Tomorrow's Medicine, Ep. 17: Math and Biological Data - Jingyuan Hu

With Jingyuan Hu, Postdoctoral Researcher, Harvard Biostatistics. Hosted by Arseniy Arsentyev. Published October 13, 2026.

https://aarsentyev.com/podcast/ep17-jingyuan-hu

Machine-transcribed and lightly edited: filler words and false starts are removed, and the wording is otherwise as spoken. Where the two differ, the recording is the record. Times are minutes into the recording. Not medical advice.

[0:00] Jingyuan Hu: Why does a cell choose a particular fate? So it turns out the dynamic information can be inferred from the snapshot. One very exciting avenue is really the virtual cell, right? If you can just predict the response of the cells, but also the next step of that would be we are able to create.

[0:18] Arseniy Arsentyev: So welcome everyone. It's the 17th episode of Tomorrow's Medicine. After a long break, we're kicking off season two and right away with a great guest, Jingyuan Hu. Jingyuan, could you please tell the audience what are you doing, who you are, what's your research, and what are the broad topics we'll talk about today?

[0:39] Jingyuan Hu: Yes, of course. My name is Jingyuan. I'm a postdoc at Harvard Biostats. I'm a computational biologist. I've been trained many years, and my research focuses on single-cell genomics and computational biology. And I build statistical and AI models for single-cell data mainly. So I've always been curious about how biology works, especially for human, as something as complex as human emerges out of 3 billion base pairs of DNA. And that motivates me to study bioinformatics, do a little bit of research in the undergrad, then pursue a PhD in quantitative and computational biology at Baylor College of Medicine. Then I basically keep going down this avenue in my postdoc at Harvard.

[1:32] Arseniy Arsentyev: Okay, thank you. And your work primarily focuses on the mathematical side of quantitative biology and computational biology. Can you outline to us what is this mathematical magic that is so important with application to biological data? And the second sub-question, could you clarify for the audience what a single-cell data is and how it differs from usual computational biology data?

[1:58] Jingyuan Hu: Yes. So I think mathematics has a very close relationship with modern biology because we are getting more and more data. For what I work on is genomics data. And I like to think of mathematics as a set of tools to use to describe ideas, just like in other subjects, like in what we see in physics or chemistry, where we can write down quantitative equations to describe a biological system. Single-cell sequencing is actually one of the most exciting research areas in modern biological science because we are able to collect the molecular profiles for each cell. For example, a human has trillions of cells, right? It will be a dream to collect a molecular profile for all the cells, from which it is a gold mine for biological information.

[2:51] Arseniy Arsentyev: And what are the limitations of this single-cell data that you're working with?

[3:01] Jingyuan Hu: Yes. So before the single-cell data, it's more common to have the so-called bulk RNA sequencing. The bulk is basically when you take a tissue and then measure the total RNA molecules within that tissue. But the limitation of that is you kind of lose the cellular resolution of it, right? You don't know what's the behavior of each cell. You just know the total RNA molecules in that one particular tissue, right? You know the expression of gene A, gene B to all 20,000 genes, but for the cellular information, it's missed. For single-cell sequencing, it's remarkable that we are able to obtain that information, right? We are able to kind of dissect each cell type and look at their RNA count at that cellular resolution.

[3:57] Arseniy Arsentyev: And this single-cell data shows us difference, I believe, thousands of genes, and it is highly dimensional kind of data. In your work, as in the work of some other great researchers, there is this highly mathematical concept of a manifold, that we can represent this kind of singular cellular data on a manifold. Can we first try to explain what the manifold is for the general audience and why is this geometrical intuition so important in context of this data?

[4:30] Jingyuan Hu: Yes, of course. So a manifold can generally be thought of as a high-dimensional surface. For example, when you have a curved surface in the 3D space, that is roughly equals to a manifold. But a manifold formally is defined mathematically as geometrical objects. So for single-cell data, that's very high-dimensional. A central hypothesis is that the data itself actually lives in some low-dimensional manifold plus noise. It just means that the data doesn't distribute randomly across all the dimensions, but rather have some structures in it, whose intrinsic dimension is much lower than the total number of features space.

[5:20] Arseniy Arsentyev: Okay. And is this concept connected to your main research and the field that you're an expert in, which is cell dynamics, RNA velocity, and the software projects that you're working on? Can you connect all the dots and tell us how these geometrical concepts are used in real-world research and especially in the research you're doing?

[5:45] Jingyuan Hu: Yes, yes. So my work, one of my main interests is to understand the cell dynamics with single-cell data. So what this means is that single-cell data, like one, actually one limitation for single-cell data is that it is a static snapshot of the cell state. Because for each time when we collect the single-cell data, we have the gene expression profile, right? But that is taken at one particular moment because all the cells need to be sacrificed when we are taking the measurements. However, the real cellular process is dynamics. Just think about our body. Our body is changing all the time.

[6:22] For all the cells, they are changing all the time. Our body is replacing hundreds of billions of blood cells every day. And that information is not really captured by the single-cell data, which is one of the limitations of it. But how do we study cellular dynamics using single-cell data? That's the key challenge. So it turns out the dynamic information can be inferred from the snapshot. Kind of one analogy for it is just kind of think about when you take a snapshot of a highway with moving cars. Even you have a snapshot, but from the directions the cars are pointing at, you kind of know where the cars are going, right? So that becomes a dynamical process. And for the single-cell data, right, it has its own geometrical structure.

[7:08] Just like a highway, there are different trajectories. Just think of the cars, where the cells are going to head out next. And that's something we want to infer from the data.

[7:20] Arseniy Arsentyev: Are you using some kind of advanced machine learning algorithms, computer vision algorithms? And what are the state of cutting-edge methods right now in this field?

[7:30] Jingyuan Hu: Yeah. So currently, for my work, the single-cell data measures a molecular profile, right? So that's actually a little bit abstract. There are actually different approaches to this problem. For example, taking the microscopy or the live microscopy, and that allows you to directly look at the cell. But for single-cells, the space, the cell state, and the geometry we're looking at is a little bit abstract. It can be thought of as a phase space in physics, because sometimes it's actually easier to look at more abstract quantities, such as the momentum, position, the acceleration, those physical quantities, rather than just looking at the physical space for its dynamics.

[8:23] And that's kind of a way to think about the single-cell data. We are looking at the cell state space defined by the molecular profiles of each cell. And that is kind of an inferred geometrical object. So, yes, we do use statistics and machine learning algorithm together to kind of explicitly infer the geometrical manifold where the cell state lives. And then we also use methods such as RNA velocity to infer the trajectories and the dynamics of how cell states are going to change.

[9:05] Arseniy Arsentyev: And so you're primarily looking on the dynamics inside the cell. Is it correct?

[9:11] Jingyuan Hu: Yes, yes. That can be thought about as a continuous process. It's like just so we can kind of think about the process of the blood generation, right? So all the blood cells, the immune cells, red cells, et cetera, et cetera, they are differentiated from the hematopoietic stem cell, right? So when you only have one cell state initially, then it gradually differentiates to other cell types. And this process is very much continuous. That's like the main observation we see from the single-cell data. The cell does not suddenly change state to red cell, but like rather it's a continuous process in terms of its molecular profile. And that's the main thing we are trying to infer from the single-cell data. So it's like the change of the molecules within that cell, as you said.

[10:03] Arseniy Arsentyev: So we can say that this transformation is continuous. Do cells have kind of a critical values, if we can say so, when certain chemical change is undergoing and it immediately changes state? Or is it more of a, as you said, continuous, maybe convex functions that we can do pretty math on?

[10:25] Jingyuan Hu: Yeah, I think it's a little bit both. For biology, reality is always a little bit complicated. For the single-cell data, roughly, we can see the transition process as continuous. Yes, there are definitely critical points where the cell states and the regulation go through rapid change from one generation to another. Because the division process is actually discrete, right? One cell divides into two cells and that process can be asymmetric.

[11:00] Arseniy Arsentyev: And Jingyuan, you've been working in this field for a long time. And what motivated you to build, I'm always forgetting the name, FlowMap? And can we spend a couple of minutes talking about this software and how it combines everything that we've talked so far into a useful tool for researchers?

[11:20] Jingyuan Hu: Yes. Okay, so I would like to start the motivation of my work, FlowMap, with an analogy called the Waddington's landscape. So it is being proposed as a geometrical metaphor that in order to study the developmental process, you can think of it as rolling a ball down a hill. So this ball will just follow the landscape of the hill and then become like various kind of cell type when it reaches the end. So kind of this metaphor is kind of the key mental picture that we want to model with single-cell data.

[12:14] The reason we can use this analogy is because just think about the developmental process. In many ways, it is deterministic, right? Because human, we started with one cell, right? So for human body, we initially started with one cell, the embryonic stem cell. Then that cell gradually divides into trillions of cells. And there are actually different stages in the developmental process where the cell states just kind of emerge in a programmable way. So that's why this analogy is useful because there seem to be different paths and trajectories that is deterministic in this process. For my work, we can actually infer this landscape from the single-cell data using the molecular profiles of each cell. And after we learn this landscape, you can also perform all kinds of inference.

[13:17] For example, why does the cell choose a particular fate? For example, there can be two stem cells that looks very similar in terms of their molecular compositions, but their fate could be predetermined. So these are the things we can kind of infer after we learn this landscape.

[13:36] Arseniy Arsentyev: Okay, and have you seen some interesting applications of it?

[13:41] Jingyuan Hu: So we actually want to expand this framework. So currently, our focus is mostly on the inference and the validation of this geometry. So we used known processes such as hematopoiesis process, the process of making blood, mainly as a way to validate this geometrical object. And also kind of for proof of concept, we apply this on the B cell activation data set. I feel like that's mostly our next step. Like currently, our main work is still to just kind of infer this geometry from the single cell data, prove that this is a valid way to interpret biology. And also kind of the known biological process, like the hematopoiesis, where we know the developmental trajectories. And we'll see if we can predict the fate for the stem cells, like what cells it's going to differentiate into.

[15:09] But that will be a next step, exactly what we are proposing and we are working on right now.

[15:13] Arseniy Arsentyev: Let's kind of separate FlowMap into different stages. First is the data that you are working on to get some outcomes. Then it's the main engine, so the software, the FlowMap. And final part for everyone in biotech and biosphere and computational biology is validation. You've touched on it just a little bit, but I want to focus on it a little more. And could you please expand, how do you validate the predictions that your models are doing? Do you have a wet lab? Do you simulate on some other artificial cells, computerized cells?

[15:55] Jingyuan Hu: Yes, I think this is a huge challenge for the community right now. Because we are generating more data and the computational tools are becoming very powerful right now. So the validation throughput is one of the main, I would say it's one of the main limitations for the entire field. Because for single cell data, it's actually easy to generate hypotheses given a process. But for the validation, sometimes we can only look at one gene at a time. There's a bottleneck for us to determine the signal we observed are true or not. But this problem is actually very relevant to the translation of science. Because at the end of the day, for the process of any downstream process like drug discovery, we sometimes need to zoom into one particular target.

[16:53] And that's the main challenge.

[16:57] Arseniy Arsentyev: And what are the challenges and what are, as you said, the recent advances in the data part of it? Are data sets becoming better? And what in statistical, maybe in biological terms, a great data set means? What is a great data set for single-cell genomics and single-cell research?

[17:16] Jingyuan Hu: Yeah, that's a great question. So for the single-cell data, I would say that our dream would be just measure all the molecular profiles across all trillions of cells for the human body. That's the dream for single-cell RNA sequencing. That's only one modality of this data. We're saying that we just take a bunch of cells and then measure the gene expressions of those cells. And that's what the RNA-seq gave us. And of course, there are many other modalities of the data. For example, the single-cell ATAC-seq that measures the chromatin accessibility. The CITE-seq can also measure the protein quantities. And also we have spatial single-cell that just not only give us gene expressions, the molecular profile, but also give us a spatial location of the cells. And also the dynamical process, as I mentioned.

[18:11] But that's even more challenging because unless we are doing live microscopies, it's very hard to measure that. Sometimes people take a snapshot at several time points as kind of a discrete way to model this dynamical process. So that would be the dream. In reality, for each modality, they always come with their own limitations. And for us computational people, sometimes we need to, like most of the time, we need to kind of integrate all the data together. Because it's also very difficult to measure many modalities in one cell. Whereas sometimes we measure cells in the same tissue, but using computational methods to kind of integrate all the data together so that it has richer molecular profiles than just one modality can provide.

[19:01] Arseniy Arsentyev: So that you're not only researching this particular field of applications of math and single-cell data and other quantum biology frameworks, you're also thinking about more global, which is scaling laws, AI and biology, and building the foundational models that truly understand the biological reality and can have statistically probable predictions. And great outcomes as a result of these predictions. And so my first question to you is, do you think that we just need more data to obtain such a general biological foundational model? And how much data do we actually need?

[19:45] Jingyuan Hu: Yeah, that is a very important question right now. So when we think of scaling law in modern AI, that actually works incredibly well for the language models. As we all see, when you increase the complexity, the model size, the performance of the model got better. And the remarkable thing about scaling law is that this is happening in a predictable way. One of the famous plots shows that how much you increase your model and the amount of computation you applied on the model, the performance of the model can get better. So biological data, you have the scaling law, but the behavior we found is actually looking different than the scaling law for the language models. Partly because the biological data is different from the language. So kind of two interesting things that we found.

[20:46] The first is that the noise of the biological data has a very different structure than the language model. Because for the single-cell data, for example, that's oftentimes used to train those foundation models. It has lots of technical noise. Whereas if you think about human language, I wouldn't say that language don't have noise. But I would say that the order of the words has a lot of structure in it. Whereas the single-cell, basically, it just has a lot of technical noise due to the measurement itself. So I feel like that's kind of one of the main differences between the two. And the other difference, I would say that the biological data only gives us a partial information.

[21:31] Even when we kind of integrate all the modalities together, but there are always information that are missing from this picture. For example, the entire dynamical process that we talked about, that's not really in the single-cell data when we train those foundation models. If you think about the language models, relatively speaking, that can be a little bit self-contained. When you're given a bunch of math equations or computer code, all the information are kind of in that training data itself.

[22:11] Arseniy Arsentyev: Great. And one of your statements is that random matrix theory helps to separate the signal from noise. And as you infer, separating signal from noise will accelerate the scaling law for single-cell and for foundational biological models. And random matrix theory for an average American, average person worldwide sounds like PhD, MIT stuff. Could we explain it in simple words and try to confer the deep meaning behind it? And why will it help to separate the noise from the signal?

[22:48] Jingyuan Hu: Yeah. So random matrix theory, the setup is actually surprisingly simple. When we have a huge matrix and the numbers in the matrix are random. That's what we study. That's the object we study in random matrix theory. So this is actually very relevant to modern data science. So basically for the tabular data that we see, right? As long as the numbers in those Excel sheets, those tables are random, we can use random matrix theory to separate out the signal and the noise of that matrix.

[23:39] And that's actually one of the most used mathematical tool in data science. It's called principal component analysis, where the spectrum, where the eigenvalues of this matrix tell us where are the signals and where are the noise. And actually the noise part of the matrix have a very predictable behavior captured by the RMT. That's kind of separated out from the signal. And using this theory, we're able to look at the scaling law, namely when the matrix becomes larger and larger, or the signal-to-noise ratio of this matrix is getting better, then how does the spectrum behavior changes accordingly? So this is kind of how we use RMT to study the scaling law.

[24:33] Arseniy Arsentyev: And do you believe that if we separate signal from noise, if we improve the quality and the quantity of data, can we achieve the very, very kind of transformative goal of creating a virtual cell and creating a model that we can test hypothesis, as we said, for validation without actually dealing with real-world wet biology?

[24:59] Jingyuan Hu: Yeah, that's a very good question. So one interesting thing we found is that using the random matrix theory, that the current way of collecting single-cell data will have a saturation point. Generally speaking, more cells can give you better performance, can give you more biological information, but there's a hard ceiling to it. Like when you sequence a cell to a certain point, and then adding more cells gives you very little additional information. And I think that kind of makes sense, right? Because the RNA sequence is only one modality of the biology, right? There are many other molecules we didn't go to measure.

[25:45] So it's really not only about collecting more cells in the same form, but rather like collecting many data across many modalities, like many different types of cells, that kind of adding more information to the model. So the way I can think of it is that it's not only about the quantity of the data, but it's about the quality of the data. And whether the new data can add more information to the existing, I think it's critical to the scaling to build AI models, such as virtual cell or other like world model for biology.

[26:22] Arseniy Arsentyev: So we can summarize that progress right now is overhyped in AI for biology general models. We can certainly see certain headlines announcing that we've solved the cell, that we've solved the genome expression. And from your monologue on this aspect of AI scaling, I can infer that we're currently not there.

[26:47] Jingyuan Hu: I think that's a little bit hard to predict. Just like in our current research environment, a few years ago, I would not imagine, as of 2026, we have all these amazing tools and all the things happening in the AI community. The biology community is moving very fast. But as of now, I would say this is an open question or open challenge. It's really about what data do we want to collect. So right now, the effort has been made mostly on the single cell. That's one of the big reasons is the cost, right? Because the sequencing is relatively cheap to perform. We can kind of easily gather a lot of data for the RNA sequencing. And also people are collecting a huge amount of perturbation data, like how the cells respond.

[27:39] For example, how the cells respond to the drugs or all kinds of interventions, right? That's all the data that are relatively cost-efficient. So there's a lot of effort being made to collect that. And the virtual cell problem you mentioned is a great example. By using this perturbation data, the idea would be we are able to simulate the cell response in a computer. Given one particular intervention, can we predict the outcome of the cell? The question that we are unsure is whether the current data set is enough or not.

[28:26] What other data do we need to collect to improve the performance? Whether there is a hard ceiling to the performance if we only use the current data we are collecting? These are the questions we are not entirely sure of. And tying back to our current work, we kind of see that across all the single-cell data, the perturbation data, there seems to be a hard ceiling for the biological information if we kind of keep collecting the data of the same type.

[29:02] Arseniy Arsentyev: And let's try to close the cell dynamics and foundational models in biology before proceeding to math and speaking about what are you doing on YouTube and other stuff related to math and biology. The last question on the topic of cell dynamics is, do you think that advances in this field can practically change the treatments of diseases like cancer in a horizon, let's say, of 5 to 10 years? Do you see that it is possible? Or is cellular dynamics a field that only pushes the primary science, the base science, and not the applications?

[29:42] Jingyuan Hu: Great question. But first, I feel like I may not be the best person to speak for this because my work is mostly in the basic science side and on the relatively theoretical side for the statistical and machine learning modeling of the genomics data. And the translational side, as we said before, one of the hard challenges is about validation. If you want to perform everything experimentally, that can be very expensive and it can be very slow. So I feel like one very exciting avenue is really the virtual cell, right? If you can just predict the response of the cells, a computer can just increase the discovery iteration so fast, right? Because we can kind of choose what one particular experiment to do and then maybe using simulation approach for all the other hypotheses in parallel, right?

[30:50] So I feel like that's one of the most exciting future directions for this.

[30:55] Arseniy Arsentyev: Great. And now let's proceed to math and biology and your channel. For everyone listening, Jingyuan Hu is the name of the guest. It's also the name of a great YouTube channel exploring some concepts in mathematics that he does apply in his research to biological data. And in preparation for this episode, I've seen a couple of his videos. It's a great content. It's explained in a very, I believe, understanding pace because most of the math videos are too fast for an average person to understand. And Jingyuan puts lots of effort to make it as accessible as possible. So can you tell us why you started this channel? Maybe tell us about your love for math and how do you believe math fits into what you're doing professionally, such as quantum biology and so on?

[31:49] Jingyuan Hu: Yeah. So how I started this? Personally, I'm a huge fan of YouTube. I feel like this is probably the best platform for anyone who are curious. And for me, my formal training and background don't really have a lot of math in it. I learned a lot of math and build my expertise from YouTube. It's a platform where you can find all kinds of content, right? From simple to really niche, really hard content. You can always find good series on YouTube. And I started this almost towards the end of my PhD, where I just become very interested in the underlying structure of the biology, where that requires a lot of math to describe the system and also describe the method we use to model biology.

[32:41] So I feel like I just kind of got addicted to YouTube about learning math. And I started from the very basic stuff, like calculus and linear algebra by Dr. Gilbert Strang from MIT. That course is very famous. And I started from basic stuff and then move up, take courses after courses on YouTube. And this whole process made me realize that I feel like right now, like anyone, if you are curious, you can build expertise outside your formal training. So I feel like I really enjoy this process and I want to share my own understanding now through YouTube. That's really the main motivation for me to create this channel.

[33:26] Arseniy Arsentyev: Well, there are some interesting things that we can see in the future of your channel. I've seen you did videos on dimension reduction. Please help me out. Anything else? And what are you thinking right now to put out in the next couple of years?

[33:42] Jingyuan Hu: Yeah. So currently, I feel like the YouTube channel for myself serves two purposes. So first is that I want to learn those concepts and topics. And also according to Feynman, if you want to really understand something, you better teach it. And I feel like that applies to me very well. So actually the topic that I made on YouTube is kind of loosely related to what I do for the research. And I want to understand those topics deeper so I can also apply it. But also they can be very useful for people who are curious about the same stuff. And I want to build it in a very structured way such that it's a complete story.

[34:31] I don't want to just put out content, just have like one piece of math here and another piece of math over there. That's the pieces and pieces are not very related. I want to create this type of full video series that has a continuous storylines. And we can follow this from the first principle in terms of how we make the derivations. And also I feel like my angle is mostly coming from the applied perspective, right? We are able to not only understand math, but also use the mathematical tools as a way to model how we see in biology and perhaps in other fields as well. That would be a goal for me. And there can be many topics related to that I'm thinking of doing in the future. I'm thinking of doing in the future.

[35:17] Arseniy Arsentyev: Yeah, and Feynman probably didn't know that the best way to learn is to invite someone for your podcast and prepare and to get a first-hand exposure. It's really like a binge study. It's crazy when you're preparing for an episode, you're delving into absolutely new topics and you have to have some understanding at least to be able to ask questions and talk about it. And yeah, it's great that you said that you also learn when you produce this content. And you probably have some kind of gut feeling what are plausible mathematical tools to use on the actual data, data in biology, data in medicine. And can you maybe share with us, do you have any ideas what are misunderstood, maybe unlooked mathematical tools that may be applied?

[36:07] Because we said that manifold differential geometry is verifiably useful, and statistics is obviously useful. It's verifiably useful in that statistics is obviously useful. Everything around us right now is statistics, is machine learning, language models and so on. Are there any other mathematical frameworks that you feel might be useful?

[36:26] Jingyuan Hu: Yeah, that's actually a great question. So the way I think about it is that when it comes to a problem, there's a whole spectrum of mathematical models that we can use. So on one end of the spectrum, I feel like we have the mechanistic models, right? So one example we can think of is the physics model, for example, Newton's second law, right? Newton's second law is just one equation, F equals ma, but we are able to describe so many things with this one, three parameters. When you have a force acting on an object, and that can cause the object to accelerate, right? There are so many things in the world we can model this, and I feel like this is closest to human understanding, right?

[37:28] And then on the other end of the spectrum, we have modern AI predictive models, right? For that, we have billions, trillions of parameters. And the main goal for the AI models is about prediction, right? Whether we can predict the behavior of the data or the real-world process through these models. And in that process, people always call it black box because we are not really trying to understand internally how does the model come up with this particular prediction, but it's more about the prediction itself, whether the end result is accurate or not. And I feel like another way to think of it is that AI models can be kind of thought of as effective models for what we observe in the reality. I would like to think that the process of a language model generates text.

[38:23] It's very different from how human generates text. But in terms of the end product, the quality of the text generated can be very similar. The process can be different versus kind of on the other end of the spectrum for the physics model that we are able to use like three parameters to describe so many things in the world.

[38:44] Arseniy Arsentyev: I can also add on this. I believe that current LLMs and current things that we try to apply to biological data are using abstractions that might not align with the real world because they train on text and train on experimental data. And experimental data are based on so many unknowns, so many dependent variables. And I believe that to truly have a mathematical system that, or not mathematical, even just a transformer, just a LLM that foundationally understands the biology, we need to automate the whole cycle and have a wet lab, constantly get feedback, and completely redefine the abstractions.

[39:29] Because biology is this field when people conjure up, they came up with different things that are imaginary, if we can say so. Not everything is imaginary. Math is imaginary, but things that are not easily testable and are not deterministic. In math, we have axioms. We can perform derivations and get the understandings. In physics, we can mostly check the physical results of a theory. In biology, it's very complex. And I believe we need to have AI restate the axioms that we have about biology right now because these axioms were just ideas and they are not proven.

[40:10] Jingyuan Hu: Yes, exactly. I feel like I can also add on to that. The biology is very much context-dependent. Every environment, every condition is different and that can cause biology to respond very differently. I would say it's very hard to reduce. I feel like that's one of the main differences between biology and math and physics. In math, we are able to reduce everything to a few axioms and for physics to a few basic principles. But for biology, I feel like everything is regulated, is connected and is conditional dependent.

[41:09] Arseniy Arsentyev: So this is one of the main challenges. And what is the problem in your fields that you wish we solve in the next like five to 10 years with modern tools?

[41:18] Jingyuan Hu: I would say that the most exciting would be just we are able to make the predictions accurately as we did for other fields using the AI models, right? AI tools has shown that we are not only able to make prediction, but also able to generate data, right? As we see in the language model and the image models. Biology, for example, the virtual cell is about prediction, whether we can predict the response accurately. But also, the next step of that could be generation.

[42:03] We are able to create new biology. And with the functions that we desire, using AI tools. I feel like that's like one of the most exciting areas I would hope to see. I'm not sure whether it's something we're going to see in like five to 10 years. Definitely looking forward to it.

[42:22] Arseniy Arsentyev: Yeah, but some people predict apocalypse in like two years with this AI boom and super intelligence and stuff. We'll see how it all plays out. And returning to the topic, you said more about basic science, more about your field, your research, and what's your take on more applied science of the stuff? And how do you see the innovations in cellular dynamics and single cellular data in foundational biological models will shape the care that we have globally, the healthcare processes, the actual medicine, and the adoption of scientific knowledge to increase humans' lifespan, healthspan, and generally improve the quality of our lives?

[43:08] Jingyuan Hu: Okay, I think that's a grand question because like understanding cellular dynamics is more on the basic side. For the health side, that's really grand. I don't know how many people are working on it. So it's a huge industry. I feel like for all the people who are doing basic science, it's more about like increasing the understanding of the biological system. I feel like that can be a huge indirect value to the entire healthcare industry. Because all the downstream stuff, all the medications, the drug discovery, all the healthcare at the end of the day is coming from our understanding of the knowledge, right? We're able to understand what is the cell, what is the cell dynamics, what's the biological process that is causing all kinds of disease, right?

[44:09] So that understanding is very important. And only after we have this understanding, we can start to think about solutions to one particular problem. For example, cancers and Alzheimer's, there has been a tremendous effort in the past 20 years where we're trying to dissect, because these are very complex diseases, right? Where the single-cell data is very useful. Because for cancer and Alzheimer's, they are the disease that's very personalized. Every person's cancer is different from another person. And only through this huge amount of data and our so-called personalized medicine can we really begin to truly develop new ways of cure.

[45:23] Modern biological data, the genomics data is really one of the key elements to create personalized medicine. You need to understand the molecular response and the profiles of each person in order to kind of develop effective treatment.

[45:41] Arseniy Arsentyev: And you know what, I think you've outlined your vision for the applications of knowledge. And I usually end the episodes with asking questions like, what's tomorrow's medicine? What do you want to see in medicine in 10 years? And you brilliantly outlined the core of it, your vision. So we covered cellular dynamics, geometry in biological data and singular cell data. Then your software projects that you were working on and how it works, how it applies to the data. What are the validation reasons? Then we covered foundational models, scaling laws in biology. And to my greatest pleasure, we covered the YouTube, how platforms works and how you're structuring your channel and how it improves your understanding by teaching other people. So I just want to say, Jingyuan, thank you very much for participating.

[46:40] I think we've spent a great hour talking about very difficult topics. And I hope the audience will understand a lot more than they did before they joined this episode.

[46:52] Jingyuan Hu: Thank you very much. This is a really fun experience. I had a lot of fun talking to you.

[46:59] Arseniy Arsentyev: Yeah, great. Everyone, follow Jingyuan's work on his YouTube channel, on his LinkedIn. He has a great personal site. And as you know, subscribe to Tomorrow's Medicine podcast, like this video, share with your friends. That will help me to get great guests into the show. Thank you very much.