthropic 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?”

Claim: The speaker says OpenAI and Anthropic initially lacked a clear way to make money, and that some people might argue they still lack one.

“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 wanna do with it.”

Claim: The speaker says some heavily funded startups have described plans to build a biology foundation model or decode biology without specifying a concrete application.

“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 Seurat and so on, right?”

Claim: The speaker has repeatedly observed that platform-based ideas are highly attractive.

“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 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.”

Claim: The speaker says it is not straightforward to turn a platform idea into a revenue-generating business.

“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 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.”

Claim: The speaker says a virtual-cell platform might generate consulting revenue, but investors, particularly in Europe's more conservative venture-capital ecosystem, more readily understand a business that eventually develops a drug.

“So I saw so many of these platform companies pivoting at some point to, "We make a drug." And yeah, at the side, of course, we 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, um, maybe you have some other counting sums.”

Claim: The speaker has seen many platform companies pivot toward drug development while retaining a platform business, using discovered drug candidates to support the case for further investment.

“And there's a few companies that are sort of still stick with the scaling up of, of the AI business, but it's not as clear, uh, how you would sort of directly turn this into a viable, sort of really scalable bu-business model, and if there's even room for many of these.”

Claim: The speaker says it is unclear whether companies focused on scaling AI can directly build viable, highly scalable business models, and whether the market can support many such companies.

“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, yeah.”

Claim: The speaker says there is concern that Astra 2.0 might already perform some of this work.

“Yeah. Th-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, just bio, so we had a whole bunch of other disciplines.”

Claim: Fabian Theis says he helped establish the Munich School for Data Science, which includes disciplines beyond biology.

“And there we had this supervisor model that there's always two PIs needed to come together to make a proposal, then sort of proposes to the students in a graduate student type of model, which has this, this Y.”

Claim: In the Munich School for Data Science's supervisor model, two PIs must jointly make a proposal for graduate students.

“So there's like a method-y, more domain one, right? And then a supervisor student. And sometimes we had two students under the Y model, so Y plus Y equals H. You get it? But, but I, I think this idea is, it's somewhat even nowadays a bit trivial, right?”

Claim: The Munich School for Data Science's Y model pairs methods and domain supervision, and sometimes places two students in the arrangement, forming what Fabian Theis describes as “Y plus Y equals H.”

“But I, I thought it was good to bake this in so hard so that students Really, I usually go to two lab meetings. One lab meeting where they need to really sort of learn and understand some deeper biological problem, understand the language maybe of a medical setup or something like that. Then they go to a computational-- Typically, they're all computation data science type of students, right? But then they go to the computation one, they sort of get pushed the method aspect of it, and thereby sucking up a, 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.”

Claim: Fabian Theis designed the training model to have students attend both a biology- or medicine-focused lab meeting and a computational lab meeting focused on methods.

“Then they go to a computational-- Typically, they're all computation data science type of students, right? But then they go to the computation one, they sort of get pushed the method aspect of it, and thereby sucking up a, 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.”

Claim: Fabian Theis says quantitatively trained computational students can formalize loosely specified biological problems as metrics and develop them into models and papers.

“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.”

Claim: Fabian Theis says his group has operated this interdisciplinary training model for approximately eight years.

“But 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 approach 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.”

Claim: Fabian Theis says the interdisciplinary approach generated many papers and exposed him to biological systems, including the lung and brain, that he had not previously understood.

“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.”

Claim: In Fabian Theis's lab, students who develop projects around biological datasets become the lab's experts on those specific biological areas, alongside a partner.

“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-cell 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 that particular domain where, you know, these essays work very well.”

Claim: Fabian Theis says his group's research has expanded from RNA to other data types, including single-cell proteomics and spatial observations.

“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 that particular domain where, you know, these essays work very well.”

Claim: Fabian Theis says additional measurement modalities create interesting research questions in domains where the assays work well.

“Yeah, I mean, this is my natural inclination is to like gathering cross-domain stuff. It-- I think it's al- often ripe for good, good generation of really good and interesting ideas.”

Claim: Fabian Theis thinks gathering ideas across domains often produces good and interesting ideas.

“But I think once, you know, students are done in, in, in, in, in Germany, Europe, I think in general, but better, I think in Europe in general, people sort of typically do a master's first before they go into a PhD, right?”

Claim: Fabian Theis says students in Europe typically complete a master's degree before beginning a PhD.

“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 this 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, in terms of domain, but then they can also play around a bit with the method side.”

Claim: Fabian Theis says an interdisciplinary graduate-school setup lets students explore different research domains and methods through different proposals.

“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 this 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, in terms of domain, but then they can also play around a bit with the method side.”

Claim: Fabian Theis says students who have completed a master's degree may know some science but may not yet know exactly what interests them.

“So I think that was kind of attractive. 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.”

Claim: Fabian Theis had the impression that trying different research areas and methods made students very good data analysts and modelers.

“And then, uh, 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 their, in their biomedical domain into pharma, but you know, some obviously stayed in research, some went more into some sort of startup direction, some went into very pure AI ty- type of work.”

Claim: Fabian Theis says graduates from the program entered diverse careers, including biomedical roles in pharma, research, startups, and AI-focused 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 the bio part that NVIDIA is now building up and there are similar other of the big companies.”

Claim: Fabian Theis says NVIDIA hired two of his people within one month, a few months before the conversation, for roles involving computation and its developing biology work.

“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 that improved maybe two percent in a Europe paper type of thing," but you also sort of had the real sort of more complex, deeper application that, uh, like, you know, I had to help find as, as a use case.”

Claim: Fabian Theis thinks job candidates benefit in interviews from describing both a model improvement and a complex application or use case they helped identify.

“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 that improved maybe two percent in a Europe paper type of thing," but you also sort of had the real sort of more complex, deeper application that, uh, like, you know, I had to help find as, as a use case.”

Claim: Fabian Theis thinks interdisciplinary experience gives students an interesting professional profile.

“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.”

Claim: Fabian Theis considers guiding experimentation one of the most exciting current research frontiers.

“So where before you-- we built these human cell atlases, we had sort of this mapping aspect. Then we talked about the world 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 the scientific method, you know, sort of putting our hypothesis in the loop and then doing something with it.”

Claim: Fabian Theis describes a progression in scientific work from mapping biological systems, to predicting them, and now toward 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 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.”

Claim: Fabian Theis says scientific papers often conclude with a model, sometimes a concept figure and limited validation, and researchers then typically begin another project rather than continuing a continual-learning cycle.

“But now, you know, I think we can actually really loop back more and more. This is happening very much so in, uh, molecule design or material design and so on.”

Claim: Fabian Theis believes research can increasingly use iterative feedback from models to guide further experiments, a pattern he says is already occurring in molecule and material design.

“You know, sort of the, 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 yet where 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?”

Claim: Fabian Theis says that, although some large pharmaceutical companies promote experiment-computation integration as a general theme, there are still few examples of repeated cycles linking experiments and computational models.

“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.”

Claim: Fabian Theis says iterative integration of experiments and models is difficult to establish because choosing the optimal next question from a model is not straightforward.

“And we actually brought out, I think it should be on, on BioRxiv today or tomorrow, a new thing we-- which we call Lab Compass, which is an inverse generative model that actually allows you, once you have a forward world model or whatever, to tell you what you need to do in order to reach a certain state.”

Claim: Fabian Theis says his group developed Lab Compass, an inverse generative model that, given a forward world model, identifies what needs to be done to reach a target 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.”

Claim: Fabian Theis says the Lab Compass project involved experimental collaborators and used three or four iterative lab-in-the-loop cycles to improve 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 that you actually get to convince experimental people to put their lifetime in, because these experiments often take some time, right?”

Claim: Fabian Theis says it is difficult to persuade experimental researchers to invest time in follow-up experiments suggested by computational predictions.

“To actually buy into and then 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 not as high. But, you know, if you like say, you know, you need to do this new cell line, you need to add these very strange things.”

Claim: Fabian Theis says experimental follow-up may be easier to secure for molecule-related predictions, where the stakes are lower, than for interventions such as creating a new cell line or adding unusual substances.

“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. That needs culture, and it might be the same person doing it.”

Claim: Fabian Theis says establishing iterative experiment-model research is difficult, requires an appropriate culture, and might be handled by one person.

“A lot of people then start throwing around robotics and so on as well, but I, I think that's a tad too early 'cause the, the cycle times in terms of our assays are still a bit on the long side. But I do think this will come.”

Claim: Fabian Theis considers robotics somewhat premature for laboratory experimentation because assay cycle times remain long, but expects robotics to become relevant in the future.

“And, you know, obviously the, the ultimate vision being that we want to RL everything. We want to do RL also in, in, in, in this loop. It's a bit out there, but you know, I think we're, we're slowly getting there.”

Claim: Fabian Theis's long-term vision is to apply reinforcement learning throughout the laboratory experimentation loop, although he says this remains a distant goal and progress is gradual.

“At the moment it's more maybe active learning.”

Claim: Fabian Theis says the current approach to guiding experiments is closer to active learning than to the long-term reinforcement-learning vision.

“... for having me. I, I think these links are important, this sort of looking beyond what maybe you do like in your own sort of smaller disciplines and, you know, these discussions help me a lot.”

Claim: Fabian Theis considers it important to look beyond one's own narrow discipline and says cross-disciplinary discussions help him.