Unlocking Cells' Secrets: Diffusion, Deconvolution, & Discovery with Siyu He, author of Squidiff & CORAL

In this episode of the Cognitive Revolution, we hear from Siyu He, a postdoc at Stanford specializing in biomedical data science. Siyu discusses the implications and methods behind their recent AI-driven biological research papers, Squidiff and CORAL. The conversation explores the use of AI models to analyze complex cellular systems and disease mechanisms, focusing on transcriptome and tissue sample analyses. Squidiff aims to simulate cellular transcriptomes to predict outcomes of various conditions, significantly expediting traditionally lengthy and expensive biological experiments. CORAL Project extends this by integrating different levels of biological data, enabling a more comprehensive understanding of tissue structures and cellular interactions. The discussion also delves into the challenges of using synthetic data for validating AI models and the potential acceleration of scientific discoveries through AI in biomedical research. The episode encapsulates the interplay between AI and biology, highlighting the future possibilities and current limitations of this innovative research front. Squidiff: https://www.biorxiv.org/content/10.1101/2024.11.16.623974v1.full.pdf CORAL: https://www.biorxiv.org/content/10.1101/2025.02.01.636038v1.full.pdf SPONSORS: Oracle Cloud Infrastructure (OCI): Oracle's next-generation cloud platform delivers blazing-fast AI and ML performance with 50% less for compute and 80% less for outbound networking compared to other cloud providers. OCI powers industry leaders like Vodafone and Thomson Reuters with secure infrastructure and application development capabilities. New U. S. customers can get their cloud bill cut in half by switching to OCI before March 31, 2024 at https://oracle.com/cognitive Shopify: Shopify is revolutionizing online selling with its market-leading checkout system and robust API ecosystem. Its exclusive library of cutting-edge AI apps empowers e-commerce businesses to thrive in a competitive market. Cognitive Revolution listeners can try Shopify for just $1 per month at https://shopify.com/cognitive NetSuite: Over 41,000 businesses trust NetSuite by Oracle, the #1 cloud ERP, to future-proof their operations. With a unified platform for accounting, financial management, inventory, and HR, NetSuite provides real-time insights and forecasting to help you make quick, informed decisions. Whether you're earning millions or hundreds of millions, NetSuite empowers you to tackle challenges and seize opportunities. Download the free CFO's guide to AI and machine learning at https://netsuite.com/cognitive PRODUCED BY: https://aipodcast.ing CHAPTERS: (00:00) About the Episode (03:37) Introduction and Guest Welcome (04:00) Setting the Big Picture Context (04:37) Exploring the Squidiff and CORAL Papers (08:31) Understanding Transcriptomes (11:17) Single Cell RNA Sequencing Technology (15:32) Motivation Behind Squidiff (Part 1) (17:14) Sponsors: Oracle Cloud Infrastructure (OCI) | Shopify (19:41) Motivation Behind Squidiff (Part 2) (25:56) Training Data and Model Architecture (Part 1) (31:38) Sponsors: NetSuite (33:11) Training Data and Model Architecture (Part 2) (37:18) Diffusion Models in Biology (46:07) In Silico Experiments and Applications (54:25) Clarifying the Validation Process (55:36) Validation Strategies and Real Data (58:26) Challenges in Modeling and Predictions (01:02:14) Accelerating Research with AI Models (01:07:31) Future Directions and Collaboration (01:10:46) Introduction to CORAL Paper (01:13:09) Spatial Transcriptomics and Proteomics (01:17:10) Challenges in Integrating Spatial Data (01:31:53) Synthetic Data and Model Validation (01:36:42) The Future of AI in Healthcare (01:43:31) Outro SOCIAL LINKS: Website: https://www.cognitiverevolution.ai Twitter (Podcast): https://x.com/cogrev_podcast Twitter (Nathan): https://x.com/labenz LinkedIn: https://linkedin.com/in/nathanlabenz/ Youtube: https://youtube.com/@CognitiveRevolutionPodcast

Speaker names marked ? are inferred from speaking time and the show’s known hosts — they are not voice-verified.

• 0:00Nathan Labenz?

  Hello and welcome back to the Cognitive Revolution. Today, I'm excited to share an in-depth conversation with Siyu He, postdoc at Stanford in biomedical data science and lead author of two notable recent papers that use intricate machine learning systems to shed light on two of the great questions in biology. What causes what at the cellular level? And how do the interactions between individual cells give rise to higher-level tissue development and disease? The first paper, Squidiff, allows us to predict how cells will respond to perturbations by conditioning a diffusion model trained to predict transcriptome states, or the level of gene expression that a cell exhibits across more than 30,000 individual genes, on semantic

• 0:45Nathan Labenz?

  directions derived from other experiments. This approach, based on a very modest amount of data, can save researchers months that would otherwise be needed to culture specific cell types, and can also shed light on intermediate states that can be extremely challenging to measure directly. The architecture will be familiar to students of image generation models, and also reminds me quite a bit of the Mind's Eye models used to reconstruct images from brain scan data that we've covered on previous episodes. This is another data point, showing just how general-purpose today's architectures really are. How models with "intuitive understanding of important problem spaces" like transcriptomics

• 1:25Nathan Labenz?

  can perform many orders of magnitude faster than physics-based simulations, and ultimately, how much future discoveries in biology will be made initially in silico and then confirmed by much slower and more costly wet lab ground-truth experiments. The second paper, Coral, makes it possible to combine low-resolution tissue data and high-resolution single-cell profiles into a single, integrated view. This is a particularly intricate system that uses graph neural networks and ultimately de-convolves the lower-resolution tissue data into a detailed cell-by-cell picture, thus enabling closer and more insightful study of the tissue as a whole.

• 2:08Nathan Labenz?

  This project makes notable use of synthetic data and reminds me of the project we covered a few months back from Michael Levin and his co-authors, which had connected gene, drug, and disease information to make previously unknown connections.

• 2:21Nathan Labenz?

  Perhaps most telling, CUN team completed both of these projects under the supervision of Professor Jane Zhao, who we recently had on the show to talk about the virtual lab project, which used a team of AI agents equipped with specialist biology models to develop new nanobodies capable of treating emerging COVID variants, and also the protein language model interpretability project, Inter PLM. That all of this work could come from one group in such a short period of time supports the broad theme that everything is working in AI. And with just a bit more acceleration, perhaps driven by the inevitable generalizations and scale ups of the techniques we discussed today, we really could see a centuries worth of biology progress in just the next decade.

• 3:07Nathan Labenz?

  As always, if you're finding value in the show, we'd appreciate it if you'd share it with friends, post about it online, or leave a review on your favorite podcast platform. Of course, we always invite your feedback and suggestions, too. You can reach us via our website, cognitiverevolution.ai, and you're always welcome to DM me on your favorite social network. Now, I hope you enjoy this deep dive into cutting-edge architectures that are advancing the frontiers of cell and tissue biology with Stanford researcher Siyu Hu, lead author of Squid Dif and Coral.

• 3:39Nathan Labenz?

  Siyu Hu, postdoc at Stanford in biomedical data science and lead author of the recent AI for Biology papers, Squid Dif and Coral. Welcome to the Cognitive Revolution.

• 3:50Speaker 01

  Thank you, Nelson. It's great to be here and very excited to be here to share my research to a broader audience.

• 3:59Nathan Labenz?

  Cool. Well, we've got a lot to unpack. So let's maybe just start with a little bit of big picture context setting. I think the audience, and I can also just speak for myself, you know, I've been obsessed with AI for the last few years, like studying it intensively. I feel like I have a pretty good general understanding of the landscape. On the biology side, you know, first of all, the landscape is even bigger and more complicated. And, you know, I haven't spent nearly as much time in it. So I always kind of like to just start off by trying to get a sense for like, where you think we are in the big picture? And where you think like this work is in the big picture? And I guess maybe just...

• 4:40Nathan Labenz?

  To you up a little bit more by saying the Squidiff paper is about the transcriptome state of the cell, operating at like the level of a cell, which is quite interesting. And then the coral paper is even zooming out to a little bit bigger scope of analysis than that and looking at like samples of tissue. And I have kind of had this general sense that like the grand challenge in biology is maybe a couple of different things. One big one is just figuring out the like super intricate web of what causes what and you know, what up regulates what, what down regulates what, how does an intervention ultimately lead to, you know, a change in outcome?

• 5:22Nathan Labenz?

  And unfortunately for all of us, you know, that's not only super complicated, but also seems to work at multiple levels of scale. So we have like, you know, the, the ability at this point to do a pretty good job of saying, well, here's a sequence, you know, what does that translate to in terms of the shape of a protein? And we're even starting to get decent at like, how do these proteins potentially fit together? But it still seems like we've got a long way to go when it comes to how does all that stuff add up to the function and evolution over time of a cell? And then even probably more work to go when it comes to understanding how that, you know, aggregates up to the tissue, the system, the whole organism level. So I've been watching this space where you're operating because it seems like that's just

• 6:06Nathan Labenz?

  a really critical frontier. And you've taken, you know, a bite out of each of those problems with these two papers. But with that kind of preface, tell me more about how you see the big picture landscape, how you understand the grand challenge and how you have decided to orient yourself with this work toward, toward solving them.

• 6:25Speaker 01

  Yeah, I think the major motivation about these two projects is like whether we can utilize the AI models really to apply on to the current technology of biology. And the goal is like trying to understand the cellular systems in trying to understand the disease mechanisms and further we will provide like. The the treatment strategies for the disease and how to really like using the AI models to the such kind of like different sort of data sets like in in technology like freedom call like single cell RNA transcriptomics that I mostly addresses like in the first paper and then the spatial transcriptomics that I mostly addresses in the second paper.

• 7:08Speaker 01

  So these are two of different studies, but like they try to into like a very similar goals like whether can we use AI models to accelerate those understanding of the cellular activity and then to figure out like the disease related stuff and then their responses to the drug uh drug treatment.

• 7:28Nathan Labenz?

  Yeah, perfect. Okay. Well, let's get into the first one then. The the first one that caught my eye was Squidiff. Am I saying that right? Squidiff Squidiff? Interested actually in where the name came from, just as a bit of trivia before getting into the science itself.

• 7:44Speaker 01

  Yes, uh right. So basically because I have a paper first like we called a starfish. And then I was thinking of like, oh, probably I want more models like they call it the sea animal. So we have like squid, but like it's a it's a diffusion model. So that's why we call it squid if but it's also can I think like it's a simulation related sort of like stimuli related quantitative influences of of the transcriptomics like using the diffusion models. Yeah. So you can trying to actually guide in the whole ideas about like what model is and coral. Yeah, is a second one, but it's also a name around like the sea pants.

• 8:24Nathan Labenz?

  Gotcha. You got a you got a theme. You've got a whole cinematic universe in the making of of various models.

• 8:30Speaker 01

  Yeah, hopefully I'll get more.

• 8:33Nathan Labenz?

  Tell me about transcriptomes in terms of, you know, maybe like.

• 8:41Nathan Labenz?

  What we need to know, you know, that's, that's not obvious. What I know coming into this is the transcriptome is basically a measure of what's, what genes in a cell are being actively transcribed at any given time. So my, my loose understanding is, and it's amazing, we can do this now at the level of a single cell, you know, the, the technical achievements of, you know, the, the resolution of some of these techniques are really super amazing. So we can take literally a single cell. I assume, but you correct me if I'm wrong, that this is like a totally destructive process. So this is the end of this cell. We can then pull out, you know, the RNA that has been transcribed from the DNA at any given time. And we can just look

• 9:24Nathan Labenz?

  at how much of these different RNA segments are there. And that basically tells us the sort of state of the cell at that moment in time. Like what is actually, what are the active processes that are happening in that cell? But that's a rookie explanation. So you know, tell me what I'm missing from that, or what more I need to know.

• 9:45Speaker 01

  Yeah, I would suggest maybe a good way to start about this story is to think of like the central dogma of the biology. We call like the cell is the smallest functional unit of the living systems. And then the key information in our cells, like we heard chromosomes, they carry the DNA. And then looking like human bodies, like all the cells or like almost all the cells that they share the same DNA, what makes them different is the transcript transcription processes that like the DNA transcribed to RNA and then RNA will further translate it into protein and then protein will fold it into specific structures like what the alpha fold is trying to predict it. And then like they will have various functions.

• 10:27Speaker 01

  So the process is like related to transcription is we could also like gene expression that is an indication of lot of like variants of the, of the cells in the human body. And one of my post advisors, Stephen Quick, has said like the cell is a bag of RNA and which actually indicated that the complexity of the cells are able to distinguish from the gene expression. Protein can

• 10:54Speaker 01

  Protein expression can distinguish the complexity too. But I think like current technology would say like it's still like maybe more prevalent in the transcription levels because like protein probably like a much noisy. Even we have like more and more advanced technology too for protein. Um. People are mostly like studying the transcript terms first. And which uh help us to understanding the cellular activity and disease mechanisms. Um. And the second um point that I also want to mention is that the development of the new technology, the single cell RNA uh sequencing. So I probably not go to details about what exactly um the whole process it look like. But the major thing is that we dissociating the tissues into individual cells.

• 11:40Speaker 01

  And then uh we will prob or like to measure the molecular expression, like the gene expression into the RNA, sorry, into the cells. And those mostly can profile like around like thirty thousands or sixty thousands of the genes, which is a very high dimension. And it's like also high support technology, which making AI models like a very good ways to address this this kind of data because usually we will have like a matrix sort of things. Uh. We have like columns and rows which are like cells and uh genes. And they are, they are kind of like indicating a lot of like things in individual cells. Yeah.

• 12:21Speaker 01

  So that's I think really a phenomenal technology came out and kind of like open areas to the quantitative biology. Yeah.

• 12:30Nathan Labenz?

  So thirty to sixty thousand different cell or sorry, different genes can be measured in just a single cell. Yeah, I mean, that is in an amazing level of resolution. And I think it's always helpful just to kind of consider these like the dimensionality of the inputs. This is true in in the AI, you know, architectures that we're going to get into. And so, you know, I guess just to be very clear, like the starting point for this whole process is a transcriptome is like thirty to sixty thousand numbers that represent the intensity of the expression of the individual genes and that's measured through the RNA that's that's pulled out of the the cell when it's going to be processed in this way.

• 13:14Nathan Labenz?

  Is there anything like, do we have any reason to think that there's like?

• 13:21Nathan Labenz?

  The transcriptome is like a complete picture? Or are there things that are sort of known to be missing from the transcriptome? You know, is there a way in which the transcriptome like doesn't tell you the whole story of what's going on with a cell?

• 13:36Speaker 01

  Of course, I think transcriptome, um, like though they will share a kind of the heterogeneity of the cells, like I think there's more information related. For example, we're looking on the epigenomics in the cells and like the chromosome things that happens in the in the cells that different across across the individual cells. And also there is like protein, as I said, which are also distinct with the gene expression, but like sometimes they kind of share similarity. And beside that, there will also like mitochondrial related or like repsomere related things, I think like cells are very complicated. That's why people haven't really figured out like what exactly

• 14:18Speaker 01

  they're going on and the transcriptome is probably like, uh, is one of the the data that we can get a whole genome level. So it's a very good way to start and to understanding like where the cell is going. But I would say like there's more other informations that can be also considered into the model design and the whole story design as well.

• 14:40Nathan Labenz?

  Cool. Okay. So now getting into the squid diff architecture. Let's first maybe just describe what it is that we're trying to do here. Cause one thing that will become clear as we go through this is this, you know, both of these papers are pretty carefully designed architectures. Actually more complicated than like what we see even from like frontier language models today, which is an interesting fact, like, you know, this is, and I want to get into as well, like, why has the bitter lesson maybe not fully come for this sort of work yet? But let's just start off with, you know, what is it that you hope to create a model to be able to do

• 15:26Nathan Labenz?

  with squid diff and then we'll like actually unpack the the strategy and the parts of it.

• 15:31Speaker 01

  Okay, yeah, sounds good. So, right, so I think the motivation of this project is like, because usually I'm a well-lab person that trying to getting those single cell sequencing from like, you know, the whole processes of experiments, which is like very kind of painful because you need to waiting the cells like to grow and then you prepare whole processes that is like you're waiting for like months. But also it's very expensive when we're trying to do the experiments, but also it's very expensive at this time point. What I'm thinking, can I actually like using the generative AI models to creating those like like a virtual or like digital creation of the transcriptomes?

• 16:11Speaker 01

  So this is the first motivation and then I realized there is a potential like we can maybe manipulate that the the major information is like for example, because we have like lot of condition generative AI models. So what if we just like manipulating these major conditions, we probably not like

• 16:33Speaker 01

  like ob- ob- like observe like exactly, but like we know what they are. And there is maybe a way that we can try then, and then that's why the we designed a squid if. And so it's not just like maybe creating the transcriptomics, but also dresses like some sort of interesting questions in biology that whether there is a perturbation happen like chemical perturbation or like a gene perturbation that making the cells like to be different and then what they look like. In in most of the cases in experiments, like we are not able to getting them immediately and also a very large scale of the data. So that's, I think, the way it can help the the whole fields.

• 17:16Speaker 02

  Hey, we'll continue our interview in a moment after a word from our sponsors.

• 17:20Speaker 01

  Yeah.

• 17:21Nathan Labenz?

  So you said it takes potentially months to culture cell type and condition of interest. And then you said it's expensive to do the sequencing. Can you cash that out? Is there like a, you know, established dollar amount? Or is it still like a relatively bespoke process where it's just like your time is the expense of doing the single cell sequencing processes?

• 17:47Speaker 01

  Yeah, so I think it depends on the exactly tissues or like the samples that you are dealing with. And usually I was working on the organoids, which is a like engineered tissues derived from like iPSC, for example, like induced proportional stem cells. We were making like those engineered brain tissues or like vessel tissues and doing like disease modeling. So the whole process is like usually take months. I know like some of the brain organoid, they need like even a year or like more, more times to create. And definitely like there will be some sort of mistakes like during the, you know, the experiments and you will need to like start over again.

• 18:32Speaker 01

  And that's why kind of like very painful. And this is like initiative that I really want to have a models that can help me to, you know, just take over the, the task. And yeah, but also there's another concern, you know, in some of the cases, there is a limitation that not experiments cannot really perform. So that's why I think such a model will be very useful.

• 18:57Nathan Labenz?

  Okay, so let's do just a little bit more on like the actual runtime, how this thing gets used. You might have a particular cell type and you might have a question of interest, like what happens if I apply this drug or I apply some radiation? Or you could correct me if I'm wrong on the assumption here, but like upregulate or downregulate a particular gene, something like that. And you basically want to know what happens? And the, the form of the answer to that question is ultimately a transcriptome. What the system ultimately outputs is a transcriptome, not that's measured, but that is, you know,

• 19:41Nathan Labenz?

  generated or predicted by the AI model. What does that do for you? So I mean, correct, again, always correct anything that I'm missing there. But then what does that do for you? Like when you now have a transcriptome put out from the model? What do you do with that?

• 19:58Speaker 01

  That's a good question. Yeah, usually like we have the transcript terms from the data like we are using kind of like some sort of analysis of the single cell RNA sequencing to understand what's going on. And I think the first step usually is like to annotate those cells. And usually people will do with the dimension reduction and also like using the manifold folding methods to seeing like how they look like, maybe like clustering technology as well, like to kind of like classify the cells into different cell types. And then because we also have the gene expression into individual cells. So we knowing like what specific genes are they expressed?

• 20:41Speaker 01

  So we know like the related signal pathways they are, and then we are understanding like what's going. So that's usually the way like people analyze the whole kind of data. Yeah, but there's multiple ways, I think, in the in the field that trying to really quantitatively understanding the whole of stuff.

• 21:01Nathan Labenz?

  And how well do we understand that sort of thing? Like if if you have a transcriptome, regardless of its source, like it could have been measured in, you know, an actual wet lab experiment, or it could have been generated now by a model.

• 21:17Nathan Labenz?

  What can we, with what level of sort of granularity or, you know, specificity or confidence can we say like what's actually going on in that cell? Okay, sure, you've got, you know, we've got a number now for thirty to sixty thousand genes. But could you say like, oh, that's definitely cancer? Or like the cancer is now, you know, it's not cancer anymore? Or, you know, it's healthy or it's not healthy or like, you know, far more? I mean, obviously you can do a lot more specific things because you can get down to this sort of, you know, gene by gene level analysis. But what is our ability to sort of aggregate that up into the things that really matter using all those techniques that you just listed?

• 22:01Speaker 01

  Right, usually like I would say if the cells are similar, like they are in a similar state, they will have like similar sort of groups of the genes that express, like we call like co-expression sort of things. And in that case, like we are able to like classify or like grouping these cells into different cluster. Each cluster will kind of usually like corresponding to a cell type. And um in that case, like we are able to um knowing like what is the high like high valuable genes in in the individual groups. And um other sort of methods people usually consider is the marker genes um to to really uh distinguish like what's going on.

• 22:43Speaker 01

  So like marker genes are those genes like identified by the whole biological community that they recognize is is a major um or like hallmarks for uh the the cell types. So then for example, there is cancer related markers. Uh people if they identify that, um they will consider, okay, they probably is like a cancerous uh um stage. And similar to what the um the pathology people that they also trying to uh using the marker like the protein. And and antibodies like to figure out like what's going on in the tissues, yeah.

• 23:18Nathan Labenz?

  Okay, cool. So there's clustering, there's these kind of indicator genes that are like reasonably well understood. And now I think we can get into, you know, how it is that you're going about training a model to to come up with these things. So I guess maybe for starters there, what does the data set look like in terms of like where does it come from? You know, how much of this stuff is sort of out there in the community for you to just grab that other people have published? How much do you, if any, have to actually gather yourself? Um and just like literally in terms of just, you know, megabytes, gigabytes, whatever, like how big

• 24:00Nathan Labenz?

  is the data set that you were working with in this project?

• 24:05Speaker 01

  Yeah, so that's a good question. Usually, I would say it really depends on the projects. And as I said, like, we we have like a matrix sort of uh format of the data that like sell by genes. And we have like thirty to, you know, thirty to uh six, like sixty thousands of the genes, which is home gene level. And then we will have the also thousands or like even actually mi- millions and billions cells. Uh but it depends on projects, I think. And in in my in my models, like um I taking those like maybe small data sets and and specific data sets instead of like the whole things altogether, uh which is because also to consider uh the complexity of the models and

• 24:47Speaker 01

  uh and and that's the the current development things. But I would also thinking of to thinking of like using a a a more uh like a larger scarce of data sets in into into the models and making it like more from foundation related. And in terms of the data resources, we usually are kind of, I mean, the community is like shareable. They will kind of like pub- publish their own data into uh some some, you know, space. And then we will uh utilize those public available data. And at the same time, we're also collaborating with Web Labs and then they uh help us to creating those like novel, maybe

• 25:30Speaker 01

  um new data sets or maybe have like validation as well, like for the model. So it's very kind of flexible. And I would say the data source is the one of the major challenger right now uh in the field. Because we will want to have like high quality data and also have, because most of the data, they they are good. But like sometimes they're missing some informations and we're making the models hard to develop to, you know, making them consistent.

• 25:59Nathan Labenz?

  So, and by the way, are the, is it annotated at all? I I understand when you described the grid of cells and transcriptomes, you know. What what I'm understanding there is basically one vector for every cell that indicates the strength of activity for each corresponding gene, you know, in the in the long list of thirty to sixty thousand. Is there additional like metadata around like what kind of tissue it was or you know what you know the patient's condition or any anything along those lines that you also tap into here? Or is it literally just pure kind of raw information?

• 26:40Speaker 01

  Yeah, we see like they have metadata and it's about like the cell types which annotated maybe by myself or by the other authors and they would say maybe annotated by publicly already. And it has meta information for them whole like what kind of tissues it belongs to and what disease stages it has. So I think there is like multiple ways to kind of like label these datas already. Yeah, that's one of the requirements in the model.

• 27:10Nathan Labenz?

  Gotcha. Okay. So did you say, sorry if I missed it, but like, is it a sensible question to ask like, how many individual cell transcriptomes constitute the training data for this project?

• 27:24Speaker 01

  Yeah, so right now I would would say like the model will taking any kind of large amount of the datas to be trained. I'm like not providing a trained model, definitely like we require people to provide their own trained data sets. So it's very depends on like how data sets you are. And then usually I think I just testing how long it it need to and around like 5,000 cells in in kind of that's kind of a data sets or it just requires like 15 minutes to learn to train. Yeah.

• 27:59Nathan Labenz?

  Wow. So you can get this working with as few as 5,000 different cells?

• 28:05Speaker 01

  Yeah, yeah. But I think too few cells will making the models like to be underfitting or maybe overfitting to.

• 28:14Nathan Labenz?

  Yeah, okay, wow. I mean, that's a small number. So do I understand correctly? Like, that's kind of where it starts to work? And in practice, you used more and in like fullness of time or, you know, maybe future work, you could use like a lot more still?

• 28:32Speaker 01

  Of course. So I think like the future direction of like the this project is I think we wanna building a bigger one, which can kind of like cover all kind of, I mean, for example, we are working on organoids as like a case studies in the paper, but we are thinking of like in what evolve all kind of like organoids together. So we will have like foundational models for the organoid things, but we can also making a furnish model for tissue or like real human tissues too. So it's very depends on how the design of the exactly data sets we want to use. But the model is very powerful to taking any kind of data, I would say, and flexible to addresses the needs of the people of the researchers.

• 29:17Speaker 02

  Hey, we'll continue our interview in a moment after a word from our sponsors.

• 29:23Nathan Labenz?

  Okay, so then let's talk about the architecture a little bit. I guess, you know, one kind of, you know, just super brief summary of the structure of the architectures is basically two core parts to it. There's the VAE, the variational autoencoder, and then there's the diffusion model portion.

• 29:44Nathan Labenz?

  The question occurred to me like before even getting into how those work. Because certainly like folks in our audience will be even more familiar with the transformer as you know, probably the whole world is like starting to become quite familiar with the transformer. Why not use a transformer for this? You know, if if somebody just came to me out of nowhere and said, let's say you want to predict, you know, the next transcriptome state for a cell, given like a current state and a, you know, a perturbation or whatever. That would be my my first instinct would be like, yeah, we could probably I do with the transformer, transformers seems to work for everything. So why did you end up going a different way with this project?

• 30:22Speaker 01

  Well, first thing I wanted to maybe apologize is there is like mistakes in the in the first manuscript that is not actually the VAE is a semantic encoder that connecting with the conditional, the diffusion models. So we have like the major, two major things. One is the DDIM models and then other is like semantic encoder. And the, the, the major generative processes is that we're taking the, the Z semantic as variable of semantic and that's kind of from the encoders and then we take a Gaussian noise X T and using the denoising processes of the diffusion into a, a final versions of transcript tones, yeah.

• 31:08Nathan Labenz?

  So I'll, I'll dig into that and unpack it a little more, but just for intuition sake, like, why not a transformer architecture? Why this more like, you know, complicated and, you know, less, less familiar architecture?

• 31:23Speaker 01

  Yeah, so transformer and diffusion models are both like very popular models in as a generative AI model. I would say like there's already a lot of applications in our medicine and with, with the transformer as well and with the diffusion. So, so the difference is, is that the diffusion model maybe takes like more stochasticity of the data. Cause you know, like the generation things is like more flexible in that case. And they will kind of maybe modeling the more complex distribution. And in terms of transformer, I think it would be powerful. And also there is a lot of similar works that are trying to generating

• 32:04Speaker 01

  the related stuff, but I think the difference is, is the import of the model is a little bit different. Because usually the transformer will take the sequence data, you know, like the sequence of the, the genes usually like in our area that we're trying to ranking the genes first. So it's not like a con values of the data, but ranking the genes and then the output is also the ranking of the genes, which is probably different with like what the original data look like. And then the transformer based models, like a lot of foundational models right now, they're trying to getting a, a embedding of the data is in their further to use those embeddings to.

• 32:50Speaker 01

  Kind of doing downstream tasks like classifications and prediction task. Right? So, so I think they are kind of like different in this time point. And another reason that I wanna using diffusion is that like I real I noticed at the moment that while I started the project, the diffusion model is not really utilized in the generation of the single cell transcriptomics that why I'm trying to really consider like it explore the the ability of that diffusion models and conditional models that really can manipulating the conditions to generating the virtual cells. Yeah.

• 33:26Nathan Labenz?

  Gotcha. Okay. So on the architecture itself, I mean, you kind of described it briefly a minute ago, but I'll try to take a shot at describing it. I'm actually not quite sure of the distinction between the VAE and the semantic encoder. You can, you know, clear that up for me in a second, but either way, when we've seen this, you know, on the on the feed, you know, I've looked at lots of different architectures over time. And actually, maybe the one that this reminds me of most might be MindZi, which was a project partially out of stability AI where

• 34:09Nathan Labenz?

  they trained a system and it was again kind of a, you know, carefully designed system to reconstruct the images that somebody was looking at based on the FMRI scan of their brain at the time they were looking at it. And that also involved a diffusion model and it involved basically creating a, in that case, a diffusion prior for the diffusion model to work from. And there's something pretty similar going on here where the first step is to try to understand what really matters in short, you know, in terms of the transcriptome of a cell.

• 34:50Nathan Labenz?

  In general, the VAEs, as I understand, are basically trying to pass the raw data through some narrow bottleneck.

• 35:03Nathan Labenz?

  Such that it can then be recreated out the other end capturing as much of the original data as possible. So typically, they're trained on a reconstruction loss, where the point is to say, you know, how can we sort of compress, abstract, you know, get this data into its its sort of most semantic form, still preserving, you know, all of the information that really matters so we can get back what we put in, but then use that highly semantic form for other things. So that's kind of part one. Anything I missed or anything you would, you know, complicate on that description?

• 35:39Speaker 01

  Yeah, I think that's a very good question. I would say like there are different kind of architectures about the models. And if we think of a combination of VAE with the diffusion, I would say like they are like the latent diffusion models. And there are also the related models that are trying to addresses of the application in transcript term as well. The difference is is that like these kind of models, like we they they usually like training separately, and then they kind of like using the VAE to do a dimension reductions and to getting a latent representation first and making the whole generation process is like more efficient. And the the way that Squidiff is trying to address is like to

• 36:24Speaker 01

  training a unified models that so they are able to train together. And then because the whole process is including the way to learning the noises, but also with the way to encoding the semantic information. So I would say it maybe have it has the powers to really like making semantic information extracted from the data and then also with the specificity that the model of the diffusions can do, and finally making a more realistic expression data with the with the cells.

• 37:01Nathan Labenz?

  Okay, that's an interesting note. So this system is trained entirely end to end under one loss function. And the big distinction there is typically in other contexts, the VAE is trained standalone separately and then can be mashed up with other things downstream as needed. Okay. So well, that brings us then to part two, which is the diffusion model, which I think folks will be fairly familiar with from an image generation context. I always like to get procedural with these sorts of things. And so the way I do that in the image context is say,

• 37:44Nathan Labenz?

  okay, let's say the goal is to train a model to say, given some starting point, what would this image look like if it was a little less noisy? And then with the conditioning, what would it look like if it was a little less noisy and a little bit more like this thing that we're trying to create? And what I think is really genius about that, of course, is it allows for mass production of training data because you can just take all the images on the internet and gradually add noise to them and then train the model to denoise them. And then eventually you get to something that can start with pure noise and just go, step, step, step, step, step, all the way to finally a high resolution image. And with conditioning,

• 38:28Nathan Labenz?

  you can get a high resolution image of something specific that you want. So a lot of same mechanics going on here with, obviously, the big difference being that the conditioning is coming out of this semantic encoder that is saying like, this is the kind of cell that we want in the sort of lower dimension, most semantic way possible. Now it's up to the diffusion model to map that back out into the fully detailed picture of a particular transcriptome. Again, what am I missing there? Or what more do I need to understand?

• 39:08Speaker 01

  Yeah, it's a very good question. Like, I think also because when I try to creating the model, so thinking of like, what's the differences between these models with the other image and base diffusion model. So I would say like, they are very different in with the data structure first. Um, because like usually most of the image and data, they are two dimension, um, have like X, Y, uh, you know, like directions, but for the single cell transcriptome data, we are dealing with uh, just one dimension data, like we the object is the cells, right, like the cell, uh, and then a gene expression. So with that, like, uh, we're not able to using most of the way that the diffusion model dealing with the image and like UNET or like image and base networks to

• 39:51Speaker 01

  uh, learning the noises. We uh, are using just like MLP to really like learning the noises, but we also consider the re- residue connections that are trying to incorporating the in- event of the the times and also the semantic features. And in terms of the semantic features, I would say it it can kind of like providing like implicating formations. So it's not just like cell types that included in this in the data. It will also have like other sort of, you know, the informations related, like for example, the the disease stage and the the conditions that related. So it's like a unified conditions, which we saw like maybe can be manipulated in the space and not in the gene space. Yeah.

• 40:35Nathan Labenz?

  Is there anything special about the noising process? I in in like images, it's pretty simple. I've seen that in contexts like protein folding, there is like a sort of highly specialized noising process that's required because like you can't just

• 40:56Nathan Labenz?

  like the naive approach of just adding noise to a protein structure just takes you to like something totally incoherent, I guess. And so there's like a a more sophisticated way of doing the noising. Does that apply here too? Or or are you able to just apply like a relatively simple noising process?

• 41:13Speaker 01

  Yeah, I think the the way to dealing with the noise is maybe quite simple. We didn't really consider too much of like the noising issues happen, but the differences we also consider is and uh you most of the expression data, they are usually in a single cell version. They are very sparse. So there are a lot of like zero in in there and we will need to kind of like uh take care of that. So I think what we're trying to address is is like we're not like studying all the genes related. We filter out some of the genes and probably not a really useful and we're looking on the high variable genes only. And another advances of the future model is that like they are not able just like to

• 41:56Speaker 01

  uh reconstructing those Gaussian or like a simple distribution. They are able to have creating like a very complex distributions that most of the gene expressions can have. So the model itself like have the powers to to model these.

• 42:14Nathan Labenz?

  So then in terms of like how we actually run a sort of in silico experiment on this, we've got the setup of this semantic encoder that takes in a raw transcriptome and converts it into a

• 42:33Nathan Labenz?

  more compact semantic, more semantic representation that abstracts away those details and you know, tries to get at what really matters.

• 42:42Nathan Labenz?

  It seems like the key thing to understand I found here as I was reading the paper is that you need to to do an in silico experiment to be able to ask the question, what would happen to this cell type if I did this sort of stimulus to it?

• 42:59Nathan Labenz?

  Is you need to be able to encode what that stimulus is in this semantic latent space. And so what I understand is happening to create that sort of direction is you're basically creating a delta or basically doing like subtraction in the in the semantic space between other pairs that you may have. So I think people probably have seen examples of this from language models or even image models. There's like the really famous example of if you have the Embedding for man and woman, for example. And then you subtract those.

• 43:40Nathan Labenz?

  You can kind of create the direction from man to woman or the direction from woman to man. Obviously, those are going to be, you know, the negative of one another in this latent space. Then you can do interesting things like if you have the embedding of king and you add the delta of man to woman, you find that, hey, that, at least in, you know, many experiments, looks like the embedding for queen. And so you have these sort of, these like delta vectors that can be synthesized from contrasting pairs and then applied to other starting points to get to, you know, some other place. So I understand that that's kind of the core mechanism, if I'm understanding correctly, as always, correct me.

• 44:20Nathan Labenz?

  But what I think you're doing in a lot of cases is saying, okay, I have cell type X and I want to know what would happen if I apply stimulus Y to that. But I would have, it would take me months to culture that cell type and then do all these experiments. But I do have cell type Z and I do have the ability to apply the same stimulus to cell type Z. And then I can, and maybe some of this stuff is just also out there as published data. You don't even necessarily have to run the experiment, you can tell me about that. But if I, if it's easier for me to do this experiment in cell type Z, then I can get that delta vector. And then I can go apply that delta vector to the cell type X. And then I can run this diffusion process.

• 45:05Nathan Labenz?

  I can see what the transcriptome would look like if I had hypothetically applied this stimulus to that cell type X without even necessarily ever having to touch the cell type X. So again, how'd I do and, you know, what I miss?

• 45:22Speaker 01

  Yeah, thank you so much for the introduction. I would say like, it's really, yeah, it's exactly what it is. It's like we using the vector operation and to address those manipulation with the semantic variable. But I would say like, this is assumption, actually. It's not exactly the real cases, I would say, because you do like the biological processes, they're complex and usually not very linear. But I would say like the latent variables probably like a more linear compare with to the gene expression space. So we're using the linear approximations to, you know, like really just, just like resemble to the real cases. And it can kind of like building with some, some certain like maybe a simple like conditions.

• 46:06Speaker 01

  For example, in, in my projects, like we are providing some scenarios, for example, this cell differentiation. Which like for example, day one, we have like IPSC, which is the stem cells. And then at the days, day three, we will have like a more, you know, like a more mature type of like mesoderm cells. And what is individual look like? And in that case, like if we have the, you know, the semi variables at day zero and then day three. We can actually using the linear interpretations to getting this like day one and day two. And to, you know, to, to kind of like creating the transcript terms look like in, in that days. And this is like sort of a, a way we, we did.

• 46:49Speaker 01

  And the second um implications, I would say like is the perturbation task. You know, like in some of the case, we will have like gene perturbation, like we making one genes to be maybe like up regulated or down regulated. And then, and in other cases, we're making gene B that kind of like regulated. And what, what's going on with like if we have A and B together? Um even like, I think the, the result will be like nonlinear. And the model of the diffusion model is able to model this like nonlinear, sorry, like nonlinear processes. But the latent variables are still able maybe to, you know, to manipulating by this like factor operations.

• 47:34Speaker 01

  Because I would suppose that the, the, the space of the latent will like more smooth and more structured. There's also a, a approximation, I would say. So it's a risky thing that we directly just like rely on the results from the, the model. So we need to kind of validating as well. But so far, um we apply on to the overnight cases and we see a very exciting result, like showing some, some cells. Uh, you know, it's a very trendy cells that during the development and during the differentiation into IP from IPSC into the endocelia. And into the fibroblast, into the bar vessel structures. Um, yeah. And then to to figure out this issue, I would say like I we will also consider variant

• 48:19Speaker 01

  of the script as our second project and instead of like, you know, just a simple semantic encoder or like simple semantic variables, we consider a a more time series analysis. Well, that kind of related to how we generating the video. So like video generated diffusion models say, we are training onto a a more complicated or like a data set that has a time series informations and we're learning implicitly about the the development things and then kind of taking more on the linearity of the data assets. Yeah.

• 48:59Nathan Labenz?

  Yeah, that's interesting. It's funny how some of these techniques that are developed for, you know, what were initially such whimsical, you know, kind of, I was following these things back in like 2021 when, you know, the legendary Twitter account Rivers Have Wings was, you know, training these diffusion models and they were just generating like very sort of bizarre, but sometimes like quite intriguing art. And, you know, it was like, is this useful for anything? It turns out it's actually like got direct application to biology. And it seems like we're maybe just a couple years behind, you know, obviously, it's the image

• 49:42Nathan Labenz?

  generation and video generation has gone like extremely fast to the point where, you know, now you, in many cases, would have a hard time distinguishing the video generator, certainly the image outputs from models from like real photographs and not quite there on video, but like sometimes it's hard to distinguish and sometimes it's still a little weird, but like, man, we've, you know, certainly come a long, long way. So to project that that could happen in the biological domain as well over the next two to three years is like, you know, nothing short of a paradigm shift, right? That, I mean, to be able to

• 50:21Nathan Labenz?

  to get that kind of accuracy out of a simulation, seems like it would just like totally upend the field. And it seems like it's almost certainly going to happen, right? I mean, what do you think is the, I mean, there is this issue of the nonlinearity, which is a fundamental one. So maybe I should just ask a little clarifying follow up question on like. I wasn't quite clear on like exactly to what degree this has been validated. I saw that there was like this I understood that there was this three day sort of cell transformation process. And this model, because you're sort of able to like essentially scale the delta, right? You could say, okay, I have a cell at day zero, something, you know, I can measure the transcript on now.

• 51:06Nathan Labenz?

  Now I can apply some perturbation. Now I can come back and measure it at day three after it's like differentiated into a different cell type or whatever. Now I have that like direction in space. Are you calculating like the two intervening days as like literally just like a third of that vector and two thirds of that vector? Is it as simple as that? And then how, if it is that simple, like, do I understand correctly that you also have like gone and actually taken cells at those one and three and one and two day time steps and like compared to see how well they they line up? And if that's all right, then like how well in fact do they line up in your, once you actually, you know, took the real measurement?

• 51:47Speaker 01

  Yeah, that's a good question. I think we have like different type of validation strategies in the model. For example, we have like some scenario cases, for example, in the diffusion applications. Sorry, in the differentiation applications, we definitely have the real experimental data from like day zero to day three. And then in the training processes, we only take days zero and day three, but we will head out those like one and two as like a test set. So when we, you know, like getting these models to be trained and we will have the results for one and three, as would you say about like manipulation with the thematic variable?

• 52:29Speaker 01

  And we got in those like predicted values for the gene expression at day one, two, and then using the real case to, you know, kind of compare and get the scores. And in terms of other cases, like the gene perturbation and drug perturbation, we also do the similar things. For example, we not, we having the whole, you know, the experimental data about like.

• 52:55Speaker 01

  Put it based on gene one and gene two, and then also one plus two, but we will had out those like one plus two. And then just testing whether they are exactly can predicting. And then the third strategy, I would say, is the real cases that we showing in the end of the paper. The blood-based organizing, so in that case, because the we need to culture, sorry, we need to culture the organites, right? So from the day zero to day eleven, if I remember correctly. So we cannot stop it, we cannot like collecting them like day by day, because the collection of this data to the single cell sequencing will actually destroy the whole sample.

• 53:37Speaker 01

  So in that case, we only take day zero and then we take the final days. And we will just using the interplay, inter, interpolation processes in to generating the intermediate stages. But but we have some like public available data that have the time series data as well. And we will compare whether we can identify similar biological things. So exactly, we discover some like novel cell states that we haven't really thought about. And in that case, we will also study about like what specifically genes are be triggered by this sort of differentiations.

• 54:19Speaker 01

  And then compare our findings to the experimental data from other publications. We will definitely seeing a consistent finding, so making the models like to be strong to like indicating that the ways to, you know, the powers to to finding these new things.

• 54:36Nathan Labenz?

  Yeah, okay, cool. So, and you mentioned too, it's an important point that like the manipulations that you can do in the semantic space are these like vector operations, which are inherently linear and.

• 54:52Nathan Labenz?

  That feels like a problem because we know how much non-linearity there is in biology. But you do have the diffusion process, which can capture nonlinearities. So it's like, sort of, you can hope that that'll work. And at least in some number of cases, it seems like it does. And I suppose it just kind of remains to be seen, like, in what cases might those sort of linear vector operations just fundamentally break down in the semantic space? Have you seen anything like that where you sort of say, yeah, well, we tried one? You know, you described the validation successes, but were there any validation failures where, you know, you're like, I guess this is just fundamentally not linear or somehow we're not capturing it in the semantic space?

• 55:36Nathan Labenz?

  Any instances like that?

• 55:39Speaker 01

  Well, I think for now, with the cases that we are tested are always like kind of efficient. I would say we're not seeing that. But we do see that the intermediate stages prediction is not as good as we thought. That's probably because we are modeling the semantic variables resemble about the factor. For example, iPSC to the mesoderm cells, we have different type of growth factors to be cultured cells. So we have actually different vectors or like different directions that making the...

• 56:23Speaker 01

  There's many ways that making the final stages of the variable. So the point is we're just doing an approximation and then making them to be like linear. So I would say if we have more information, for example, we learn what the semantic variables for the individual growth factors look like, the whole models will be treated or like being better a lot. But for now, we don't have that information. So this is the way that we try to address this. And that's a point that I probably mentioned, not probably related to this question, but we also realize the limitation of the models. For example, if we're having some data like...

• 57:06Speaker 01

  About like drug A and drug B and we want knowing like what's the result look like. We training on A and B and then we can kind of like using that vectors A and B to other cell types, right? So we will having A and B on any type of cell types and knowing what the responses is. But what what the case that we have drug C? Cause we don't know exactly the vector or like the semantic factor of drug C. So we can not like to model this like unseen drug. And to address these issues actually, because also I was asked by the reviewers in the paper. Uh we thinking of like consider another variant of the model. So we having uh now I like adapter that can encoding the drug informations.

• 57:51Speaker 01

  That is using, we using this mouse structure of the drug and also have this dosage information of the drug. So we can embedding this drug information related to structures and dosage. And then also concatenating with the semantic variable. We will have like a new version of the semantic variables. And then in that case, uh we even we don't have the application or like training onto the drug C. We can still kind of predicting the results if we having the drug components of the drug C.

• 58:24Nathan Labenz?

  Yeah, interesting. Okay. So do you think that this is, as it stands today? Something that can't effectively accelerate science like? The, I I understand how it would if it's accurate enough. I mean, like at least I think I do. The way I always kind of think about these things is, you know, long term, this can change. But like short term, we have sort of a finite amount of wet work experiments that we can do. We've got finite lab space. We've got finite, you know, grad students and post docs, you know, to be hands on and actually do the cell culturing. And there's like all this know how. And so the.

• 59:05Nathan Labenz?

  You know, for a short period of time, that throughput is relatively fixed and sort of what determines like how much actual scientific progress we make is like, do we run the right experiments that actually, you know, yield the interesting results that we can learn from? And so the great hope for these sort of in silico experiment models is like, obviously, they run, you know, orders of magnitude faster. I assume that maybe you could tell me like how long it takes to do one diffusion process, but I would assume it's like less than one one thousandth, maybe less than one one ten thousandth of the time that it would take to actually do the physical experiments. Can you do like, is there a single answer to like, how long does it take to do one of these?

• 59:51Speaker 01

  Yeah, I would say like this model will actually help in the researchers to accelerate in their research. I would say, you know, the, the way that we getting this like single cell RNA data, I probably like need at least like one weeks to, you know, from the dissociation to exactly getting these like raw, raw files about like the, the sequence. But if you want to getting the data sets like with other conditions, like can we just using this creative model? And using maybe like an hour to getting what's the, what's going on with other conditions look like? And then in that case, like people maybe have some intuition about like what's, what's the direction of the project look like?

• 1:00:35Speaker 01

  I would not say like trust the model like, you know, totally about like we can having some, something that information is happen. And, you know, in some of the projects, for example, we people develop those like organoids things. Cause organoid is a, a good way to being like a drug screening platforms. And so, so they, they are, they are studying about like different ways to creating those organoids. So I guess like if we having the vectors for the individual growth factors, we can able to like predicting what's going on with, with going on with the development of the organoids and also providing information about like those organoids respond to the, the drug, the different drug components.

• 1:01:18Speaker 01

  And that's one example, but in other cases for them, uh, for example, in a real case for them in the patient's level, you know, like for the pa- pa- patient with the cancer and uh, they want to know what kind of drugs are mostly useful for for them and it it's kind of like risky for them to just like randomly taking some drugs, right? So we need to testing and then I think it's showing powerful applications in in predicting how uh, these these cells respond and whether they will be useful for that patients or like will be have uh, side effects as well. Yeah, I guess we'll provide like, you know, efficient uh, outcomes and predictions

• 1:01:59Speaker 01

  and not necessarily to waiting, you know, long times to to see how the outcome look like.

• 1:02:06Nathan Labenz?

  Yeah, yeah, the the context of a single patient is really interesting, too. I hadn't even really considered that. I was just kind of thinking like general scientific inquiry. You can sort of run a bunch of these simulations and identify the ones that seem most promising and then prioritize your wet lab work accordingly. And that in and of itself seems like it could really accelerate the pace of discovery making that you could do. But then it becomes like, you know, like even even potentially more of a game changing technology at the point where you're applying it to an individual who, you know, has potentially their own like idiosyncratic uh, situation going on in in their own cells. So, yeah, that's cool.

• 1:02:51Nathan Labenz?

  Do you think that, so I guess next steps for this are like, scale up is obviously always one good candidate. And you mentioned the sort of, sort of separating out the the interventions to have like a distinct semantic encoder for them. And then you also mentioned the sort of video to image kind of time series to like single, you know, sort of before and after analog. I mean, I guess that's enough, right? You'll have plenty to work on to tackle those. Do all those, I guess, you know, one, one thing I'm always struck by when I talk to people who are doing this sort of work is, you know, the pace of like papers is just crazy.

• 1:03:34Nathan Labenz?

  And, you know, you're evidence of that because we're not just talking about one today. We've got a whole other one that we're going to spend a little time on, you know, very soon as well. When you look ahead to all those next steps, do you feel like it's just a matter of work? You know, at my software company, the software engineers used to say it's just a simple matter of programming. And what they would mean by that is like, it's going to take us some work. It's going to take us some time. We're going to probably find some things that are going to challenge us a little bit along the way. But like, we'll definitely overcome those. And in the end, like, if it's a simple matter of programming, like, we'll definitely be able to get it done. And it's just kind of a question of putting in the work and, you know, getting over the little stumbling points. Is that how you kind of think about these next steps?

• 1:04:16Nathan Labenz?

  Are they like, are you that confident that these are going to work? Not to say that there won't be little surprises along the way, but like, overall, would you say you feel very confident that these next steps will all kind of lead to better models and, you know, better predictions and all that good stuff that we want?

• 1:04:35Speaker 01

  Yeah, I, I think like it really depends on the whole community, like the collaboration between scientists and not just like the all our own projects, because you know, you know, we need like a amount of the data and

• 1:04:53Speaker 01

  I mean, people have their own expertise in their domain area. And then it will be very helpful to accelerate in the whole work to be, you know, into the next step together. So like, we need definitely biologists to understanding the whole things, like whole mechanisms of disease. And we need like machine learning person like to creating those models. And then we also need more people like statisticians and other even artists like to really kind of making our like, you know, are maybe our model, models and, and works like to be more fashion.

• 1:05:30Speaker 01

  Yeah, I would say it's really kind of like a working together things and the collaboration is very important. And another cases I remember, I think because we are dealing with the application AI models into healthcare and about medicines, we already care about like how the accuracy is and we uh we will be responsible for the the patients and you know it's it's serious that we we need to figure out like how the model exactly can can do and not in the opposite ways. So I think a lot of processes are required before it can goes to exactly um clinical available and testable. I think we need to experimental test and clinical test and

• 1:06:14Speaker 01

  a lot of like pre uh testing before exactly to publish the whole models to be applying to the real cases in the patients, I would say. But in the lab things, I think there will be more flexibility to, you know, um to testing because we will have ways to, you know, um getting more experiments to to testing the whole processes. Yeah, so it's all about the teamwork since.

• 1:06:38Nathan Labenz?

  Any last thoughts on Squidiff before we move on to talk about coral a little bit?

• 1:06:46Speaker 01

  Yeah, I think that's that's all the same. I I really appreciate that you ask so many questions. I think already catch the the key points of the model. Thank you so much.

• 1:06:56Nathan Labenz?

  Yeah, my pleasure. All right. Well, then let's do the coral paper a little bit. This one honestly is even like a little more challenging, I would say, for a non biology background person like myself to understand. But what intrigued me about it, I wrote down several things that we can go through that sort of intrigued me about it. But maybe the most fundamental one is this challenge of connecting different levels of resolution or different levels of scale together into an integrated understanding that in in so many, you know, ways that's like sort of the fundamental challenge of of science. It's like, you know, we going down to the very lowest level.

• 1:07:39Nathan Labenz?

  I don't know exactly what we want to call the lowest level, but maybe it's string theory, maybe it's quantum mechanics or whatever, right? We have like pretty good theories, certainly at the quantum mechanics level at this point of like how a very small number of particles will interact. And then we can go up and, you know, look at atoms and we can look at molecules and we can look at proteins and we can look at cells and we can look at, you know, all the way up, right, to like the economy. But There's often a gap between those where it's like, yeah, we don't know how to necessarily aggregate up these like smaller units of analysis into the bigger thing that we care about. You know, that's like even in economics, the micro and macro economics divide, right?

• 1:08:19Nathan Labenz?

  Like, we don't have a way of aggregating all the, you know, individual economic actors into an economy. So we sort of have this other like top-down approach that's like measuring these aggregate statistics and, you know, trying to predict where they're going. And it seems like we have a lot of, you know, I don't know how many layers you would identify, but it seems like we have that problem in biology, you know, at maybe the highest level of difficulty. And so I'm always intrigued by anything that I see that kind of tries to take multiple levels and, you know, integrate them and create a, an understanding that sort of bridges these like scale gaps. So with that prompt, maybe tell me like how you, you

• 1:09:04Nathan Labenz?

  know, came to this, sort of what the motivation is behind this work and, you know, what you're trying to do with it.

• 1:09:10Speaker 01

  Yeah, sure. I would say like this one is like more technical. So it kind of like addresses a technical issues when we're dealing with the spatial data. So maybe I will first like to start by introduce about the spatial transcriptomics that majorly like addresses in this project. So we have discussed about the single cell RNA sequencing data in the first paper. But as you know, like the single cell RNA data, they have one of the very obvious shortage is like we dissociate in the tissues and which means that we don't have the location informations of individual cells. This is very important because you know that cells are actually not individuals, like they are interacting with each other.

• 1:09:53Speaker 01

  And it's very important about like knowing what's the cell's neighbor is. So in that case, like we lost this information and it kind of like making the study of like architecture of the tissues or like the communications very difficult. So with that, people creating the new technology, we called the spatial transcriptomics and also proteomics as well. Totally, we can see like spatial all mixed data. Recently, you know, just came out like maybe seven years, seven years ago. And now there is like emerging of these current technologies about the spatial mixing. And then more and more data can came out and requires computational ways to understanding these data.

• 1:10:36Speaker 01

  And one of the in very significant issue of uh current uh ways to um understanding this thing is like, we, we have spatial mix, but like those data are not have the same resolution. And they are also not the same uh slices. You can consider as this, like you having a imaging of yourself, like a photo of, of yourself, like uh it's a gray scale, high resolution um images. And you have another eh photo of yourself, like at a different time point or like a different uh, you know, this eh like space and then eh and then it's colorful, but it's like a mosaic um um level.

• 1:11:20Speaker 01

  Can you actually under, can you actually making it into a, a better way? Like for example, you can getting a colorful, high resolution images. So that is actually the goal of these, these uh projects. Well, but I think the photo is different with what we dealing with, because the photo always have um like three channels, like RGB channel. But in other cases, we have a very high dimension channels. For example, uh the spatial transcutomics, as I said, we have whole genome level, thirty to sixty solars of the genes. But because of that, they will lost information of the spatial resolution. So it's like mosaic um and uh low resolution.

• 1:12:01Speaker 01

  But then in some other cases, we will have like high dimension uh or high resolution of uh the proteomics. So, so for the proteomics, like we, the resolution is in the space is very high, but the limitation is that we can only get limited number of the proteins to be measured. So in that case, like they are not perfect uh in both of the data sets. And then we will have limitations to really know, understanding or, or like. Or like dissecting the spatial structures with the low resolution of the space. And very difficult to distinguish those like cellular heterogeneity with low level of the proteins. Can we actually combine them together?

• 1:12:42Speaker 01

  And then we can have like a more comprehensive understanding about the tissue architectures and then how they are coordinated with each other and how they respond to the disease. And how can we provide a new strategies in the treatment? Yeah, that's the the whole legal of of of this project. So we wanted to address this issue with two of the tech uh technology. Um I would say there's new came out like have the high resolution uh in both protein and in RNA. But still, there is a high need that we need a way back addresses this like resolution differences.

• 1:13:20Speaker 01

  Oh, sorry, I forgot something. So in other other cases, when we get when we measure these like spatial proteomics and transcriptomics, the difficulty is that we cannot actually using the same slices. So usually these slices are at a different location. They are very adjacent, but still they are kind of like shifting in in these cases. Like what I said in the photos, like we taking photos with the different locations. And then we're making some more difficulties in the in the data dealing.

• 1:13:53Nathan Labenz?

  Yeah, it's it's just a good primer on the difficulty of doing all this sort of stuff, especially if it has to come out of actually a living person. I mean, it's a little easier, I guess, if it's coming out of a an organoid or whatever that's um that's developed in a lab. But you've got, you know, a again, in a living person, you've got like a procedure to actually get the tissue sample. There's, you know, uh different, as you said, like different times that that might, you know, that those measurements might be happening at. They they might be actually like just distinct parts of the tissue nearby, but not the same. Then you have the the challenge of. The different kinds of measurements.

• 1:14:35Nathan Labenz?

  And so this like basically creates like a very confusing picture that is not integrated into a single understanding. This kind of reminds me also a little bit of, I'm not by any means an expert on this, but there in an episode, maybe a year ago, where we looked at a bunch of different applications of the new, what was then the new Mamba architecture to biomedical imaging problems. There was a similar thing that came up where it was like, you know, we want to be able to integrate different kinds of scans, you know, whether it's whatever, there's all kinds of different scans. You've got your MRIs, you know, you've got your ultrasounds, whatever. You've got all these kinds of different ways of trying to get a picture of the body and

• 1:15:20Nathan Labenz?

  they have their different strengths and weaknesses. And if you have two of them side by side, taken at different times, you know, when the person was like in a little bit of a different bodily position, whatever, now you can have a real challenge to try to figure out, how do I integrate these pictures? Right? So one of the interesting applications that we saw in the, in that Mamba use cases episode was taking different scans of different types and figuring out how they needed to be deformed in space so as to create one, you know, coherent, single view that could allow a clinician to like, look at one thing and kind of see these different signals in an integrated way.

• 1:16:04Nathan Labenz?

  So it's sort of a similar problem here where you've just got these like different angles of trying to see what's going on and it's like very hard for us. This is another example of just where AI is like, so, and there's other techniques, not all the techniques here are learned techniques. It's quite a mix of things, but it's a, it's another indication. This has been like a huge theme of this. Podcast experience for me is just seeing all the different types of problems for which humans just don't have good intuitions. And at best, we can get like a few experts to like,

• 1:16:44Nathan Labenz?

  with a lot of time and effort, you know, develop the intuition, but then they can only process, you know, even if they do develop that intuition, they can only like process, you know, so much, right? And what we see just over and over again in so many different domains is that AIs are like able to learn the sort of intuition that is needed to make sense of these things. And then obviously, once they do, they become like dramatically more scalable. So those are, those are just like big themes that I think like very much are echoed in this work. Another, I guess, maybe you could, I do want to go to the synthetic data question in just a second, but before we do that, because that is another big theme in distinct from the ones I was just

• 1:17:29Nathan Labenz?

  waxing on about, what are like the big sort of contributions here? One big one that I noticed was simply taking one of these higher scale things and de-convolving it, which is a word you don't hear too often. We hear a lot about convolving and less about de-convolving, but taking something that is higher dimensional but lower resolution and actually going down and saying like, okay, this is what we now think each of the individual cells from this like larger sample actually looked like. So maybe give us a little bit more on sort of what the like practical upshot and value is that comes out of this whole setup.

• 1:18:08Speaker 01

  I would say like the de-convolving definitely is the major task that the model is trying to address. But the thing is also the model tried to providing a comprehensive analysis of across like integration of the different modality. And, you know, transcript terms and proteomics is a wing example. We can also, you know, transfer to other type of modality as well. So besides this like high resolution P two B like deconvolve, I would say like the in terms of the comprehensive analysis. As I said, like it's very important to understanding like how the cells are communicating with each other. So in that case, the model will also modeling the interaction level of like individual cells.

• 1:18:51Speaker 01

  So we we creating a graph new network based model and to also inferring the interaction between the cells. Another point is that once we have a high resolution of the of of this data, we can al- always like to exploring the single cell level modality. And you know, the and and latent features of them. And then making the, you know, the spatial regions to be more like organized and then figure out like the structures that happens in the tissues. And then study the tissue architectures. And and here we call like we identify the functional domains. Um. So I think this is the major things that in in models, but you know, I also forgot something, maybe.

• 1:19:34Speaker 01

  We can also to investigating this spatial variability, kind of like happening inside of the tissues, yeah.

• 1:19:43Nathan Labenz?

  Yeah, and maybe just for people like me who are not well schooled in all this. I don't have a good intuition for like how.

• 1:19:54Nathan Labenz?

  You know, obviously, the body is like highly differentiated, you know, we it's a. A wonder of nature that we start off as even just one cell at one point and then, you know, a bunch like a blob of like. You know, not super different differentiated cells and then all this differentiation happens and then we've got like. In the end, you know, just many, many, many different types of cells and like some of them seem to have like. Sort of gradual transitions, but then other times they have like more abrupt transitions and so. You know, from my like rudimentary study of biology in years past, I I do have some sense of like how segmentation sort of happens. Based on like.

• 1:20:37Nathan Labenz?

  gradients of certain signaling factors. And it seems like a key assumption, if I understand this work all correctly, is that like, this is going to be fairly gradual. There's like a couple of ways, I guess, that that happens. One is that the the graph neural network is sort of modeling cells interacting like directly on their neighbors. And I I didn't, if there is a way that they're also like modeling sort of longer distance interactions, I didn't see that. So that was interesting that it it was seemed to be like a very sort of local causal model. And then there seemed to be also a part of the loss function that was meant to like, try to keep things smooth so that, you know, as you go through a tissue and you see like,

• 1:21:21Nathan Labenz?

  okay, well, over here, it's like the tissue is this way and over here it's this way. There's some sort of an assumption that like, this is going to happen gradually through space.

• 1:21:32Nathan Labenz?

  So maybe, you know, need to be de-confused about some of that. But the thing that also kind of, I'm having a little bit of a hard time sort of reconciling in my own mind is like, for certain like parts of the body, it seems like the, you know, the the sort of barrier or the sort of the boundary, let's say, boundary is better word than barrier, is like, pretty clear, you know, like, when I think of like a bone and then like the, you know, tissue away from the bones, like, seems like I know, like, right where that bone stops and where the other tissue starts. And it doesn't seem like there's like a, you know, smooth transition. Maybe there is, and I'm just not like zooming in, you know, close enough to see it.

• 1:22:12Nathan Labenz?

  But yeah, I mean, that's more of a prompt than a question, but like, what am I missing there, or, you know, what what sort of additional information would like, help me have a better intuition going forward for all this stuff?

• 1:22:26Speaker 01

  Of course, I think like the tissue actually depends on like the type of the tissues that we are studying. Like for them, as you said, like the bone is having more boundaries and then also like the blood vessel have like boundary as well. But in some of the cases, for example, in the like tumors, for example, they have maybe not that kind of clear boundaries. They are, they have some like infiltration of the tumor cells into, you know, the normal tissues. So I guess like it's very depends on the the tissues that we are studied and and in some of the cases, like, it's not clear to identify their like structures directly like, you know, from the morphologies, but if you're looking on.

• 1:23:06Speaker 01

  The molecular, they actually having some uh interesting patterns that existed and and we are only able to see in that ways the measurement of the molecular information. Um and in terms of the the thing that you concerned you mentioned, advise advise that are related to how um the models can uh capture these boundaries and also addresses these uh interactions. Um I would say like we we are based on the quality of um the spatial data we can get uh like from, you know, both the high resolution and uh high spatial resolution and uh high uh gene resolutions.

• 1:23:44Speaker 01

  Um you know, like um the model itself um is able to modeling these interactions between the neighbor cells, but uh probably like not a little bit detailed in the models that um we are not only just taking the most adjacent, we but also taking the nearest neighbor cells. And then we can adjust it these like K values and then making this like, you know, the network graph like more more bigger, uh but I would say um it's not necessary to make it too long, cause uh we were just focusing um the most uh the ne- neighbor cells in communications in these cases. Um because the there is like so many complicated things happen in the tissues, so the way

• 1:24:28Speaker 01

  we trying to simplify is very to understanding the cell itself and also their neighbors look like. And in terms of uh the the I think it's the graph Laplace uh loss that I added in the model. Um we are making them into like more smooth, but I think it's like a a regulation like regulation terms that we trying to make in the models to be more uh realistic and we have some other um laws related to the reconstruction of uh the data and as well as like the K L D divergence uh things. And in that case, it will have a balance between the smoothness and the sharpness of the bound boundary. Because usually if we have a single cell level of the data, they will have

• 1:25:13Speaker 01

  clear boundary because the high resolution of the single cells are able to distinguish a cellular level. So we will exactly see a boundary of the cells that related to the bone and also related to the blood vessels, which even give us like more higher or detailed structures that indicating the inner structures of the of that organs and tissues.

• 1:25:40Nathan Labenz?

  Gotcha. So the reconstruction loss incentivizes accuracy and and sometimes there is a sharp boundary. And so that incentivizes capturing that. And then the smoothness sort of tries to smooth things out in general. And you said that sort of acts as like a regularization. And the, I mean, we've seen this, you know, plenty of different setups, like a clever compound loss function can take you far. So you hope to get the best of both worlds there. And then did I understand also that like, basically, the sort of K value that you mentioned, that's like a hyperparameter for how many nearest neighbors each cell

• 1:26:21Nathan Labenz?

  will be connected to. So you could sort of scale up or down with just that single hyperparameter. Like, I just want to look at, you know, the very small number of the most adjacent cells, or I could like, look a little bit, or even more, you know, kind of expand the radius with that, with that hyperparameter. Yeah, okay.

• 1:26:39Nathan Labenz?

  Cool. How about the, on the synthetic data side, you know, this is something that I think in the context of like, language modeling has been, like, the community's been on a bit of a roller coaster with this, where it's been like, oh, you know, we have a data wall. Oh, don't worry, like synthetic data will take us there. Then it was like, oh, you know, synthetic data creates like, bad models that sort of have mode collapse or whatever. That seems to have been, you know, that that idea got a lot of traction for a little while. I would say at this point, it's like, we're pretty well past that notion. And it's pretty clear that like, just, you know, having models solve problems and doing reinforcement learning on that is like, working very well. Thank you very much. But in the biology context.

• 1:27:23Nathan Labenz?

  Yeah, I've just had a lot less exposure and I have a lot less intuition for what sort of synthetic data exists? How like, how much should we trust it? You know, can we, can we train models on that and be confident? Also, like even just the mix here, I, I don't think it was all synthetic data that was used, but there was like some of it. So I'd love to just understand kind of the role that the synthetic data plays and sort of what are the limits today to like the quality of that data? And then, you know, as a result of that, like how much of it we can really use in a, in a project like this?

• 1:28:04Speaker 01

  Yeah, in my project, I also utilize the synthetic data as well as like experimental data to validate like how the model can work as what we supposed. And in terms of the simulation part, we usually like having those like well-known technology or like models in the field that people creating and for, you know, for this, for the both the single cells or like spatial as well. They are more realistic in related resemble to experimental observations. And mostly probably similar like with other fields, like we are interested in like how their distribution look like.

• 1:28:43Speaker 01

  And then in those of light simulation models, they, they are actually to resemble the similar type of light distributions. And the sample that from the design, you know, structured distribution. And in, in my cases, squid if is like, you know, using this letter, it is like well-known simulation ways to the single cell RNA. But in the spatial, you know, the color model, which building with a lot of like spatial data, we are actually working with our collaborator, who is a assistant professor now in the University of like Conitex. And he's designed a model called I C design, which published in Nature of BaoTec, I think two years ago.

• 1:29:26Speaker 01

  And then this model is like able to sampling from a, a design distributions that having a spatial information as well as having a. A similar um observations with the spatial uh and gene expression. Yeah, so this is the about the synthetic data we are working on. We having a real data as well, and that is a little bit uh different because we have preknowledge about like what's this tissue look like and then can we actually identify uh some interesting patterns that existed and we're always having a domain knowledge like to validating the model.

• 1:30:04Nathan Labenz?

  So it's safe to assume that just the reason we need the synthetic data, of course, I guess, is that it's just hard to gather the real data. Sometimes these things sort of feel like a little bit of a a hall of mirrors and you hope it's not a house of cards like uh there there feels like there's something uh that can kind of be circular about it, I guess, where you're sort of saying, well, we've got this ability to generate this synthetic data, can we learn from it? But I'll when I see that pattern, I also sometimes think, well, geez, if we can generate that synthetic data, like, don't we in some sense already know what we need to know, you know, so? Yeah, again, I I feel like I'm just constantly asking you to help me sort of get less confused, but

• 1:30:47Nathan Labenz?

  what is it that you are learning from the synthetic data that is not already in some sense learned by the process that generates the synthetic data in the first place?

• 1:30:58Speaker 01

  That's a very interesting question, I would say, like, we actually well, using this synthetic data to validating our models like ready to capture the results we want, so um we will not like exactly using the same type of uh generation process, we were having like a uh data generated, I know synthetic data generated from a uh well-known or like a common way um that people observed. And uh the good thing about this simulation data is we having the ground truth of that, like for example, we are interested these uh interactions or like we are interested in the domain things we know like the region or like the domain look like or like the region look like.

• 1:31:42Speaker 01

  So we can kind of like validating whether the model have the ability to exactly identify that domains and also can deconvolve into high resolutions. But the real thing is, you know, like in the real cases, we don't have that ground truth. And then once we can validate in the model are able to perform well in synthetic data, we can transfer the models into real cases and then we can able to trust the results came out for the real data. And then we don't have any further informations about that. So this is all about the way that we first using the synthetic data to validating the model and then we will transfer the model to real data. And then they need to training, you know, separately.

• 1:32:26Speaker 01

  It's very, it's totally different things. And as I mentioned again, maybe emphasize is that most of the synthetic way to getting, you know, is different ways what we designed. So we will make sure that there's like no loop actually happens, you know, in the, in the whole project.

• 1:32:44Nathan Labenz?

  Yeah, gotcha. Okay. Yeah. When ground truth is hard to come by, things can, can certainly get really tricky. Do you feel like the, you know, there's talk these days about like a Manhattan project for superintelligence and whatnot? And, you know, sometimes, and then when, when it's like very fuzzy, often like, well, what is that superintelligence supposed to look like? And like, what's it supposed to do? And, you know, are we going to be able to keep control of it and so on? And I sometimes feel like when I talk to folks like you who are doing such, you know,

• 1:33:24Nathan Labenz?

  detailed work in, especially in biology, I feel like maybe a worthy Manhattan project would be. Just to get all the data that we might really need to scale these things up and figure them out. And I suppose that could come about in like multiple different ways. I have to imagine a lot of the data that we need is sort of locked up in like medical record systems and it's like hard to share because it's, you know, privacy rules and, and obviously those have their place, not to suggest that they don't, but you know, it seems like liberating some of that data in some responsible way could be just a huge unlock. And then maybe even just like a major investment,

• 1:34:09Nathan Labenz?

  you know, in the form of potentially just like government subsidy, if we're going to like subsidize AI development, because I didn't finish the thought on the superintelligence, but like, you know, what's it going to do? Whatever? Can we keep control of it? The, the number one most compelling answer to me generally, when people like get excited about superintelligence is it's going to, you know, cure all the diseases. We're going to live longer, healthier lives. It's going to be amazing, you know, maybe double your lifespan, whatever. Okay. So I'm ready. I want that for sure. And I'm just kind of wondering like, maybe instead of building a trillion dollar data center and trying to like, create a superintelligence that can solve all our problems, maybe what we should do is get all the data so that people like you have a lot easier time of

• 1:34:53Nathan Labenz?

  like, making the discoveries directly and we don't necessarily have to go through this like, undefined sort of amorphous superintelligence. And maybe we end up with both. But I guess the question there is like, how different could your life be in, you know, say a year or two years, if we had a like strategic nation level project to make the biological data available that would, you know, facilitate the next couple of levels of scale up and improvement in the sort of work that you're doing?

• 1:35:29Speaker 01

  Yeah, this is a very interesting question. I would say like there are actually many or like more and more projects like came out like trying to kind of like making the models like big and then you know the making the super intelligent models to very resemble the the human bodies and like for example, I think like for example, I think Google DeepMind, they have creating projects like a virtual cells and and um they wanna having a unified models that uh can making all these like different type of you know cells in organs and and then uh in different scales together uh to really you know having a digital sort of like copies of you know all the cells.

• 1:36:11Speaker 01

  And also I know that C V I, the Tron Darkberg Initiative, they also launched another project, they call like Billion Cells, and they will have like lot of cells together to you know building such like a similar model as well. So I I think there is like very exciting, this is a very exciting period that people are really utilize those like formation models and then large data, like large scale models to address these questions. And you know, we have G P T and and people are also creating the biological version of the G P T, you know, to answer any questions related to the cells and then in other cases, like people interesting the, you know, the clinical questions or like the

• 1:36:55Speaker 01

  health healthy questions. So we can have like virtual doctors and there is like many ways and I would say there's like lot of applications, very excited time points that we we are trying to working on now and you know, people are also working on digital twins start projects. We we can maybe have a copy of our own self and then we knowing like what what's going on if we you know, making different choices at some time point and then what we will look like after that. That is uh also very interesting. And yeah, I think there's like lot of uh very exciting and interesting questions that came out and I would say we are excited to see how the model will goes into into one to two amount, well, sorry, one to two years and um

• 1:37:41Speaker 01

  looking forward to the fields, how they will develop.

• 1:37:46Nathan Labenz?

  Yeah, it's an exciting time for sure and a little scary. But this sort of work is much more pure upside, it feels like to me. I mean, obviously, anything like in in some possible scenario could be abused. But I don't think we have to worry about squid if or coral, you know, getting out of control. It is, you know, it's it's focused work that is answering like specific problems. And, you know, it's it's a good reminder that such work, you know, still happens and still has value. And it's not all just about embracing the bitter lesson and, you know, maximizing the cluster size that you can throw at something. So I really appreciate that about both of these projects.

• 1:38:28Nathan Labenz?

  I think that's about all the questions that I had for you. Is there, you know, any other next steps you want to sort of tease or any other thoughts you want to leave us with?

• 1:38:41Speaker 01

  Yeah, I would say like, I think I'm very optimistic about like how the AI can help with the healthcare and also about medicine. I know people may be like scared about like how how like fast they can grow. And at least like for now, I don't think there is like a way ready to, you know, scare about like how the AI can destroy, you know, the the world of probably like for now. Yeah, I hopefully I think the researchers like who are working at the intersection of the AI and plus the healthcare is we're making our development of the healthcare like to be more to be quicker and then people will have a better life, you know, the styles after the studies like all of our projects.

• 1:39:26Speaker 01

  And that's all I think I want to share. And yeah, thank you so much for this great opportunity and very excited and the delightful discussion with you. Yeah, thanks so much.

• 1:39:36Nathan Labenz?

  My pleasure. See you. Huh. Thank you for being part of the cognitive revolution.

• 1:39:41Speaker 01

  Thank you.

• 1:39:42Nathan Labenz?

  It is both energizing and enlightening to hear why people listen and learn what they value about the show. So please don't hesitate to reach out via email at [email protected] or you can DM me on the social media platform of your choice.